Detecting Offshore Drilling Rigs with Multitemporal NDWI: a Case Study in the Caspian Sea

Home , Drilling rig, Offshore drilling, Oil platform, Storey

remote sensing

Article Detecting Offshore Drilling Rigs with Multitemporal NDWI: A Case Study in the Caspian Sea

Hui Zhu 1,2,3, Gongxu Jia 1,2,3, Qingling Zhang 4,*, Shan Zhang 1,2,3, Xiaoli Lin 1,2,3 and Yanmin Shuai 2,5,6

1 State Key Laboratory of Desert and Oasis, Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi 830011, China; [email protected] (H.Z.); [email protected] (G.J.); [email protected] (S.Z.); [email protected] (X.L.) 2 Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi 830011, China; [email protected] 3 University of Chinese Academy of Sciences, Beijing 100039, China 4 School of Aeronautics and Astronautics, Sun Yat-sen University, Shenzhen 518107, China 5 Research Center for Ecology and Environment of Central Asia, Urumqi 830011, China 6 College of Surveying and Mapping and Geographic Science, Liaoning Technical University, Fuxin 123000, China * Correspondence: [email protected]; Tel.: +86-181-2709-5863

Abstract: Offshore drilling rigs are the foundation of oil and gas exploitation in water areas. Their spatial and temporal distribution, state attributes and other information directly reﬂect the exploitation level of regional oil and gas resources. Therefore, it is very important to build an automatic detecting method for offshore drilling rigs with good performance to accurately capture the temporal

and spatial distribution and state of oil and gas exploitation activities. At present, there are two main groups of methods for offshore drilling rigs: invariant feature-based methods and nighttime

Citation: Zhu, H.; Jia, G.; Zhang, Q.; firelight-based methods. Methods based on invariant location are more subjective in terms of their Zhang, S.; Lin, X.; Shuai, Y. Detecting parameter settings and require intensive computation. Nighttime light-based methods are largely Offshore Drilling Rigs with unable to identify offshore drilling rigs without associated waste gas ignition. Furthermore, multiple Multitemporal NDWI: A Case Study offshore drilling rigs in close proximity to one another cannot be effectively distinguished with in the Caspian Sea. Remote Sens. 2021, low spatial resolution imagery. To address these shortcomings, we propose a new method for the 13, 1576. https://doi.org/10.3390/ automatic identification of offshore drilling rigs based on Landsat-7 ETM+ images from 2018 to 2019, rs13081576 taking the Caspian Sea as the research area. We build a nominal annual cloud and cloud shadow-free Normalized Difference Water Index (NDWI) composite by designing an optimal NDWI compositing Academic Editor: Lalit Kumar method based of the influence of cloud and cloud shadow on the NDWI values of water, bare land (island) and offshore drilling rigs. The classification of these objects is simultaneously done during Received: 22 February 2021 the compositing process, with the following rules: water body (Max_NDWI > 0.55), bare land (island) Accepted: 2 April 2021 − Published: 19 April 2021 (Min_NDWI < 0.05) and offshore drilling rig (0 < Mean_NDWI < 0.4). A threshold segmentation and postprocessing were carried out to further refine the results. Using this method, 497 offshore

Publisher’s Note: MDPI stays neutral platforms were automatically identified using a nominal annual cloud and cloud shadow-free NDWI with regard to jurisdictional claims in composite image and Google Earth Engine. Validation using Sentinel-2 Multispectral Imager (MSI) published maps and institutional affil- and Google Earth images demonstrated that the correct rate of offshore drilling rig detection in the iations. Caspian Sea is 90.2%, the missing judgment rate is 5.3% and the wrong judgment rate is 4.5%, proving the performance of the proposed method. This method can be used to identify offshore drilling rigs within a large water surface area relatively quickly, which is of great significance for exploring the exploitation status of offshore oil and gas resources. It can also be extended to finer spatial resolution Copyright: © 2021 by the authors. optical remote sensing images; thus small-size drilling rigs can be effectively detected. Licensee MDPI, Basel, Switzerland. This article is an open access article Keywords: energy production; natural resource; optical remote sensing; time series analysis; oil distributed under the terms and market; NDWI; Caspian Sea; Landsat-7 ETM+ conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

Remote Sens. 2021, 13, 1576. https://doi.org/10.3390/rs13081576 https://www.mdpi.com/journal/remotesensing Remote Sens. 2021, 13, 1576 2 of 26

1. Introduction Petroleum is often regarded as the blood of modern industry. With the development of industry, world demand for energy, including oil and gas, has been constantly increasing. As onshore oil and gas exploration becomes increasingly mature, the growth of global oil and gas reserves slows down. The scale of newly-discovered oil and gas resources becomes smaller and their exploitation more difficult [1]. Consequently, more and more attention has been paid to exploring and exploiting offshore oil and gas resources in recent years. According to the statistical data of the U.S. Energy Information Administration (EIA), offshore production accounted for nearly 30% of global crude oil output in 2015. Therefore, investigating and monitoring oil and gas exploitation activities in the sea is of great significance to understanding the development of oil and gas resource extraction and understanding global energy development. Traditional reports of oil and gas resource data are mainly included in national-scale statistics [2], such as BP World Energy Statistics [3], International Energy Statistics of the U.S. EIA [4], World Bank database [5], etc. Such statistical reports often poorly reflect the temporal and spatial distribution of oil and gas exploitation activities due to human factors. In contrast, remote sensing has the characteristics of a short revisiting period, timeliness and large area synchronous observation, which makes it possible to monitor the development of offshore drilling rigs with temporal and spatial details. Furthermore, the powerful spatial data management and spatial analysis functions of modern Geographic Information Systems (GIS) provide powerful technical support to manage related spatial data. At the same time, with the development of cloud platform technology, e.g., remote sensing cloud computing platforms such as the Google Earth Engine (GEE), have recently emerged, creating the possibility of rapid and accurate extraction of geographic information on a large scale from big remote sensing data [6]. Therefore, remote sensing technologies and GIS are being more and more widely used to extract data about offshore drilling rigs. Existing algorithms for identifying and monitoring offshore drilling rigs with remote sensing mainly focus on the following two aspects: (1) eliminating false alarms to determine offshore drilling rigs according to their position and shape invariant characteristics in multitemporal remote sensing images; and (2) high temperature and brightness of the associated waste gas flames at night (Figure1). However, there are other advanced techniques such as Machine Learning algorithms. With the wide application of Convolutional Neural Networks (CNNs) in ship target detection on the sea [7–10], this technology has been applied to oil and gas platform recognition [11]. A few studies have been carried out to identify offshore drilling rigs based on their location invariant feature in Synthetic Aperture Radar (SAR) data. These researchers developed a two-parameter false alarm algorithm to identify the location of potential offshore drilling rigs with SAR images first, and then eliminate the false alarm with multiphase SAR data to improve the predictions of offshore drilling rigs in the South China Sea [12–14]. Cheng et al. [15] proposed a method based on the triangular invariant rule to identify offshore drilling rigs with SAR and reported improved detection accuracy of offshore drilling rigs in the South China Sea. Wan et al. [16] compared the shape differences between ships and offshore drilling rigs in SAR data to improve detection accuracy. Wong et al. [17] developed a cloud-native geoprocessing algorithm to automatically detect and extract offshore oil platforms in the Gulf of Mexico using synthetic aperture radar and the Google Earth Engine. Zhang et al. [18] proposed an automatic extraction method for offshore platforms in single SAR images based on a dual-step-modified model. Passing ships are the main source of false alarm in offshore drilling rig identification. Meng and Xing [19] attempted to map offshore drilling rigs and ships using Landsat remote sensing images taken within a single time phase, but the results showed that single time phase optical images could not distinguish the two well. Cheng et al. [20] utilized the high brightness of lighting facilities and associated high temperature of waste gas combustion flames in oil and gas exploitation activities at night to map the target area of offshore drilling rigs in the South China Sea based on the Defense Meteorological Satellite Pro- Remote Sens. 2021, 13, 1576 3 of 26

gram/Operational Linescan System (DMSP/OLS) nighttime light images. They performed multilayer screening based on multitemporal Landsat-8 OLI (band 6) images to determine the presence of offshore drilling rigs. Liu et al. [21] adopted a multilayer screening optimization and multitemporal overlapping comparison strategy to automatically identify offshore drilling rigs in the Persian Gulf, the Gulf of Mexico and the Gulf of Thailand, with full consideration to the characteristics of spectra, texture, shape and location of offshore drilling rigs in Landsat-8 OLI (band 6). Zhao et al. [22] identiﬁed offshore drilling rigs in the South China Sea by using a multisliding window threshold and multitemporal images superposition comparison based on Landsat images. Li et al. [23] proposed a critical value algorithm for convolution operation based on phase 2 NPOESS Preparatory Project/Visible Infrared Imaging Radiometer Suite (NPP/VIIRS) imaging products. They identiﬁed offshore drilling rigs in the Northern part of the South China Sea. Sun et al. [24] conducted a long time series analysis of offshore drilling rigs in the South China Sea based on multitemporal and multisource remote sensing image data, including Landsat and SAR data. An exploratory analysis of the relationship between the total nighttime light of offshore drilling rigs and oil and gas production was established based on DMSP/OLS nighttime light images. Fan et al. [25] proposed a new automatic extraction algorithm for offshore platforms in the Bohai area based on multispectral image of the Gaofen-2 (GF-2) Remote Sens. 2021, 13, x FOR PEER REVIEWMSS sensor. As is known to all, existing detection methods of offshore platforms based on 3 of 28

optical remote sensing images often rely on a large number of time series images. In order to improve the accuracy of platform detection with single images, Zhu et al. [26] proposed a platform detection method based on a Harris detector and intensity texture feature image from a Sentinel-2 L2A Image.

Figure 1. Research framework of offshore drilling rig identiﬁcation.

Waste gases during oil and gas extraction processes are often hard to collect, and burn- Figureing 1. them Research is a routine framework way to preventof offshore environment drilling rig damage. identification In 1978,. the associated waste gas flame was first observed on DMSP/OLS night-time light image data [27]. Thereafter, StroppianaA few studies et al. [28 have] applied been a contextcarried automatic out to identify recognition offshore algorithm drilling to generate rigs based the on their locationfirst global-scale invariant spatiotemporalfeature in Synthetic distribution Aperture map of fire Radar points (SAR) based ondata. thermo-infrared These researchers de- velopedAdvanced a two Very-parameter High Resolution false Radiometeralarm algorithm (AVHRR) to image identify data. Chowdhurythe location et al.of [ 29potential] offshore drilling rigs with SAR images first, and then eliminate the false alarm with multiphase SAR data to improve the predictions of offshore drilling rigs in the South China Sea [12–14]. Cheng et al. [15] proposed a method based on the triangular invariant rule to identify offshore drilling rigs with SAR and reported improved detection accuracy of offshore drilling rigs in the South China Sea. Wan et al. [16] compared the shape differences between ships and offshore drilling rigs in SAR data to improve detection accuracy. Wong et al. [17] developed a cloud-native geoprocessing algorithm to automatically detect and extract offshore oil platforms in the Gulf of Mexico using synthetic aperture radar and the Google Earth Engine. Zhang et al. [18] proposed an automatic extraction method for offshore platforms in single SAR images based on a dual-step-modified model. Passing ships are the main source of false alarm in offshore drilling rig identification. Meng and Xing [19] attempted to map offshore drilling rigs and ships using Landsat remote sensing images taken within a single time phase, but the results showed that single time phase optical images could not distinguish the two well. Cheng et al. [20] utilized the high brightness of lighting facilities and associated high temperature of waste gas combustion flames in oil and gas exploitation activities at night to map the target area of offshore drilling rigs in the South China Sea based on the Defense Meteorological Satellite Program/Operational Linescan System (DMSP/OLS) nighttime light images. They performed multilayer screening based on multitemporal Landsat-8 OLI (band 6) images to determine the presence of offshore drilling rigs. Liu et al. [21] adopted a multilayer screening optimization and multitemporal overlapping comparison strategy to automatically identify offshore drilling rigs in the Persian Gulf, the Gulf of Mexico and the Gulf of Thai- land, with full consideration to the characteristics of spectra, texture, shape and location of offshore drilling rigs in Landsat-8 OLI (band 6). Zhao et al. [22] identified offshore drilling rigs in the South China Sea by using a multisliding window threshold and multitemporal images superposition comparison based on Landsat images. Li et al. [23] proposed

Remote Sens. 2021, 13, 1576 4 of 26

attempted to test the possibility of detecting the associated waste gas combustion sources in Alberta using Sequential Maximum Angle Convex Cone (SMACC) with the short-wave infrared of 1.6 µm and 2.2 µm Landsat-8 images. The results show that the recognition effect of daytime Landsat-8 images for detecting the associated waste gas flames is not good. Casadio et al. [30] extracted offshore drilling rigs in the North Sea area in 2008 by using the short-wave infrared band (1.6 µm) at nighttime Along Track Scanning Radiometer (ATSR), and the extraction results were highly correlated to the reported oil and gas production in the North Sea region. Anejionu et al. [31] studied the spatial and temporal distribution of oil-associated waste gas combustion sources in the Niger Delta from 2000 to 2014 with a double threshold segmentation algorithm applied to the mid-infrared bands (3.96 µm) of nighttime Moderate-resolution Imaging Spectroradiometer (MODIS). Elvidge et al. [32] used the characteristics of associated waste gas flames on DMSP/OLS false color composite images, and visually interpreted the spatial distribution of relatively stable waste gas ignition points in the world at the country scale from 1992 to 2008. Elvidge et al. [33] proposed a network function virtualization (VNF) algorithm based on the short and medium infrared bands of National Polar-orbiting Operational Environmental Satellite System Preparatory Project/Visible infrared Imaging Radiometer (NPP/VIIRS) images and the Planck fitting formula to automatically extract global thermal anomalies. Elvidge et al. [34] refined the thermal anomalies extraction VNF algorithm, extracted global waste gas combustion sources and classified them into upper, downstream and liquefied natural gas (LNG) site ignition points with high-resolution NPP/VIIRS images. However, there are some major drawbacks with the approaches mentioned above. For the algorithms based on SAR images and Landsat images, the parameters were set subjectively and required intensive computation. At the same time, multiphase data comparison also poses a new challenge to the quality and quantity of remote sensing data. Offshore drilling rig identification based on associated waste gas flames mainly uses the near-infrared and short-wave infrared bands of NPP/VIIRS, MODIS and AVHRR images, causing those without gas combustion to be largely unidentifiable. Moreover, because the spatial resolutions of the mentioned remote sensing images are generally low, it is especially difficult for multiple offshore drilling rigs gathered closely together to be accurately differentiated. We propose a new method for offshore drilling rig extraction in the Caspian Sea, aiming to overcome these drawbacks. We build our method based on an observed phe- nomenon: a drilling rig set up above the sea surface will cause a significant drop of the Normalized Difference Water Index (NDWI) calculated with satellite observations in that specific water area. We first examine the maximum, minimum and average NDWI values of seawater, offshore drilling rigs and bare land (islands) in the Landsat-7 ETM+ images based on samples collected for these three groups, and then build nominal annual cloud and cloud shadow-free image composites, based on the analysis of the influence of cloud and cloud shadow on the NDWI values of water, bare land (island) and offshore drilling rigs. Based on these cloud and cloud shadow-free NDWI composites, we design a set of classification rules to distinguish offshore drilling rigs from the other two groups. To test this method, we extract the offshore drilling rigs in 2018 in the Caspian Sea based on our method and verify the extraction accuracy. The results show that this method has good performance.

2. Study Area 2.1. Physical Geographical Background The Caspian Sea, the largest saltwater lake in the world, is located between 44.69◦~54.07◦ E and 36.58◦~47.11◦ N. It sits at the junction of Europe and Asia, bordering Kazakhstan, Turkmenistan, Iran, Azerbaijan and Russia (Figure2). The length of the lake is about 1200 km from north to south, with an average width of 320 km from east to west. The shoreline of the lake is about 7000 km in length, covering 371,000 square kilometers. The Remote Sens. 2021, 13, 1576 5 of 26

Remote Sens. 2021, 13, x FOR PEER REVIEW 6 of 28

Caspian Sea has an ecosystem similar to that of the ocean, with a thriving shipping industry, and is a “sea trail lake” geographically [35].

Figure 2. The location of Caspian Sea. Figure 2. The location of Caspian Sea.

2.2. Oil and Gas2.2. Resource Oil and Endowment Gas Resource Endowment The Caspian SeaThe region Caspian consists Sea region of four consists major geological of four major basins: geological the northern, basins: central the northern, central and southern Caspianand southern basins, Caspian and the basins, Ustel basinand the in Ustel the north. basin Territorial in the north. disputes Territorial and disputes and limited explorationlimited techniques exploration in offshore techniques areas in makeoffshore it difficult areas make to determine it difficult the to Caspiandetermine the Caspian Sea oil and gas reserves. Based on field data, onshore and offshore fields in the broad Sea oil and gas reserves. Based on field data, onshore and offshore fields in the broad Caspian Basin are thought to hold 48 billion barrels of oil and 292 trillion cubic feet of waste Caspian Basin are thought to hold 48 billion barrels of oil and 292 trillion cubic feet of gas, respectively (U.S. EIA, year) (Table1). Since the reserves figure includes both proven waste gas, respectively (U.S. EIA, year) (Table 1). Since the reserves figure includes both and probable reserves, it tends to be overestimated. Most of these reserves are offshore proven and probable reserves, it tends to be overestimated. Most of these reserves are or near the Caspian Sea coasts, especially near the north coast. The EIA estimates that offshore or near the Caspian Sea coasts, especially near the north coast. The EIA estimates 41% of the total Caspian oil and associated gas condensates (19.6 billion barrels) and 36% that 41% of the total Caspian oil and associated gas condensates (19.6 billion barrels) and of the waste gas (106 trillion cubic feet) are offshore, with most of the offshore oil in the 36% of the waste gas (106 trillion cubic feet) are offshore, with most of the offshore oil in northern Caspian and most of the offshore gas in the southern Caspian. Another 35% of oil the northern Caspian and most of the offshore gas in the southern Caspian. Another 35% (16.6 billion barrels) and 45% of waste gas (130 trillion cubic feet) can be found onshore of oil (16.6 billion barrels) and 45% of waste gas (130 trillion cubic feet) can be found on- within 100 miles of the coasts, particularly in Russia’s North Caucasus. The remaining shore within 100 miles of the coasts, particularly in Russia’s North Caucasus. The remain- 12 billion barrels of oil and 56 trillion cubic feet of waste gas are scattered farther onshore in the large Caspianing 12 Basin,billion mainly barrels in of Azerbaijan, oil and 56 trillion Kazakhstan cubic andfeet Turkmenistan.of waste gas are scattered farther onshore in the large Caspian Basin, mainly in Azerbaijan, Kazakhstan and Turkmenistan. The Caspian Sea region is one of the oldest oil-producing regions in the world and an increasingly important source of global energy production. The region is rich in oil and gas resources, both from offshore deposits in the Caspian Sea and onshore oil fields in the Caspian Basin. The legal status of the Caspian Sea region is complicated by the lack of consensus on whether it should be defined as a “sea” or a “lake”. In each case, different international laws would apply. As a result, changing legal and regulatory frameworks

Remote Sens. 2021, 13, 1576 6 of 26

Table 1. Caspian basins proved and probable reserves

Crude Oil and Lease Natural Gas (Billion Country Condensate (Billion Barrels) Cubic Feet) Azerbaijan 8.5 51 The Caspian offshore parts 6.8 46 The Land parts 1.7 5 Iran 0.5 2 The Caspian offshore parts 0.5 1 The Land parts (s) 1 Kazakhstan 31.2 104 The Caspian offshore parts 15.7 36 The Land parts 15.5 68 Russia 6.1 109 The Caspian offshore parts 1.6 14 The Land parts 4.5 95 Turkmenistan 1.9 19 The Caspian offshore parts 1.1 9 The Land parts 0.8 10 Uzbekistan (s) 7 The Caspian offshore parts 0 0 The Land parts (s) 7 The Caspian basin 48.2 292 The Caspian offshore parts 19.6 106 The Land parts 28.6 186 Note: (s) indicates that the number is extremely small relative to the statistical value. The Caspian offshore parts refer to the oil and gas ﬁelds in the waters of the Caspian Sea. The Land parts refer to the oil and gas ﬁelds in Caspian Basin but not in the waters. Data source: the U.S. Energy Information Administration (EIA) [36].

The Caspian Sea region is one of the oldest oil-producing regions in the world and an increasingly important source of global energy production. The region is rich in oil and gas resources, both from offshore deposits in the Caspian Sea and onshore oil fields in the Caspian Basin. The legal status of the Caspian Sea region is complicated by the lack of consensus on whether it should be defined as a “sea” or a “lake”. In each case, different international laws would apply. As a result, changing legal and regulatory frameworks create uncertainty for foreign companies to invest in natural resources. For example, the lack of an agreed maritime boundary between Turkmenistan, Azerbaijan and Iran hampers geological exploration in the southern Caspian Basin. The existing data related to oil and gas resources in the Caspian Sea that can be shared are mainly based on national statistical data. A timely, accurate and comprehensive grasp of the status quo, history and trends of the exploitation of oil and gas resources in the region is of great significance to the regional security of central Asia. Therefore, we chose the Caspian Sea as the study area.

3. Methodology The offshore drilling rigs in the Caspian Sea are usually small in size (generally 120 m × 70 m), and most of them have no associated gas vent burning torch equipment, such as the Shah Deniz Oil Field and the Azeri-Chirag-Guneshli (ACG) Oil field At present, there are some studies on the automatic identification of offshore drilling rigs based on remote sensing images. Centering on the features of offshore drilling rigs with unchanged spatial position, shape and size, multistage and multisource remote sensing data is adopted to extract offshore drilling rigs [20,21,23,37]. These kinds of methods show good detection performance for offshore drilling rigs on the whole, but relatively poor performance for small platforms without associated waste gas burning flames. The current study proposes a new algorithm based on NDWI composed with multitemporal Landsat-7 ETM+ images to automatically detect offshore drilling rigs. Remote Sens. 2021, 13, 1576 7 of 26

3.1. Data Sources and Data Preprocessing The remote sensing data sources used in this study mainly include Landsat-7 ETM+ data, Sentinel-2 Multispectral Imager (MSI) data, Google Earth image, and global ignition point dataset generated with the VIIRS data. Except for the ignition point data from the data set published in the existing research [34], the other data are from GEE database data (Table2 ). According to the location of study area, we chose 596 Landsat-7 ETM+ images taken from January 2018 to December 2019 and 98 Sentinel-2 MSI images talk from 2016 to 2019. The Landsat series of sensors are widely used for long-time continuous observation with high geometric accuracy (less than 30 m), even in areas with large topographic relief [38]. The Landsat-7 ETM+, one of the Landsat series, has the longest observation time (1999–Present). Although 22% of the pixels were missing because of the Scan Line Corrector (SLC) failure in 2003, the ETM+ data is also widely used in land-cover change analyses [37,39,40]. Zhang et al. [41] and Morel et al. [39] proposed an effective method to fill in the missing data. Landsat-7 ETM+ data is the main remote sensing data source used in the offshore drilling rig identification and monitoring algorithm proposed in this study. It was noted that some areas of our study area were not covered by Landsat-7 ETM+ images, and these areas were not considered for testing the proposed method. Sentinel-2 MSI data and Google Earth image data are mainly used to verify the accuracy of identifying offshore drilling rigs, considering their relatively finer spatial resolutions. In the analysis of dynamic monitoring results of offshore drilling rigs in the Caspian Sea, a Google Earth image was used for partial verification of the near shore area, and Sentinel-2 MSI images were used for visual interpretation and verification. The 2018 VIIRS ignition point data was used to identify the Caspian offshore drilling rigs and evaluate the results of the present study. The data used in this study and the Landsat-7 ETM+ images coverage of the Caspian Sea for our method are shown in Table2 and Figure3.

Table 2. Remote Sensing data used in this study.

Data Set Data Name Spatial Resolution Data Source Landsat data Landsat-7/ETM+ 30 m GEE [42] Night-light data VIIRS/DNB About 500 m NOAA [43]

Sentinel-2 MSI 10 m GEE [44] High-resolution data DigitalGlobe imagery / Google Earth Note: Google Earth Engine (USGS), National Oceanic and Atmospheric Administration (NOAA).

Considering the relatively large scale of the Caspian Sea region, most of our remote sensing data processing and analysis are carried out on the Google Earth Engine (GEE) cloud computing platform. GEE provides not only a mass remote sensing data archive, but also big data computing facilities, enabling large-scale remote sensing studies. The Landsat- 7 ETM+ data are provided by the Earth Resources Observation and Science (EROS) Center and are archived on the GEE Cloud platform. We use only Landsat 7 Collection 1 Tier 1 and Real-Time data to reduce the impact of registration errors, critical for time series synthesis. Six bands of Landsat ETM + were used in this study: blue (band 1: 0.45–0.52 µm), green (band 2: 0.53–0.61 µm), red (band 3: 0.63–0.69 µm), near-infrared (band 4: 0.78–0.90 µm) and two mid-infrared (band 5 and 7: 1.55–1.75 µm and 2.09–2.35 µm). Each Landsat-7 ETM + L1T scene is approximately 185 km × 85 km, and since May 2003, 22% pixels have been lost in the repeated stripe pattern along with the scan due to ETM+ Scan Line Corrector (SLC) failure [45]. In ETM+ images, low-quality pixels often lack full spectral bands, especially at the image edges. Before generating the NDWI composite, these pixels must be ﬁrst identiﬁed and removed. We generate a low-quality ETM + pixel mask (including gaps caused by SLC-OFF) by examining bands 1–5 and 7: If the product is equal to 0, the pixel is of low quality and is discarded [41]. Additionally, clouds and shadows seriously contaminate Remote Sens. 2021, 13, 1576 8 of 26

Remote Sens. 2021, 13, x FOR PEER REVIEW 9 of 28

Landsat images in the late stage [46]. We adopt the FMASK algorithm to remove clouds and shadows from Landsat images to reduce their inﬂuence on the NDWI index [47].

Figure 3. The total count of Landsat ETM+ images coverage of Caspian Sea in 2018 and 2019. Note: there are no data in the Figure 3. The total count of Landsat ETM+ images coverage of Caspian Sea in 2018 and 2019. Note: there are no data in white areas. the white areas. 3.2. Offshore Drilling Rig Appearance in Remote Sensing Images Considering the relatively large scale of the Caspian Sea region, most of our remote sensingAn data offshore processing drilling and rig analysis is the infrastructure are carried out for drilling,on the Google extracting, Earth processing Engine (GEE) and cloudtemporarily computing storing platform. crude oil GEE and provides natural gas.not only It is thea mass carrier remote of offshore sensing oil data and archive gas pro-, butduction also activitiesbig data [computing25]. The metal facilities, concrete enabling structure large and-scale associated remote waste sensing gas studies. incineration The Landsatcharacteristics-7 ETM+ of the data offshore are provided drilling rigsby the make Earth them Resources show high Observation reﬂectivity, in and contrast Science to the seawater background in remote sensing images [25]. However, due to the small size (EROS) Center and are archived on the GEE Cloud platform. We use only Landsat 7 Col- of offshore drilling rigs, weak image features and the vast ocean background with noise lection 1 Tier 1 and Real-Time data to reduce the impact of registration errors, critical for (cloud, cloud shadow, etc.) and false alarms (passing vessels), the accurate detection of time series synthesis. Six bands of Landsat ETM + were used in this study: blue (band 1: spatial information of offshore oil and gas production platforms from images still faces 0.45–0.52 μm), green (band 2: 0.53–0.61 μm), red (band 3: 0.63–0.69 μm), near-infrared great challenges [21]. (band 4: 0.78–0.90 μm) and two mid-infrared (band 5 and 7: 1.55–1.75 μm and 2.09–2.35 Offshore drilling rigs are mostly metal structures, and their NDWI value is much μm). Each Landsat-7 ETM+ L1T scene is approximately 185 km × 85 km, and since May smaller than that of water, which is reﬂected in the obvious difference in brightness between 2003, 22% pixels have been lost in the repeated stripe pattern along with the scan due to offshore drilling rigs and the water background in remote sensing images (Figure4d–f). ETM+ Scan Line Corrector (SLC) failure [45]. Figure4d–f shows the minimum, mean and maximum NDWIs, calculated from a series of In ETM+ images, low-quality pixels often lack full spectral bands, especially at the images taken within one year, offshore drilling rigs and the background water, respectively. image edges. Before generating the NDWI composite, these pixels must be first identified In the maximum NDWI composite image (Figure4f), the difference between the offshore anddrilling removed. rigs and We the generate background a low (water)-quality is mostETM obvious.+ pixel mask Thus, (incl the maximumuding gaps value caused NDWI by SLCcomposites-OFF) by can examining be used tobands extract 1–5 offshore and 7: If drilling the product rigs. is equal to 0, the pixel is of low quality and is discarded [41]. Additionally, clouds and shadows seriously contaminate Landsat images in the late stage [46]. We adopt the FMASK algorithm to remove clouds and shadows from Landsat images to reduce their influence on the NDWI index [47].

Remote Sens. 2021, 13, x FOR PEER REVIEW 10 of 28

3.2. Offshore Drilling Rig Appearance in Remote Sensing Images An offshore drilling rig is the infrastructure for drilling, extracting, processing and temporarily storing crude oil and natural gas. It is the carrier of offshore oil and gas production activities [25]. The metal concrete structure and associated waste gas incineration characteristics of the offshore drilling rigs make them show high reflectivity, in contrast to the seawater background in remote sensing images [25]. However, due to the small size of offshore drilling rigs, weak image features and the vast ocean background with noise (cloud, cloud shadow, etc.) and false alarms (passing vessels), the accurate detection of spatial information of offshore oil and gas production platforms from images still faces great challenges [21]. Offshore drilling rigs are mostly metal structures, and their NDWI value is much smaller than that of water, which is reflected in the obvious difference in brightness between offshore drilling rigs and the water background in remote sensing images (Figure 4d-f). Figure 4d–f shows the minimum, mean and maximum NDWIs, calculated from a series of images taken within one year, offshore drilling rigs and the background water, Remote Sens. 2021, 13, 1576 respectively. In the maximum NDWI composite image (Figure 4f), the difference between9 of 26 the offshore drilling rigs and the background (water) is most obvious. Thus, the maximum value NDWI composites can be used to extract offshore drilling rigs.

FigureFigure 4.4. Caspian offshore drilling rig: ( (aa)) is is offshore offshore drilling drilling rigs rigs of of the the ACG ACG oil oil and and gas gas fields ﬁelds in in GoogleGoogle high-resolutionhigh-resolution image; image; ( b(b,c,c)) are are sizes sizes of of offshore offshore drilling drilling rigs; rigs; (d (,de,ef),f are) are offshore offshore drilling drilling rigs inrigs different in different NDWI NDWI synthesis synthesis methods. methods.

3.3. Establishment of NDWI Classification Rule 3.3. Establishment of NDWI Classiﬁcation Rule 3.3.1. Distinguishability Analysis of Different Water Indexes on Drilling Platform, Water 3.3.1. Distinguishability Analysis of Different Water Indexes on Drilling Platform, Water andand BareBare LandLand InIn thisthis section,section, we we analyze analyze the the distinguishability distinguishability of NDWI,of NDWI NDWI-NDVI, NDWI-NDVI (Normalized (Normal- Differenceized Difference Vegetation Vegetation Index) Index) [48] and [48] Modiﬁed and Modified NDWI NDWI (MNDWI) (MNDWI) [49] to [49] drilling to drilling platform, plat- waterform, andwater bare and land. bareUsing land. Sentinel-2Using Sentinel L1C- image2 L1C image on GEE on platform, GEE platform, we randomly we randomly select 60select sample 60 sample areas of areas offshore of offshore drilling rigs,drilling water rigs, and water bare and land. bare Then, land. we Then, calculate wethe calculate mean valuethe mean of three value water of three body water indexes body for indexes each small for each area small and obtainarea and the obtain water the index water changes index ofchanges three features of three (drillingfeatures rig,(drilling water rig, and water bare and land) bare (Figure land)5). (Figure 5). For offshore drilling rigs, we can see that the changes of NDWI and MNDWI are more consistent; the NDWI value is slightly lower than MNDWI, but the ndwi-ndvi value is lower. For the water body, the values of NDWI-NDVI, NDWI and MNDWI increase in turn. For the bare land, the values of NDWI-MDVI, NDWI and MNDWI decrease in turn. It can be seen that for the drilling rigs and water body, the minimum values of the three water body indexes on the drilling platform are greater than the maximum values on the water body; for the drilling rigs and bare land, the NDWI and MNDWI values of most drilling platforms are larger than those of bare land, but the difference of NDWI-NDVI values between these drilling platforms and bare land is not obvious. In conclusion, NDWI- NDVI has the lowest discrimination, and NDWI and MNDWI have similar discrimination. Therefore, we choose NDWI as the water index for our method.

3.3.2. NDWI NDWI is the normalized difference between the green band (band 2) and near-infrared band (band 4) [50], which can enhance water in remote sensing images. NDWI is widely used in water recognition. Its formula is: ρ − ρ NDWI = Green NIR (1) ρGreen + ρNIR Remote Sens. 2021, 13, 1576 10 of 26

where ρNIR represents the reﬂectivity of the near-infrared band and ρGreen represents the Remote Sens. 2021, 13, x FOR PEER REVIEW reﬂectivity of the green band. Water bodies often show high NDWI values. As a comparison,11 of 28 the NDWI values of offshore drilling rigs are much lower, and there is a large difference between them (Figure4d–f).

Figure 5. Changes of NDWI,Figure MNDWI 5. Changes and of NDWI, NDWI MNDWI-NDVI and on NDWI-NDVI drilling platform, on drilling platform, water body water bodyand andbare bare land land..

For offshore drilling rigs, we can see that the changes of NDWI and MNDWI are more consistent; the NDWI value is slightly lower than MNDWI, but the ndwi-ndvi value is lower. For the water body, the values of NDWI-NDVI, NDWI and MNDWI increase in turn. For the bare land, the values of NDWI-MDVI, NDWI and MNDWI decrease in turn. It can be seen that for the drilling rigs and water body, the minimum values of the three water body indexes on the drilling platform are greater than the maximum values on the water body; for the drilling rigs and bare land, the NDWI and MNDWI values of most drilling platforms are larger than those of bare land, but the difference of NDWI-NDVI values between these drilling platforms and bare land is not obvious. In conclusion, NDWI-NDVI has the lowest discrimination, and NDWI and MNDWI have similar discrimination. Therefore, we choose NDWI as the water index for our method.

3.3.2. NDWI NDWI is the normalized difference between the green band (band 2) and near-infrared band (band 4) [50], which can enhance water in remote sensing images. NDWI is widely used in water recognition. Its formula is:

Remote Sens. 2021, 13, 1576 11 of 26

Since 1999, multispectral Landsat-7 ETM+ data with a spatial resolution of 30 m have been archived. However, cloud and cloud shadow pollution, as well as the lack of algorithms to analyze them, greatly hindered their use [41]. At the same time, clear images are often rare, especially in tropical areas with persistent thick cloud cover [51]. Therefore, good image pixels with useful information in some polluted images are often not well utilized. Many algorithms have been developed to automatically detect and remove clouds and associated shadows in Landsat images [52,53]. These advanced algorithms detect clouds and cloud shadows with higher and higher accuracy, but none can achieve a 100% screening effect. Consequently, residual clouds and cloud shadows, especially thin clouds and associated shadows, can still remain. To make better use of high-quality pixels in images that may be partly polluted by clouds and cloud shadows, we propose a multitemporal image compositing algorithm. Pixel-scale satellite image compositing was initially developed for Advanced Very High-Resolution Radiometer (AVHRR) time series by automatically picking up image pixels free of cloud or cloud shadow pollution. This technology aims to generate a high-quality composite image by picking up clear pixels from other periods, even if there is no single clear image in a specific period. All the pixels that are not completely polluted can be better utilized. The largest normalized difference vegetation index (NDVI) and the maximum apparent temperature method [54] are commonly used. In 2015, Zhang et al. [41] modified the maximum NDVI value compositing method to consider to the underlying earth surface (vegetated land, bare land and water) and applied different compositing criteria individually to generate a final NDVI composite image. In this study, we modify the algorithm of Landsat’s nominal annual NDVI composite proposed by Zhang et al. to generate the optimal cloud-free NDWI images based on Google Earth Engine (GEE) using Landsat-7 ETM+ data.

3.3.3. Effects of Clouds and Cloud Shadows on NDWI To remove the influence of clouds and cloud shadows on NDWI, we first examine their disturbance to NDWI values. The effect of cloud and cloud shadow on NDWI values depends largely on the underlying surface (Figure6). There are three major types of objects in the region of the Caspian Sea: water, bare land (mainly islands) and offshore drilling rigs. Among them, bare land mainly refers to the islands inside the Caspian Seawater and seasonally exposed lands along the Caspian Sea coasts. Over the water surface, clouds will reduce the NDWI value, while cloud shadow has relatively little influence on its NDWI value (Figure6a). Over the bare land surface, the NDWI value will increase to different degrees due to the influence of cloud and cloud shadow (Figure6b), and cloud thickness has different degrees of influence on the NDWI value. Compared with water and bare land, the spatial distribution of offshore drilling rigs is discontinuous. To compare the influence of cloud and cloud shadow on offshore drilling rigs, we select images with cloud (shadow) and without cloud (shadow) in the same area in order to conduct a statistical analysis on the NDWI values of different offshore drilling rigs (Figure6). In Figure6c, the blue line represents the NDWI value of the offshore drilling rigs under cloudless conditions and the orange line represents the NDWI value of the offshore drilling rigs under cloudy (shadowy) conditions. It can be seen that due to the influence of cloud and shadow, the NDWI value of offshore drilling rigs will decrease to a certain extent. Compared with the NDWI curve of water and bare land, the NDWI curve of offshore drilling rigs fluctuates greatly, mainly due to the mixed noise of water around the offshore drilling rigs. The NDWI values of a water body, bare land and offshore drilling rig pixels under different clouds and cloud shadows were calculated. The results showed that the maximum NDWI value should be used to select clear water pixels in a time series, but for bare land, the criteria of minimum NDWI value should be adopted. For offshore drilling rigs, the NDWI value is between water and bare land, and due to strip and other problems of Landsat 7 ETM+ images, the criterion of mean NDWI value is adopted (Table3). Based on these observations, we develop the optimal NDWI criteria rather than the single maximum Remote Sens. 2021, 13, x FOR PEER REVIEW 13 of 28

RemoteRemote Sens.Sens. 20212021,, 1313,, xx FORFOR PEERPEER REVIEWREVIEW 1313 ofof 2828

extent. Compared with the NDWI curve of water and bare land, the NDWI curve of off- Remote Sens. 2021, 13, x FOR PEER REVIEW shoreextent.extent. drilling ComparedCompared rigs fluctuateswithwith thethe NDWINDWI greatly, curvecurve mainly ofof waterwater due to andand the barebare mixed land,land, noise thethe 13of NDWINDWI waterof 28 curve curve around ofof off-off-the offshoreshoreshore drillingdrilling drilling rigsrigs rigs. fluctuatesfluctuates greatly,greatly, mainlymainly duedue toto thethe mixedmixed noisenoise ofof waterwater aroundaround thethe offshoreoffshore drillingdrilling rigs.rigs. Remote Sens. 2021, 13, 1576 12 of 26

extent. Compared with the NDWI curve of water and bare land, the NDWI curve of offshore drilling rigs fluctuatesNDWI criteria greatly, to describe mainly the NDWI due ofto different the mixed ground noise objects of ofwater water around in the clear the offshore drilling rigs.sky (Table3).

Figure 6. Clouds and cloud shadows and their effects on NDWI. (a): over the water surface, (b ): overFigureFigure bare 6.6. CloudsCloudsland surface, andand cloudcloud (c): over shadowsshadows offshore andand drilling theirtheir effectseffects rigs. (onon1) TrueNDWI.NDWI. color ((aa):): Landsat overover thethe 7 water waterETM+ surface,surface, image. ((b2b)):): Landsatoverover barebare 7 landlandETM surface, surface,+ NDWI (( ccimage.):): overover Theoffshoreoffshore images drillingdrilling (c) on rigs.rigs. the (left(11)) True Trueand right colorcolor sides LandsatLandsat are 7images7 ETM+ETM+ of image.image. the off- ((22)) shoreLandsatLandsat drilling 77 ETMETM rigs ++ NDWINDWI with and image.image. without TheThe imagescloudsimages (shaded). ((cc)) onon thethe leftleft andand rightright sidessides areare imagesimages ofof thethe off-off- shoreshore drillingdrilling rigsrigs witwithh andand withoutwithout cloudsclouds (shaded).(shaded). The NDWI values of a water body, bare land and offshore drilling rig pixels under differentTheThe NDWIcloudsNDWI valuesandvalues cloud ofof aa shadows waterwater body,body, were barebare calculated landland andand. T heoffshoreoffshore results drillingdrilling showed rigrig that pixelspixels the underundermaxi- mumdifferentdifferent NDWI cloudsclouds value andand should cloudcloud be shadowsshadows used to wereselectwere calculatedcalculatedclear water.. T Tpixelshehe resultsresults in a time showedshowed series, thatthat but thethe for maxi-maxi- bare landmummum, theNDWINDWI criteria valuevalue of shouldshouldminimum bebe used usedNDWI toto selectselectvalue clearclearshould waterwater be adopted. pixelspixels inin Foraa timetime offshore series,series, drilling butbut forfor rigs,barebare thelandland NDWI,, thethe criteriacriteria value ofisof betweenminimumminimum water NDWINDWI and valuevalue bare shouldshouldland, and bebe adopted.dueadopted. to strip ForFor and offshoreoffshore other drillingdrillingproblems rigs,rigs, of thethe NDWINDWI valuevalue isis betweenbetween waterwater andand barebare land,land, andand duedue toto stripstrip andand otherother problemsproblems ofof Figure 6. Clouds and cloud shadows and theirLandsat effects 7 onETM+ NDWI. images, (a): over the thecriterion water surface,of mean ( bNDWI): over v barealue land is adopted surface, (Table 3). Based on Figure 6. Clouds and cloud shadowsLandsatLandsat and 77 their ETM+ETM+ effects images,images, on thethe NDWI. criterioncriterion ( ofaof): meanmean over NDWINDWI the water vvaluealue surface, isis adoptedadopted (b (Table(Table): 3).3). BasedBased onon (c): over offshore drilling rigs. (1) True colorthese Landsat observations, 7 ETM+ image. we (2) develop Landsat 7the ETM optimal + NDWI NDWI image. criteria The images rather (c) onthan the single maxi- thesethese observations,observations, wewe developdevelop thethe optimaloptimal NDWINDWI criteriacriteria ratherrather thanthan thethe singlesingle maxi-maxi- theover left bare and right land sides surface, are images (c): of over the offshore offshoremum drilling NDWI drilling rigs criteria with rigs. andto describe ( without1) True the clouds color NDWI (shaded). Landsat of different 7 ETM+ ground image. objects ( of2) water in the clear Landsat 7 ETM + NDWI image. Theskymummum images (Table NDWINDWI 3). ( criteriacriteriac ) on the toto describedescribe left and thethe right NDWINDWI sides ofof differentdifferent are images groundground of objectsobjectsthe off- ofof waterwater inin thethe clearclear skysky (Table(Table 3).3). shore drilling rigs witTableh and 3. Thewithout impacts clouds on NDWI (shaded). from clouds and cloud shadows over different landcover types. Table 3. The impacts on NDWI from clouds and cloud shadows over different landcover types. Table 3. The impacts on NDWI from clouds and cloud shadows over different landcover types. TableGround 3. The object impacts on NDWI from clouds Cloud and Effectcloud shadows Shadow over Effectdifferent landcover Optimal NDWI types. The NDWIGround values object of a waterNDWI body, bare landCloud and E offshoreffect drillingShadow rig Effect pixels underOptimal NDWI GroundGround objectobject NDWINDWI CloudCloud EEffectffect ShadowShadow EEffectffectMaximum OptimalOptimal NDWINDWI Water WaterHigh High Maximum NDWI different clouds and cloud shadows were calculated. The results showed thatNDWI the maxi- BareWaterWater land Low High(Highnegative ) MaximumMinimumMaximum NDWINDWI Minimum mum NDWI BarevalueBare landland should Bare land be usedLowLow (( Lownegativenegativeto select (negative))) clear water pixels in a time series, but forMinimumMinimum bare NDWINDWI Between water and bare NDWI Offshore drilling rigs Between water and bare Mean NDWI land, the criteria of minimumOffshore drillingBetween NDWI waterlandBetween value and water bare should be adopted. For offshore drilling rigs, OffshoreOffshore drillingdrilling rigsrigs Mean NDWI MeanMean NDWINDWI the NDWI value is betweenrigs waterlandland andand bare bare land land, and due to strip and other problems of 3.3.4. NDWI Feature Extraction for the Classification of Different Objects Landsat 7 ETM+ images, the criterion3.3.4. NDWI of meanFeature NDWIExtraction v aluefor the is C adoptedlassification (Table of Different 3). Based Objects on 3.3.4. NDWI Feature3.3.4.Offshore NDWI Extraction F eaturedrilling for theE rigsxtraction Classification are fixed for the in oflocationClassification Different and Objects can of Dbeifferent operated Objects for a long time once these observations, weOffshore develop drillingbuilt the OffshoreOffshore(generally rigs optimal are drilling fixeddrilling around NDWI in locationrigsrigs 20 years). areare criteria andfixedfixed Although can inin be locationlocationrather operated there andandthan are for can canoften a longthe bebe many operatedoperated timesingle ships once forformaxi- builtin a athe longlong Caspian timetime once onceSea, mum NDWI criteria(generally to describe aroundthebuiltbuilt their 20 (generallypositions(generally years).NDWI Although are aroundaroundof generallydifferent 20 there20 years).years). not are ground oftenfixed AlthoughAlthough manyover objects thereathere shipsperiod areare inof oftenof theoften water time. Caspian manymany The in shipsships Sea,optimalthe their inclearin thethe NDWI Caspian Caspian compo- Sea,Sea, sky (Table 3). positions are generallysitingthetheirir positions positionsmethod not fixed proposed areare overgenerallygenerallya in period this notnot paper offixedfixed time. based overover The aaon period optimalperiod Landsat ofof NDWI time.time.images TheThe compositing for optimaloptimal two consecutive NDWINDWI compo-compo- years method proposedcansitingsiting inalso this methodmethod effectively paper proposedproposed based avoid on inin Landsatfalse thisthis alarmpaperpaper imagess based basedcaused for onon two by LandsatLandsat consecutiveships. Inimagesimages the years analysis forfor can twotwo alsoof consecutiveconsecutive the influen yearsyearsce of effectively avoidcloudcancan false alsoalso and alarms effectivelyeffectively cloud caused shadow avoidavoid by ships.on falsefalse the alarm Inalarmwater, thess analysis barecausedcaused land ofbyby and the ships.ships. influenceoffshore InIn thethe drilling ofanalysisanalysis cloud rigs andofof thethein Section influeninfluen 3.cece3 .3ofof, Table 3. The impacts on NDWI fromcloud clouds shadow anditcloudcloud oncloud can the be andand water, shadowsseen cloudcloud that bare shadowshadow the landover influence and ondifferenton thethe offshore ofwater,water, clouds landcover drilling barebare and landland rigscloud andtypesand in shadows Section offshoreoffshore. 3.3.3 ondrillingdrilling the, it canthree rigsrigs be intypesin SectionSection of objects 3.3.33.3.3,, seen that the influenceisitit cancandifferent. bebe of seenseen clouds To thatthat eliminate and thethecloud influenceinfluence or shadowsminimize ofof cloudsclouds on the the andandimpact three cloudcloud types of cloudsshadowsshadows of objects and onon cloud is thethe different. threethree shadows typestypes on ofof objectsNDWIobjects Ground object NDWI To eliminateCloud ortoisis minimize different.different.composeEffect the TotheTo impact eliminate eliminateoptimal ofShadow clouds NDWI oror minimizeminimize and images, cloudEffect thethe different shadows impactimpact ofofground on Optimalcloudsclouds NDWI object andand to compose cloud cloudNDWItypes shadowsshadowsin the the Cas ononpian NDWINDWI Sea Water High optimal NDWIsurfacetoto images, composecompose should different thethe first optimaloptimal ground be distinguished. NDWINDWI object typesimages,images, inThen, the differentdifferent Caspianthe optimal groundgroundMaximum Sea surfaceNDWI objectobject should should types typesNDWI bein firstin theselectedthe CasCaspianpian accord- SeaSea be distinguished.surfacesurface Then, shouldshould the optimal firstfirst bebe NDWI distinguished.distinguished. should be Then,Then, selected thethe optimaloptimal according NDWINDWI to the shouldshould strategies bebe selectedselected accord-accord- Bare land Low (negative) ing to the strategies listed in Table 2 for different groundMinimum objects NDWI based upon the NDWI listed in Table2timeing ingfor differenttoto series thethe strategies strategiesof ground each pixel. listed objectslisted Finally, inin based TableTable this upon 22 NDWIforfor the differentdifferent NDWI image timeground groundis segmented series objectsobjects of each with basedbased pixel. threshold uponupon thethe values NDWINDWI to Between water and bareFinally, this NDWItimetime image seriesseries of isof segmentedeacheach pixel.pixel. Finally, withFinally, threshold thisthis NDWINDWI values imageimage to identify isis segmentedsegmented offshore withwith drilling tthresholdhreshold valuesvalues toto Offshore drilling rigs identify offshore drilling rigs. To find the NDWI featuresMean that distinguish NDWI water, bare land land rigs. To find the(islands)identifyidentify NDWI offshore offshoreand features offshore drillingdrilling that drilling distinguish rigs.rigs. ToTo rigs, findfind water, we thethe used NDWINDWI bare 10 land featuresmfeatures resolution (islands) thatthat distinguish distinguish andSentinel offshore-2 MSI water,water, images barebare landlandand drilling rigs, we(islands)(islands) used 10 andand m resolution offshoreoffshore drillingdrilling Sentinel-2 rigs,rigs, MSI wewe images usedused 1010 and mm Googleresolutionresolution high-resolution SentinelSentinel--22 MSIMSI imagesimages andand images to vectorize the boundary of water, offshore drilling rigs and bare land (islands) samples as the area of interest in the north, central and south Caspian Sea. Then, the areas 3.3.4. NDWI Feature Extraction for the Classification of Different Objects of interest of the three types of ground objects are superimposed onto the maximum NDWI, Offshore drillingminimum rigs NDWIare fixed and meanin location NDWI composite and can images be operated to calculate thefor distributiona long time range once of built (generally aroundtheir NDWI 20 years). values. Although In this sample there selection, are thereoften are many 100,000 ships water samples,in the Caspian 1050 offshore Sea, their positions are drillinggenerally rig samples not fixed and 10,000 over bare a period land (islands) of time. samples. The Figure optimal7 shows NDWI the histogram compositing method proposed in this paper based on Landsat images for two consecutive years can also effectively avoid false alarms caused by ships. In the analysis of the influence of cloud and cloud shadow on the water, bare land and offshore drilling rigs in Section 3.3.3, it can be seen that the influence of clouds and cloud shadows on the three types of objects is different. To eliminate or minimize the impact of clouds and cloud shadows on NDWI to compose the optimal NDWI images, different ground object types in the Caspian Sea surface should first be distinguished. Then, the optimal NDWI should be selected according to the strategies listed in Table 2 for different ground objects based upon the NDWI time series of each pixel. Finally, this NDWI image is segmented with threshold values to identify offshore drilling rigs. To find the NDWI features that distinguish water, bare land (islands) and offshore drilling rigs, we used 10 m resolution Sentinel-2 MSI images and

Remote Sens. 2021, 13, x FOR PEER REVIEW 14 of 28

Google high-resolution images to vectorize the boundary of water, offshore drilling rigs and bare land (islands) samples as the area of interest in the north, central and south Cas- pian Sea. Then, the areas of interest of the three types of ground objects are superimposed onto the maximum NDWI, minimum NDWI and mean NDWI composite images to cal-

Remote Sens.culate2021, 13, 1576 the distribution range of their NDWI values. In this sample selection, there 13 are of 26 100,000 water samples, 1050 offshore drilling rig samples and 10,000 bare land (islands) samples. Figure 7 shows the histogram statistical distribution of the NDWI composite images of the maximum,statistical minimum distribution and of mean the NDWI of various composite ground images objects of the maximum, samples minimumin the wa- and ters of the Caspianmean Sea. of various ground objects samples in the waters of the Caspian Sea.

FigureFigure 7. Distributions7. Distributions of maximum, of maximum, minimum, minimum, mean, NDWI mean, values NDWI of bare values land, water of bare and offshoreland, water drilling and rig. offshore drilling rig. From the statistical distribution based on the maximum, minimum and mean NDWI From the statisticalimages of distribution water, offshore based drilling on rigs the and maximum, bare land (islands)minimum samples and (Figuremean NDWI7a–c), we images of water,found offshore that waterdrilling can rigs be easily and separatedbare land from (islands) bare land samples and offshore (Figure drilling 7a– rigc), inwe the maximum NDWI statistics, Its maximum NDWI value (Max_NDWI) is usually greater than found that water0.55, can among be easily which separated the maximum from NDWI bare value land of bare and land offshore (islands) drill is lessing than rig 0.2, in and the the maximum NDWINDWI statistics, value ofIts most maximum offshore drilling NDWI rigs value is between (Max_NDWI) the water and is bareusually land (greaterFigure7a ). than 0.55, amongTherefore, which the we maximum can draw the NDWI line at thevalue maximum of bare NDWI land value(islands) of 0.55 is toless separate than 0.2, water and the NDWI valuefrom the of othermost two offshore types. The drilling minimum rigs NDWI is between value of morethe water than 70% and pixels bare of offshoreland drilling rigs and water is equal to 0 (Figure7b), while the minimum NDWI value of bare (Figure 7a). Therefore,land (island) we can is usually draw the less thanline −at0.05 the (Figure maximum7d). Therefore, NDWI usingvalue a of minimum 0.55 to NDWIsep- arate water from valuethe other of −0.05, two we types. can separate The minimum bare land fromNDWI water value and offshoreof more drilling than 70% rigs. Similarly,pixels of offshore drillingthe meanrigs and NDWI water value is of offshoreequal to drilling 0 (Fig rigsure (Mean_NDWI) 7b), while the is between minimum 0–0.4; thisNDWI can be value of bare landused (island) to distinguish is usually offshore less drilling than − rigs0.05 from (Fig bareure land7d). (islands)Therefore, (Figure using7c). Ita ismini- worth mum NDWI valuenoting of − that0.05, in Figurewe can7a, theseparate maximum bare NDWI land value from of somewater pixels and of offshore offshore drilling drilling rigs is greater than 0.55, mainly due to the mixing of water body pixels in the vectoring process rigs. Similarly, theof offshoremean NDWI drilling value rig samples. of offshore In Figure drilling7b, the valuerigs of(Mean_NDW water and offshoreI) is between drilling rig 0–0.4; this can be samplesused to in distinguish the minimum offshore NDWI statistics drilling is equalrigs from to 0, which bare island possibly (islands) due to (Fig the qualityure 7c). It is worth notingreduced that ETM+ in Figure image pixels 7a, the caused maximum by clouds, NDWI shadows, value strips, of etc.some Such pixels pixels of must off- be shore drilling rigsmasked is greater out for than further 0.55, analysis. mainly In Figuredue to7c, the the meanmixing NDWI of water statistics body of water pixels samples in the vectoring process of offshore drilling rig samples. In Figure 7b, the value of water and

Remote Sens. 2021, 13, 1576 14 of 26

are relatively low (0.2–0.4). The main reason is that the existence of cloud, shadow, strip and other problems leads to partial pixels were masked (NDWI = 0), which reduces the value range of the mean NDWI of water. From the statistical results of sample pixels, using any single one of maximum, minimum, or mean NDWI would not make a good distinction between water, bare land (islands) and offshore drilling rigs. However, if we combine different NDWI features (maximum, minimum and mean), we can effectively distinguish water, bare land (islands) and offshore drilling rigs. Therefore, we design a set of rules based on these statistics to depict water bodies (Max_NDWI > 0.55), bare land (islands) (Min_NDWI < −0.05) and offshore drilling rigs (0 < Mean_NDWI < 0.4) to classify the three types (Figure7e).

3.4. Offshore Drilling Rig Extraction Based on Optimal NDWI Composite 3.4.1. Preliminary Extraction of Offshore Drilling Rigs In Section 3.3.3, through the statistical analysis of sample pixels, we build a set of rules to effectively distinguish water, bare land (islands) and offshore drilling rigs based on NDWI time series (Maximum, Minimum, Mean). The NDWI time-series features proposed in Section 3.3.3 effectively reduce if not completely remove residual clouds and cloud shadows and other disturbing factors on the NDWI value. Because the optimal NDWI composite image reflects the true NDWI value of the water, bare land (islands) and offshore drilling rigs, the threshold segmentation can then be used to identify the offshore drilling rigs. False alarms such as passing ships and clouds show similar reflectivity characteristics to offshore drilling rigs, which constitutes a major difficulty in identifying of offshore drilling rigs. To better identify offshore drilling rigs, the optimal NDWI composite images were constructed using Landsat-7 ETM+ images for two consecutive years in this paper. The general process is shown in Figure8. First, the Landsat-7 ETM+ images of the Caspian Sea for two consecutive years were preprocessed to remove most clouds and shadows, and the NDWI was calculated. The map algebra statistics are then performed based on the maximum, minimum and mean NDWI of pixels. Finally, the optimal composite NDWI image is generated according to the following rules: water body (Max_NDWI > 0.55), bare land (island) (Min_NDWI < −0.05) and offshore drilling rig (0 < mean_NDWI < 0.4). The optimal NDWI compositing strategy is based on this fact: On the one hand, once the offshore drilling rig is built, it will be used for a long time, i.e., generally, about 20 years, and its location and size will remain unchanged. However, passing ships do not stay at a fixed position for a long time, so the optimal NDWI composite method can easily avoid the influence of moving ships. On the other hand, with two consecutive years of Landsat images, we can obtain cloud-free or cloud-less composite images to the greatest extent. The construction of an offshore drilling rig can then be completed within two years at most (acquired through long-term image observation). Therefore, the optimal NDWI composite data established in this paper include all images from one year and the following year. Adopting this strategy, we will avoid ships and reduce the effects of clouds, cloud shadows, etc. Due to a large amount of calculation for the synthesis of optimal composite NDWI images, we implement our algorithm on the GEE cloud platform. After the optimal NDWI composite is synthesized according to the process shown in Figure8, the offshore drilling rig can be identified with a threshold segmentation method. Image threshold segmentation is one of the most commonly used image segmentation methods. This study assumes that the optimal composite NDWI image is F (x, y), which is classified into the water, bare land and offshore drilling rigs by setting the corresponding threshold value. The classification image is g (x, y) and the specific formula is as follows:

 < −  Bare land , NDWI 0.05 g(x, y) = offshore drilling rig , 0 < NDWI < 0.4 (2)  Water , NDWI > 0.55 Remote Sens. 2021, 13, x FOR PEER REVIEW 16 of 28 Remote Sens. 2021, 13, 1576 15 of 26

Figure 8. The procedure to generate NDWI composites from multiyear Landsat-7/ETM + imagery. Figure 8. The procedure to generate NDWI composites from multiyear Landsat-7/ETM + imagery.

3.4.2.After Postprocessing the optimal NDWI of Offshore composite Drilling is synthesized Rig Detection according Results to the process shown in FigureThe 8, the Caspian offshore Sea drilling covers rig acan wide be identified area (about with 370,000 a threshold square segmentation kilometers), method. with more Imageislands threshold and sandbars segmentation along the is coast,one of especiallythe most commonly in the shallow used water image area segmentation in the northern methods.part. However, This study the assumes NDWI value that the of theoptimal interface composite between NDWI land image edges, is islands, F (x, y), orwhich sandbars isand classified water into is similar the water, to that bare of land offshore and offshore drilling drilling rigs, which rigs by tends setting to bethe another correspond- source of ingfalse threshold alarms. value. Furthermore, The classification there are image many is offshoreg (x, y) and drilling the specific rigs in formula the Caspian is as fol- Sea and lows:their structures are complex. The offshore drilling rigs with close distances extracted by the threshold segmentation may be different Bare land parts ， ofNDWI one< platform. −0.05 Therefore, postprocessing of the extracted results푔(푥 is, 푦 required) = {offshore to improve drillingthe rig detection， 0 < NDWI accuracy,< 0.4 and it mainly includes(2) two parts: (1) Generating a land maskWater to eliminate ， NDWI the> interference0.55 of islands, sandbars and near-shore areas from the extraction results; (2) Reﬁning and combining offshore drilling 3.4.2.rigs Postprocessing with similar distances. of Offshore The Drilling land mask Rig D isetection generated Results with a spatial buffer toward the centerThe of Caspian the Caspian Sea covers Sea based a wide on thearea Caspian (about 370,000 Sea vectorized square kilometers), boundary data. with While more most islandsoffshore and drilling sandbars rigs along are the located coast, 4 especially km away in from the shallow the boundary water area line in of the the northern Caspian Sea part.(based However, on visual the NDWI interpretation value of withthe interface high-resolution between land images edges, on islands, Google or Earth), sandbars some of andthem water are relativelyis similar to closer that toof theoffshore islands drilling and sandbars rigs, which in thetends water. to be If another the buffer source distance of of islands and sandbars is set too large, offshore drilling rigs could be excluded. Therefore, the buffer distance of the Caspian Sea boundary is set as 3.5 km, while the buffer distance of islands and sandbars is set as 60 m.

Remote Sens.Remote 2021, 13Sens., x FOR2021 ,PEER 13, x FORREVIEW PEER REVIEW 17 of 28 17 of 28 Remote Sens.Remote 2021, 13Sens., x FOR 2021 PEER, 13, x REVIEWFOR PEER REVIEW 17 of 28 17 of 28

false alarms.false Furthermore, alarms. Furthermore, there are there many are offshore many drillingoffshore rigs drilling in the rigs Caspian in the SeaCaspian and Sea and falsetheir alarms.structuresfalsetheir Furthermore, alarms.structures are complex. Furthermore, are there complex. The are offshore theremany The are drillingoffshoreoffshore many rigs drillingdrillingoffshore with rigs rigs closedrilling within distances the rigsclose Caspian in distances extractedthe SeaCaspian andextracted by Sea and by theirthe threshold structurestheirthe threshold segmentationstructures are complex. segmentation are may complex.The beoffshore differentmay The bedrillingoffshore differentparts rigs drillingof onepartswith platform. rigs closeof one with distances platform. Therefore,close distances extracted Therefore, postpro- extracted by postpro- by thecessing threshold ofthecessing the threshold extractedsegmentation of the extracted segmentationresults may is resultsrequired be different may is requiredto be improve differentparts to of theimprove oneparts detection platform. of the one detection accuracy, platform. Therefore, accuracy, and Therefore, itpostpro- mainly and itpostpro- mainly cessingincludes of cessingincludestwo the extractedparts: of two the (1) extracted parts: resultsGenerating (1) is resultsrequiredGenerating a land is torequired mask improve a land to to eliminatemask theimprove detection to eliminate the detection interferenceaccuracy, the interferenceaccuracy, and of it islands, mainly and of it islands, mainly Remote Sens. 2021, 13, 1576 includessandbars includessandbars twoand parts: near-shore twoand (1) parts:near-shoreGenerating areas (1) fromGenerating areasa landthe fromextraction mask a landthe to extraction eliminate maskresults; to (2)eliminate results;the Refining interference (2) the Refiningand interference16 ofcombiningof 26islands, and combiningof islands, sandbarsoffshore sandbarsoffshoredrillingand near-shore rigs drillingand with near-shore areasrigs similar with from areasdistances. similar the fromextraction distances. The the landextraction results; The mask land (2) results;is Refininggeneratedmask (2) is Refininggeneratedand with combining a spatialand with combining a spatial offshorebuffer toward drillingoffshorebuffer the toward rigs centerdrilling with the of rigs centerthesimilar Caspianwith of distances. thesimilar SeaCaspian based distances. The Sea on land basedthe The maskCaspian on land theis generated SeamaskCaspian vectorized is generatedSea with vectorized boundary a spatial with boundary a spatial bufferdata. While towardbufferdata. most theWhile toward centeroffshore most the of thedrillingcenteroffshore Caspian of rigs drillingthe Sea areCaspian basedlocated rigs Seaare on 4 basedthelocated km Caspian away on 4 the km from Sea Caspian away vectorizedthe fromboundary Sea vectorizedthe boundary boundary line of boundary line of Based on field photos,data.the Caspian While high-resolutiondata.the mostSeaCaspian While (based offshore mostSea on (based remotedrillingoffshore visual on interpretationrigs drilling sensingvisual are located interpretationrigs images, are with 4 locatedkm high-resolution EIA,away with 4 km from andhigh-resolution away the their images fromboundary shape, the on images boundary Googleline ofon Google line of theEarth), Caspian sometheEarth), Sea Caspianof them some(based areSeaof on them relatively(based visual are on interpretationrelatively closer visual to interpretation closerthe islandswith to thehigh-resolution and islandswith sandbars high-resolution and sandbars inimages the water. on inimages Googlethe If water.the on Google If the structure and combinationEarth),buffer characteristics,distancesomeEarth),buffer of ofdistance themsome islands are of of them relativelyand theislands sandbarsare offshore relatively andcloser sandbars is toset drilling thecloser too islands islarge, toset rigsthe too offshoreand islands large, in sandbars the drilling offshoreand Caspian sandbarsin rigs thedrilling water.could Seain rigsthe beIf thewater.couldex- beIf theex- can be grouped into threebuffercluded. categories: distance Therefore,buffercluded. ofdistance Therefore, islands the single buffer of and islands offshorethe distance sandbars buffer and ofdistance drilling sandbarsis the set Caspian too of rigs,is large,the set SeaCaspian too largeoffshore boundary large, Sea offshore drillingoffshore boundary is set rigs as drilling 3.5 iscould setkm, rigs as bewhile 3.5 couldex- km, bewhile ex- rig groups and artificialcluded.the islands buffer Therefore,cluded.the distance formedbuffer Therefore, the ofdistance islands buffer by offshore theof distanceand islands buffer sandbars drilling ofdistanceand the sandbars is Caspian set of rigs as the 60 is SeaCaspian (Table m.set boundary as 60 4Sea ).m. Asboundary is set shown as 3.5 is setkm, in as while 3.5 km, while Table4, the large offshorethe drillingbufferBasedthe distance on buffer rigBased field group ofdistancephotos, on islands field is high-resolutionof connectedphotos, and islands sandbars high-resolution and by sandbars isremote multipleset as sensing60 isremote m.set offshore as images, sensing60 m. drillingEIA, images, and EIA,their rigs and shape, their shape, Based on Basedfield photos, on field high-resolution photos, high-resolution remote sensing remote images, sensing EIA, images, and EIA,their and shape, their shape, through “corridors”. However,structurestructure and due combination toand the combination narrowcharacteristics, characteristics, width the of offshore the the “corridor” drillingoffshore rigs drilling in (about the rigs Caspian in 9 the m), SeaCaspian can Sea can structurebe grouped structurebeand groupedinto combination three and into combinationcategories: three characteristics, categories: single characteristics, offshore the single offshore drillingoffshore the drillingoffshore rigs, drilling rigs largedrilling inrigs, offshorethe rigslarge Caspian in drillingoffshorethe SeaCaspian can rig drilling Sea canrig only the platforms canbegroups be grouped successfully andbegroups groupedartificialinto andthree islandsartificial identified.into categories: three formed islands categories: single The by formed offshore offshore artificial single by drillingoffshore drillingoffshore island rigs drilling rigs, drilling (Table is large equipped rigs rigs,4). offshore (Table As large shown 4). with drillingoffshore As in shown Table rig drilling in Table rig corresponding auxiliarygroups4, facilities the largeandgroups4, theartificial offshoreand large and the islands artificial offshoredrilling basic formed equipmentislands rig drilling group by formed rigoffshore is group connectedfor by oildrillingoffshore is production connected by rigs drilling multiple (Table by rigs (gas), multiple4). offshore (Table As andshown 4). offshoredrilling theAs in shown Table rigs drilling in Table rigs identification results are4,through often the large “4,through broken.corridors the offshore large “corridors” Therefore,. offshoredrillingHowever,” . rigHowever, drilling due we group establishto rigthe isdue group connectednarrow to thethe is width connectednarrow following by multipleof widththe by “ rules corridor multiple offshoreof the for “”corridor offshorelarge(aboutdrilling ”9 rigs(about drillingm), 9 rigsm), offshore drilling rig groupsthroughonly the and “throughonlyplatformscorridors artificial the “platforms corridors”can. However, island be successfully ”can. formedHowever, duebe successfullyto the byidentified. due offshorenarrow to theidentified. Thewidth narrow drilling artificial of The widththe rigs artificialisland“corridor of detected the is islandequipped“”corridor (about inis equipped” 9with (aboutm), 9with m), onlycorresponding the onlycorrespondingplatforms the auxiliary platforms can be auxiliaryfacilities successfully can be facilitiesand successfully the identified. basicand the equipmentidentified. The basic artificial equipment Thefor oil islandartificial production for is oil islandequipped production (gas), is equipped with and (gas), with and remote sensing images,correspondingthe to identification bettercorrespondingthe identification determine auxiliary results auxiliary facilitiesare theresults often type areandfacilitiesbroken. andoften the basicTherefore, numberandbroken. theequipment basicTherefore, ofwe offshoreestablishequipment for we oil establishtheproduction drilling for following oil theproduction rigs: (gas),following rules and for (gas), rules and for (1) For large offshore drillingthelarge identification offshore rigs,thelarge identification vectorizeddrillingoffshore results rig drilling are groups results “corridors”often rig and arebroken. groups oftenartificial Therefore,andwill broken. islandartificial be Therefore, built. weformed island establish A by 60weformed offshore m establishthe buffer followingby drillingoffshore the zone following rules rigs drilling de-for rules rigs de-for will be built on both sideslargetected with offshore in largetectedremote the drillingoffshore vectorizedin sensing remote rig drilling groups images,sensing corridor rig and togroupsimages, betterartificial as and thetodetermine betterislandartificial center. determineformed the island Offshore type by formed theoffshoreand drillingtype number by drillingandoffshore of rigsnumber offshorerigs drilling de- of offshore rigs de- in the buffer zone will betecteddrilling grouped in rigs:(1) tecteddrillingremote into Forin rigs:(1)sensing remote large the largeFor offshore images,sensing large offshore todrillingoffshoreimages, better drillingrigs, drillingtodetermine better vectorized rigs, rigdetermine the group,vectorized “typecorridors theand while “type numbercorridors” will offshoreand be of number”built. willoffshore Abe 6 ofbuilt.0 offshore A 60 drilling rigs:(1)drillingm buffer For rigs:(1) zone large will Foroffshore belarge built drillingoffshore on both rigs, drilling sides vectorized withrigs, thevectorized “ corridorsvectorized “corridors” willcorridor be ”built. willas the Abe center.6 built.0 AOff- 60 drilling rigs outside them buffer buffer zone will will be bebuilt counted on both sides as single with the offshore vectorized drilling corridor rigs. as the (2) center. For Off- mshore buffer drilling mshorezone buffer rigs willdrilling zone inbe thebuilt rigswill buffer on inbe the bothbuilt zone buffer sides on will both zone withbe sidesgrouped willthe vectorizedwithbe groupedinto the thevectorized corridor largeinto theoffshore ascorridor large the center.drillingoffshore as the Off- rig center.drilling Off- rig the artificial island of offshoreshoregroup, drilling whileshoregroup, drilling offshorerigs drilling while in rig, the drillingoffshorerigs each buffer in polygontherigs drillingzone buffer outside will rigs zone of be the outside thegrouped willbuffer artificial be the zonegroupedinto buffer thewill island largezone intobe counted isthewilloffshore identified large be ascounted drillingoffshore single as off-rig drilling single off-rig to build a 150 m buffergroup,shore zone drilling while andgroup,shore thenoffshore rigs. drilling while fused(2) Fordrilling offshorerigs. the and (2) artificialrigs Fordrilling counted outsidethe island artificialrigs asthe outside of onebuffer islandoffshore offshorethe zone of buffer drillingoffshore will drillingzone be rig, drilling counted will each rig.be rig,polygon ascounted eachsingle ofpolygon as off-the single of off-the shoreartificial drilling shoreislandartificial rigs. drilling is islandidentified (2) For rigs. is the identified (2)to artificial buildFor the ato island150artificial build m ofbuffera 150islandoffshore mzone bufferof drilling offshoreand zone then rig, drilling andfused each then andpolygonrig, fused countedeach and ofpolygon the ascounted of the as artificial islandartificial is identifiedisland is identified to build ato 150 build m buffera 150 mzone buffer and zone then andfused then and fused counted and ascounted as Table 4. Caspian offshoreone drilling offshoreone rig drillingoffshore types rig. drilling and image rig. characteristics (the red border in remote one offshoreone drillingoffshore rig. drilling rig. sensingTable images4. CaspianTable is4. offshore theCaspian preliminary drilling offshore rig drilling extractiontypes andrig types image result). and characteristics image characteristics (the red border (the red in borderremote insensing remote images sensing is imagesthe is the Tablepreliminary 4. CaspianTablepreliminary extraction 4. offshore Caspian extractionresult). drilling offshore result). rig drilling types andrig typesimage and characteristics image characteristics (the red border (the red in borderremote insensing remote images sensing is imagesthe is the preliminaryPlatformpreliminary extraction Type extractionresult). result). Remote Sensing Image Related Images PlatformPlatform Type Type Remote SRemoteensing ISmageensing Image Related Images Platform PlatformType Type Remote SRemoteensing ISmageensing Image Related IRelatedmages Images

Single offshoreSingle drillingoffshore drilling Single offshore drilling rig Single offshoreSinglerig drillingoffshorerig drilling rig rig

Large offshoreLarge drillingoffshore rig drilling rig LargeLarge offshoreLargegroup offshore drillingoffshore group drilling rig drilling rig rig group groupgroup Remote Sens.Remote 2021, 13Sens., x FOR2021 ,PEER 13, x FORREVIEW PEER REVIEW 18 of 28 18 of 28

ArtificialArtiﬁcial Artificialisland island of islandoff- of offshore of offshore drillingdrillingshore rig drilling rig rig

4. Results4. and Results Discussion and Discussion 4. Results and Discussion4.1. Identification4.1. Identification Accuracy AccuracyAnalysis Analysis 4.1. Identification Accuracy AnalysisTo exploreTo the explore effect the of theeffect optimal of the NDWI optimal algorithm NDWI algorithm on offshore on drillingoffshore rig drilling identifi- rig identifi- To explore the effectcation, of thewecation, evaluate optimal we evaluatethe NDWI accuracy the algorithm accuracyof offshore of ondrillingoffshore offshore rig drilling identification drilling rig identification rig in the identifi- Caspian in the SeaCaspian Sea waters inwaters 2018. Becausein 2018. thereBecause is no there ground is no truth ground data truth and datafield and verification field verification is difficult is todifficult to cation, we evaluate thecarry accuracy out,carry we ofadopt out, offshore we the adopt following drilling the following verification rig identification verification strategy. strategy.It includes in the It Caspianincludespartial verification partial Sea verification in in waters in 2018. Becausenear-shore therenear-shore is areas no ground with areas Google truthwith Earth Google data images, andEarth fieldvisualimages, verificationinterpretation visual interpretation isof difficultSentinel-2 of Sentinel-2 toMSI and MSI and carry out, we adopt theself-consistency followingself-consistency verification verification verification of strategy.Sentinel-2 of Sentinel-2 MSI It includes (mainly MSI for (mainly partial relatively for verification relatively remote and remote small in and plat- small plat- near-shore areas withforms Google offshore).forms Earth offshore). For images, this purpose,For visual this purpose,a total interpretation of a7405 total potential of 7405 of potentialSentinel-2targets are targets extracted MSI are and extracted based on based on the Landsat-7the Landsat-7 ETM+ optimal ETM+ NDWIoptimal composite. NDWI composite. From these, From 6883 these, targets 6883 of targets the extracted of the extracted self-consistency verification of Sentinel-2 MSI (mainly for relatively remote and small targets aretargets excluded are excluded using the using buffer the mask buffer of land,mask islandsof land, and islands sandbanks. and sandbanks. Finally, 522 Finally, 522 platforms offshore). Forpotential this purpose, potentialoffshore drillingoffshore a total rigs drilling of are 7405 confirmed rigs potential are confirmed by refining targets by therefining are processing extracted the processing rules based of offshore rules of offshore on the Landsat-7 ETM+drilling optimal rigs.drilling NDWI After rigs. verificationcomposite. After verification with From high-resolution these,with high-resolution 6883 images, targets the images, of finalthe extracted the algorithm final algorithm deter- deter- mines 497mines offshore 497 drillingoffshore rigs, drilling including rigs, including 486 single 486 platforms, single platforms, 3 large offshore 3 large drillingoffshore drilling rig groupsrig and groups 8 artificial and 8 islandartificial platforms island platforms (Figure 9 ).(Figu re 9).

Figure 9. OffshoreFigure 9. drillingOffshore rig drilling extraction rig extractionresults, figures results, (a )figures, (b), and (a )(,c ()b are), and respectively (c) are respectively enlarged areasenlarged (a), (areasb), and (a )(,c ()b in), and (c) in the left image.the left image. Remote Sens. 2021, 13, x FOR PEER REVIEW 18 of 28

Artificial island of offshore drilling rig

4. Results and Discussion 4.1. Identification Accuracy Analysis To explore the effect of the optimal NDWI algorithm on offshore drilling rig identification, we evaluate the accuracy of offshore drilling rig identification in the Caspian Sea waters in 2018. Because there is no ground truth data and field verification is difficult to carry out, we adopt the following verification strategy. It includes partial verification in near-shore areas with Google Earth images, visual interpretation of Sentinel-2 MSI and Remote Sens. 2021, 13, 1576 self-consistency verification of Sentinel-2 MSI (mainly for relatively remote and17 small of 26 platforms offshore). For this purpose, a total of 7405 potential targets are extracted based on the Landsat-7 ETM+ optimal NDWI composite. From these, 6883 targets of the extracted targets are excluded using the buffer mask of land, islands and sandbanks. Finally, 522 targets are excluded using the buffer mask of land, islands and sandbanks. Finally, 522 po- potentialtential offshoreoffshore drilling drilling rigs rigs are are confirmed confirmed by by refining refining the the processing processing rules rules of offshore of offshore drillingdrilling rigs. rigs. After After verification verification with with high-resolution high-resolution images, images, the final the algorithm final algorithm determines deter- mines497 offshore497 offshore drilling drilling rigs, including rigs, including 486 single 486 platforms, single 3platforms, large offshore 3 large drilling offshore rig groups drilling rigand groups 8 artificial and 8 island artificial platforms island (Figure platforms9). (Figure 9).

FigureFigure 9. Offshore 9. Offshore drilling drilling rig rig extraction extraction results results,, figuresfigures ( a(),a) (,b (),b and), and (c) ( arec) are respectively respectively enlarged enlarged areas ( aareas), (b), ( anda), ( bc)) in, and the (c) in the leftleft image. image. We further check the extracted offshore drilling rigs through Google Earth’s high- resolution images (Figure 10). Careful visual inspection reveals that the 3.5 km buffer distance of the Caspian Sea and the 60 m buffer distance of the islands preclude a large number of false alarms, such as those of docks, berthing vessels, islands or sandbars at the water junction (brown dots in the image). But at the same time, some offshore drilling rigs (13) off the coast of Azerbaijan have been excluded, because they are too close to shorelines. Based on Google Earth images, in the water area beyond the buffer distance, there are 248 offshore drilling rigs (including two offshore drilling rig groups), 238 of which were extracted correctly, while 10 were undetected and 23 were misidentified. The 23 misidentified targets are all small and narrow pieces of land in the water (generally no more than 2 pixels in width, i.e., 60 m). Remote Sens. 2021, 13, x FOR PEER REVIEW 19 of 28

We further check the extracted offshore drilling rigs through Google Earth’s high- resolution images (Figure 10). Careful visual inspection reveals that the 3.5 km buffer distance of the Caspian Sea and the 60 m buffer distance of the islands preclude a large number of false alarms, such as those of docks, berthing vessels, islands or sandbars at the water junction (brown dots in the image). But at the same time, some offshore drilling rigs (13) off the coast of Azerbaijan have been excluded, because they are too close to shorelines. Based on Google Earth images, in the water area beyond the buffer distance, there are 248 offshore drilling rigs (including two offshore drilling rig groups), 238 of which Remote Sens. 2021, 13, 1576 were extracted correctly, while 10 were undetected and 23 were misidentified. The18 23 of mis- 26 identified targets are all small and narrow pieces of land in the water (generally no more than 2 pixels in width, i.e., 60 m).

FigureFigure 10. 10.Google Google Earth Earth image image veriﬁcation verification results. results.

Sentinel-2Sentinel-2 MSI MSI images images are are mainly mainly used used for for visual visual interpretation interpretation and and self-consistency self-consistency verificationverification (Figure (Figure 11 11).). All All offshore offshore drilling drilling rigs rigs not not confirmed confirmed by by Google Google Earth Earth images images areare loaded loaded onto onto the the high-quality high-quality Sentinel-2 Sentinel-2 true-color true-color images images (spatial (spatial resolution resolution 10 10 m). m). ThroughThrough careful careful visual visual verification, verification, a a total total of of 265 265 offshore offshore drilling drilling rigs rigs were were found, found, among among whichwhich 259 259 offshore offshore drilling drilling rigs rigs (including (including one one offshore offshore drilling drilling rig rig group) group) were were correctly correctly extractedextracted by by the the proposed proposed algorithm, algorithm, while while six offshoresix offshore drilling drilling rigs wererigs were omitted omitted and two and weretwo misidentified.were misidentified. To summarize, there are 526 offshore drilling rigs in the Caspian Sea covered by the Landsat images, among which 497 offshore drilling rigs were automatically identified by the proposed algorithm, while 29 offshore drilling rigs were omitted (including 13 offshore drilling rigs being masked out) and 25 were misidentified (Table5). The accuracy rate was 90.2%, the false-negative rate (including the buffer mask platforms) was 5.3%, and the misidentification rate was 4.5%.

4.2. Comparison with Other Methods 4.2.1. Comparison with the Method Based on SAR Images Wong et al. [17] extracted the offshore drilling rigs using Sentinel-1 SAR data with GEE. The method may be divided into ﬁve steps: (1) Generating median composite image; (2) Calculating the difference between median composite image and mean ﬁlter image; (3) The difference image is segmented by static or dynamic threshold; (4) Eroding and dilating the segmentation result; and (5) The center of the connected area is extracted as the location of the drilling platform. We applied their method to the test area. The test area, as shown in Figure 12a, contains a large offshore drilling rig group and a number of scattered individual offshore drilling rigs. Therefore, it is suitable for testing the performance of the algorithm. RemoteRemote Sens. Sens.2021 2021, 13, 13, 1576, x FOR PEER REVIEW 1920 of of 26 28

FigureFigure 11. 11.Sentinel-2 Sentinel-2 veriﬁcation verification results. results.

Table 5.ToIdentification summarize, results there ofare offshore 526 offshore drilling drilling rigs. rigs in the Caspian Sea covered by the Landsat images, among which 497 offshore drilling rigs were automatically identified by Verification Visual Inter- Successfully False Identi- the proposed algorithm, while 29 offshore drillingMissed rigs were omitted (includingAccuracy 13 (%) off- Image pretation Identified fication shore drilling rigs being masked out) and 25 were misidentified (Table 5). The accuracy rateGoogle was Earth 90.2%, the 248false-negative 238rate (including 23 the buffer mask 23 platforms) was 5.3%, andSentinel-2 the misidentification 265 rate was 2594.5%. 6 2 90.2% Total 513 497 29 25 Table 5. Identification results of offshore drilling rigs. The results are shown in Figure 12b. We can see that due to the erosion and dilation Verification Visual Inter- Successfully False Identifi- operation for the preliminary extraction results,Missed the interconnected platformsAccuracy are inter- (%) ruptedImage and identifiedpretation as a single drillingIdentified rig. We adopt different strategiescation to postprocess theGoogle preliminary Earth extraction248 results and238 set buffers to merge23 adjacent23 platforms, so that the numberSentinel of platforms-2 is265 more accurate.259 Sensitivity to parameters6 is the2 biggest disadvantage90.2% of Wong’sTotal method. The513 size of filtering497 window and29 the shape and25 size of morphological operation structure elements will affect the results. In [17], 460 parameter combinations were tested and the optimal parameter combination was selected. However, the portability of the optimal parameter combination in other regions is uncertain.

Remote Sens. 2021, 13, 1576 20 of 26 Remote Sens. 2021, 13, x FOR PEER REVIEW 22 of 28

Figure 12.12. The results of our method comparedcompared withwith thethe otherother twotwo methods. methods. ( a(a)) is is a a color color composite composite image image of of the the Sentinel-2 Sentinel- MSI2 MSI image image of of the the test test area. area. (b ()b is) is the the identification identification result result of of TSRS TSRS method. method. (c ()c) is is the the identification identification result result based based on on Sentinel-1 Sentinel- SAR1 SAR image. image (.d ()d is) is the the identification identification result result of of our our method. method.

The first step of the TSRS method is to detect the candidate targets of the offshore 4.2.2.drilling Comparison rigs from individual with the Method images. Based In order on Optical to achieve Remote this Sensingpurpose, Data the TRSR method uses Liuthe Order et al. [ 21Statistic] extracted Filtering offshore (OSF) drilling method rigs to estimate using Landsat the sea time-series background, and and multire- adap- finementtively sets strategies. the threshold Specifically, according they to the extracted OSF value the seaof the surface sliding targets window according and the to con- the relativelystant value high of noise. reflectance It is worth compared noting to thethat surrounding the percentage water. value They of then OSF adopted and noise an values image time-seriesare set subjectively. strategy The to eliminate second step moving of the vesselTSRS method to get potential is to segment candidate the offshore targets drilling with a simplerigs from overlay the time analysis series by images. using pairwiseIn this step, comparison, the binary according images of to the the detection position-invariant results of principleall individual of offshore images drilling are accumulated. rigs. Finally, The a compound fixed threshold refinement, segmentation in which for all the pairwiseaccumu- comparisonlated image binarywill be imagesapplied were to obtain accumulated the final to detection calculate results. the occurrence Because frequencyTSRS method and setuses the more double than threshold 20 years of to multi extendsource the position- data, in addition and size-invariant to detecting principle the drilling to a platform, temporal dimension,it also monitors was usedthe existence to maximize status the of exclusion the drilling of random platform. residual Our purpose clouds. Inis to this propose method, a thedrilling final platform determination detection of offshore algorithm drilling based rigs on requires two consecutive a pairwise years comparison of Landsat between images images,without and paying each comparison attention to requires its dynamic mathematical changes. morphologyTherefore, we operation. only use Thentwo consecutive the double thresholdyears of Landsat and area data variation in the modified amplitude area need when to betesting manually TSRS set.method. These Figure steps 12 increasec presents the computation and subjectivity of the method, such as the setting of the structural elements the results obtained using the TSRS method. Compared with our method (Figure 12d), it of mathematical morphology operation, double thresholds and the fluctuation range of can be seen that in the middle part of the test area, the TSRS method does not extract the offshore drilling rig area. As a comparison, our proposed algorithm is simple in operation, whole platform, and the extraction effect of the connecting channel of the drilling platform and the discriminant rules of offshore drilling rig are based on the statistical analysis of group is not as good as in our method. The biggest disadvantage of the TSRS method is samples, which is less subjective. its sensitivity to various parameter settings; there are many artificial parameters. In [37], Liu et al. [37] proposed the method to assess the platform state based on the time-series the author conducted experiments on parameter setting and selected the optimal param- remote sensing (TSRS). This method includes three steps: the first two are to identify the eters. However, once the study area changes, the suitability of this parameter combination location of the offshore drilling rigs, and the last is to determine their current state. In cannot be determined. The result of TRSR method obtained in the test area may also be contrast to our detection method, it does not involve the assessment of the status of the rig, obtained through experiments with different parameters. The optimal NDWI synthesis and does not take into account the different data sources used. We opted to use the first

Remote Sens. 2021, 13, 1576 21 of 26

two steps of the TSRS method to detect the offshore drilling rigs in the experimental area in order to perform a comparison with our method. The first step of the TSRS method is to detect the candidate targets of the offshore drilling rigs from individual images. In order to achieve this purpose, the TRSR method uses the Order Statistic Filtering (OSF) method to estimate the sea background, and adap- tively sets the threshold according to the OSF value of the sliding window and the constant value of noise. It is worth noting that the percentage value of OSF and noise values are set subjectively. The second step of the TSRS method is to segment the offshore drilling rigs from the time series images. In this step, the binary images of the detection results of all individual images are accumulated. The fixed threshold segmentation for the accumulated image will be applied to obtain the final detection results. Because TSRS method uses more than 20 years of multisource data, in addition to detecting the drilling platform, it also monitors the existence status of the drilling platform. Our purpose is to propose a drilling platform detection algorithm based on two consecutive years of Landsat images without paying attention to its dynamic changes. Therefore, we only use two consecutive years of Landsat data in the modified area when testing TSRS method. Figure 12c presents the results obtained using the TSRS method. Compared with our method (Figure 12d), it can be seen that in the middle part of the test area, the TSRS method does not extract the whole platform, and the extraction effect of the connecting channel of the drilling platform group is not as good as in our method. The biggest disadvantage of the TSRS method is its sensitivity to various parameter settings; there are many artificial parameters. In [37], the author conducted experiments on parameter setting and selected the optimal parameters. However, once the study area changes, the suitability of this parameter combination cannot be determined. The result of TRSR method obtained in the test area may also be obtained through experiments with different parameters. The optimal NDWI synthesis required by our method comes from the statistical analysis of samples, and does require a subjective selection of parameters. Elvidge [34] detected global combustion flames of associated waste gas (VIIRS ignition point data) during the oil and gas extraction and treatment processes with the VIIRS VNF algorithm and generated a set of global VIIRS ignition point data set. The VIIRS ignition point data in 2018 showed only a total of 19 associated waste gas ignition points in the Caspian waters. Sentinel-2 image verification found that 17 of the 19 associated waste gas ignition points were offshore drilling rigs with associated waste gas combustion, while two associated waste gas ignition points were misidentified (Figure 13a,b). In terms of the number of offshore drilling rigs in Caspian waters, the 17 offshore drilling rigs identified by the VIIRS VNF algorithm were also successfully identified by our proposed algorithm (Figure 13). Moreover, our algorithm can identify offshore drilling rigs more effectively (497 vs. 17) and is not limited to those with associated waste gas combustion flames. The position accuracy of offshore drilling rigs extracted by our proposed algorithm is also higher than that of the VIIRS ignition point data. When the distance of multiple offshore drilling rigs is relatively close, the positional accuracy of the VNF algorithm is much poorer. Furthermore, by comparing VIIRS ignition point data of different years in Caspian Waters from 2012 to 2016, 2017 and 2018, it is also found that the location accuracy needs to be further optimized (Figure 13r,s).

4.3. Missed and False Identification of Offshore Drilling Rigs A total of 29 offshore drilling rigs are found to be missing by Google Earth and Sentinel-2 image verification, among which 13 were masked out by the buffer zones near the Caspian Sea coasts, while the other 14 were mainly due to their small-size, i.e., length or width generally less than a single 30 m by 30 m image pixel (Figure 14a). The remaining two offshore drilling rigs cannot be examined due to the low spatial resolution of Sentinel-2 images. This shows that the wrongly identified offshore drilling rigs are mainly either small narrow islands or small exposed islands with shallow near shore water in the shape of a circle. For the narrow small islands, the width is generally nearly or less than a pixel Remote Sens. 2021, 13, x FOR PEER REVIEW 23 of 28

required by our method comes from the statistical analysis of samples, and does require a subjective selection of parameters. Elvidge [34] detected global combustion flames of associated waste gas (VIIRS ignition point data) during the oil and gas extraction and treatment processes with the VIIRS VNF algorithm and generated a set of global VIIRS ignition point data set. The VIIRS ignition point data in 2018 showed only a total of 19 associated waste gas ignition points in the Caspian waters. Sentinel-2 image verification found that 17 of the 19 associated waste gas ignition points were offshore drilling rigs with associated waste gas combustion, while two associated waste gas ignition points were misidentified (Figure 13a,b). In terms of the number of offshore drilling rigs in Caspian waters, the 17 offshore drilling rigs iden- Remote Sens. 2021, 13, 1576 tified by the VIIRS VNF algorithm were also successfully identified by our proposed al- 22 of 26 gorithm (Figure 13). Moreover, our algorithm can identify offshore drilling rigs more effectively (497 vs. 17) and is not limited to those with associated waste gas combustion flames. The position accuracy of offshore drilling rigs extracted by our proposed algorithm is also higher than that of the VIIRS ignition point data. When the distance of mul- sizetiple (30offshore m), drilling and the rigs length is relatively is generally close, the positional larger than accuracy a pixel of the size. VNF They algorithm are mainly distributed inis much the west poorer. of Furthermore, the Caspian by Sea comparing near the VIIRS coast ignition (Figure point 14datab). of To different solve years these problems, remote sensingin Caspian imagery Waters from with 2012 a to much 2016, ﬁner2017 and spatial 2018, it resolution is also found is that required. the location accuracy needs to be further optimized (Figure 13r,s).

Remote Sens. 2021, 13, x FOR PEER REVIEW 24 of 28

Figure 13. A comparison of results from different algorithms on extracting offshore rigs in the Caspian Sea. (a–r) are the comparison between the recognition results of VNF algorithm and those of our algorithm, it can be seen that VNF algorithm has a large number of error recognition.

4.3. Missed and False Identification of Offshore Drilling Rigs A total of 29 offshore drilling rigs are found to be missing by Google Earth and Sen- tinel-2 image verification, among which 13 were masked out by the buffer zones near the Caspian Sea coasts, while the other 14 were mainly due to their small-size, i.e., length or width generally less than a single 30 m by 30 m image pixel (Figure 14a). The remaining two offshore drilling rigs cannot be examined due to the low spatial resolution of Sentinel- 2 images. This shows that the wrongly identified offshore drilling rigs are mainly either small narrow islands or small exposed islands with shallow near shore water in the shape Figureof 13. a circle.A comparison For the narrow of resultssmall islands, from the different width is algorithmsgenerally nearly on or extracting less than a offshore pixel rigs in the Caspiansize Sea. (30( m),a–r )and are the the length comparison is generally between larger than the a recognition pixel size. They results are mainly of VNF distributed algorithm and those of our algorithm,in the west it of can the be Caspian seenthat Sea near VNF the algorithm coast (Figure has 14 ab). large To solve number these ofproblems, error recognition. remote sensing imagery with a much finer spatial resolution is required.

Figure 14. Validation of undetectedFigure platforms 14. Validation and misidentiﬁed of undetected platforms platforms. and misidentified (a) are the platforms omitted. ( offshorea) are the omitted drilling rigs of our offshore drilling rigs of our method, (b) are the misidentified offshore drilling rigs of our method. method, (b) are the misidentiﬁed offshore drilling rigs of our method. 5. Conclusions The world is in an era of global climate change; at the same time, we are witnessing a rapidly growing process of globalization. However, in recent years, the globalization process has arrived at a crossroads, with increasing uncertainties in many areas. The global oil market is one of those areas; it has dramatically affected the world crude oil market, which has been on something of a roller coaster ride in recent years. There is a pressing need to understand with accuracy global fossil fuel production worldwide. Re- mote sensing data have been an effective resource for monitoring oil and gas extraction activities. But existing methods for the detection of offshore drilling rigs from optical remote sensing images are still limited, in terms of either their automation or spatial details. The availability of cloud computing facilities calls for high-efficiency algorithms to extract

Remote Sens. 2021, 13, 1576 23 of 26

5. Conclusions The world is in an era of global climate change; at the same time, we are witnessing a rapidly growing process of globalization. However, in recent years, the globalization process has arrived at a crossroads, with increasing uncertainties in many areas. The global oil market is one of those areas; it has dramatically affected the world crude oil market, which has been on something of a roller coaster ride in recent years. There is a pressing need to understand with accuracy global fossil fuel production worldwide. Remote sensing data have been an effective resource for monitoring oil and gas extraction activities. But existing methods for the detection of offshore drilling rigs from optical remote sensing images are still limited, in terms of either their automation or spatial details. The availability of cloud computing facilities calls for high-efficiency algorithms to extract information from fine resolution optical remote sensing images obtained by an increasing number of satellite sensors. This study proposed a new automatic method to synthesize nominal annual optimal NDWI images from multitemporal Landsat-7 ETM+ data to distinguish offshore drilling rigs from the water background. The advantages of this algorithm are as follows: 1. A NDWI characteristics statistical analysis was carried out on the main disturbance ground object (water and bare ground) in order to identify offshore drilling rigs against the background of ocean or water, and a set of rules was established to effectively distinguish three objects to depict water (Max_NDWI > 0.55), bare ground (Min_NDWI < −0.05) and offshore drilling rig (0 < Mean_NDWI < 0.4). These rules can not only effectively distinguish water, bare land (islands) and offshore drilling rigs, but can also effectively select clean pixels from images partially polluted by clouds and cloud shadows to generate high quality NDWI composites. These high quality NDWI composites form the basis of a method to identify offshore drilling rigs with an overall accuracy reaching 90.2%. 2. The optimal NDWI compositing process considers images that were taken over two consecutive years, successfully excluding passing ships, clouds and cloud shadows, and other moving objects. 3. The algorithm uses free ETM+ images to facilitate the monitoring of long time series. The optimal NDWI compositing rules are set based on a statistical analysis of the sample pixels in the region of interest, avoiding human subjectivity. Meanwhile, the algorithm is simple and easy to implement, and the GEE platform provides powerful computation. Furthermore, a spatial resolution of 30 m can effectively avoid missed identification with a coarse spatial resolution of night-light data. It also overcomes the defect that night-light ignition point data cannot identify offshore drilling rigs without a waste gas flame. 4. One current limitation is the robustness of the method, which needs to be further confirmed. The algorithm was only used in the Caspian Sea with good water quality; future work should focus on the large-scale research of other sea areas such as the Gulf of Mexico or the Persian Gulf. Another limitation is that Landsat-7 ETM+ images are only used for static identification of oil and gas platforms in the Caspian region. The next step is to use multisource remote sensing images, such as Sentinel-2 imagery, Landsat-8 OLI imagery and SAR imagery, to conduct more comprehensive research of offshore oil and gas platforms so as to improve the recognition accuracy and time length in order to obtain temporal and spatial dynamic attribute information and establish a more complete global oil and gas platform information management system. Lastly, it is rather difficult to determine a distance for near-shore areas. An unsuitable buffer distance can miss rigs or make false identifications. As such, the method yielded results that were highly accurate in open waters but relatively inaccurate near the shore. The composite NDWI synthesis method proposed in this paper can be used to identify and dynamically monitor offshore drilling rigs within a large water surface area relatively quickly, which is of great significance for exploring the exploitation status of offshore oil Remote Sens. 2021, 13, 1576 24 of 26

and gas resources. It can be extended to finer spatial resolution optical remote sensing images; thus, small-sized drilling rigs can also be effectively detected. Of course, there are still some issues with this algorithm. For example, the threshold segmentation of the optimal composite NDWI image for offshore drilling rig identification will simultaneously identify some islands, sandbars, parts of the boundary between coastal areas and water, and scattered exposed islands around them. Furthermore, an offshore drilling rig may be identified as multiple discrete rigs, which requires postprocessing of preliminary results, increasing the workload. We only tested this algorithm in the Caspian Sea, and it needs further validation whether in order to be applied in other regions.

Author Contributions: Conceptualization, H.Z., G.J. and Q.Z.; methodology, H.Z. and Q.Z.; software, H.Z. and G.J.; validation, H.Z., G.J. and S.Z.; formal analysis, H.Z., G.J. and X.L.; data curation, H.Z.; writing—original draft preparation, G.J.; writing—review and editing, G.J. and Q.Z.; visualization, H.Z. and G.J.; supervision, Q.Z.; project administration, Q.Z. and Y.S.; funding acquisition, Q.Z. and Y.S. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by the National Key Research and Development Program of China (Grant No. 2017YFB0504204); the One Hundred Talents Program of the Chinese Academy of Sciences (2015, Grant No. Y674141001); and supported in part by the National Natural Science Foundation of China (General Program, Grant No. 42071351), the National Key Research and Development Program of China (Grant No. 2020YFA0608501), the Liaoning Revitalization Talents Program (Grant No. XLYC1802027) and the One Hundred Talents Program of Chinese Academy of Sciences (Grant No. Y938091). Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Acknowledgments: The authors would like to thank the Google Earth Engine for providing the computing platform and integrating free satellite imagery data and the editors and reviewers for their comments and suggestions. Conﬂicts of Interest: The authors declare no conﬂict of interest.

References 1. Jiang, W.; Zhou, W.; Jia, H. Potential of Global Offshore Petroleum Resource Exploration and Utilization Prospect. Nat. Gas. Geosci. 2010, 21, 988–995. 2. Sun, C.; Liu, Y.; Zhao, S.S.; Song, J. Estimating offshore oil production using DMSP-OLS annual composites. ISPRS J. Photogramm. 2020, 165, 152–171. [CrossRef] 3. BP. Energy Outlook 2019 Edition. 2019. Available online: https://www.bp.com/content/dam/bp/business-sites/en/global/ corporate/pdfs/energy-economics/energy-outlook/bp-energy-outlook-2019.pdf (accessed on 13 December 2020). 4. EIA. International Energy Outlook 2018. 2018. Available online: https://www.eia.gov/pressroom/presentations/capuano_0724 2018.pdf (accessed on 19 November 2019). 5. World Bank Database. Available online: https://databank.worldbank.org/home.aspx (accessed on 13 December 2020). 6. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens. Environ. 2017, 202, 18–27. [CrossRef] 7. Liu, Y.; Cui, H.Y.; Kuang, Z.; Li, G.Q. Ship detection and classiﬁcation on optical remote sensing images using deep learning. ITM Web Conf. EDP Sci. 2017, 12, 05012. [CrossRef] 8. Gallego, A.J.; Pertusa, A.; Gil, P. Automatic Ship Classiﬁcation from Optical Aerial Images with Convolutional Neural Networks. Remote Sens. 2018, 10, 511. [CrossRef] 9. Gao, L.; He, Y.; Sun, X.; Jia, X. Incorporating Negative Sample Training for Ship Detection Based on Deep Learning. Sensors 2019, 19, 684. [CrossRef][PubMed] 10. Huang, G.A.; Wan, Z.A.; Liu, X.A.; Hui, J.B. Ship detection based on squeeze excitation skip-connection path networks for optical remote sensing images. Neurocomputing 2019, 332, 215–223. [CrossRef] 11. Falqueto, L.E.; Sá, J.A.; Paes, R.L.; Passaro, A. Oil rig recognition using convolutional neural network on Sentinel-1 SAR images. IEEE Geosci. Remote Sens. Lett. 2019, 16, 1329–1333. [CrossRef] 12. Peng, C.; Wang, J.; Li, D. Oil platform investigation by multi-temporal SAR remote sensing image. In Proceedings of the Conference on SAR Image Analysis, Modeling, and Techniques XI, Prague, Czech Republic, 21–22 September 2011. 13. Wang, J.; Liu, Y.; Li, M.; Yang, K.; Cheng, L. Drilling platform detection based on ENVISAT ASAR remote sensing data: A case of southeastern Vietnam offshore area. Geogr. Res. 2013, 32, 2143–2152. Remote Sens. 2021, 13, 1576 25 of 26

14. Wang, Q.; Zhang, J.; Su, F. Offshore Platform Extraction Using RadarSat-2 SAR Imagery: A Two-Parameter CFAR Method Based on Maximum Entropy. Entropy 2019, 21, 556. [CrossRef] 15. Cheng, L.; Yang, K.; Tong, L.; Liu, Y.; Li, M. Invariant triangle-based stationary oil platform detection from multitemporal synthetic aperture radar data. J. Appl. Remote Sens. 2013, 7, 073537. [CrossRef] 16. Wan, J.; Yao, P.; Meng, J. Research on Detection Method of the Offshore Oil Platform Based on SAR Images. Bull. Surv. Map. 2014, 1, 56–59. 17. Wong, B.A.; Thomas, C.; Halpin, P. Automating offshore infrastructure extractions using synthetic aperture radar & Google Earth Engine. Remote Sens. Environ. 2019, 233, 111–412. 18. Zhang, J.; Wang, Q.; Su, F. Automatic extraction of offshore platforms in single SAR images based on a dual-step-modified model. Sensors 2019, 19, 231. [CrossRef][PubMed] 19. Meng, R.; Xing, Q. Detection of offshore ship and well platform based on optical remote sensing images. J. Comput. Appl. 2013, 33, 708–711. [CrossRef] 20. Cheng, W. Research on Extraction of Oil and Gas Drilling Platforms with Remote Sensing in the South China Sea. Master’s Thesis, Nanjing University, Nanjing, China, 2015. 21. Liu, Y.X.; Sun, C.; Yang, Y.; Zhou, M.; Zhan, W.; Cheng, W. Automatic extraction of offshore platforms using time-series Landsat-8 Operational Land Imager data. Remote Sens. Environ. 2016, 175, 73–91. [CrossRef] 22. Zhao, S.; Sun, C.; Wang, H.; Cheng, W. Extraction and Monitoring of Offshore Oil and Gas Platforms Based on Landsat Imagery. Trop. Geogr. 2017, 37, 112–119. 23. Li, Q. Oil and Gas Platforms Detection and Oil and Gas Resources Security Situation Evaluate Based on Multi-Source Data in the South China Sea. Master’s Thesis, Lanzhou Jiaotong University, Lanzhou, China, 2016. 24. Sun, C. Dynamic Monitoring of Oil/Gas Development in the South China Sea Based on Long-Period Time-Serious and Multi- Source Remote Sensing Image. Ph.D. Thesis, Nanjing University, Nanjing, China, 2018. 25. Fan, Y.; Wang, Q. A Connected Component Based Offshore Platforms Extraction Method of GF2 Multi-channel Optical Images. In Proceedings of the 2019 IEEE 4th International Conference on Signal and Image Processing (ICSIP), Wuxi, China, 19–21 July 2019; IEEE: New York, NY, USA, 2019; pp. 916–919. 26. Zhu, Z.; Tang, Y.; Han, T. Offshore Platform Detection Based on Harris Detector and Intensity-Texture Image from Sentinel_2A Image. In Proceedings of the 2018 10th IAPR Workshop on Pattern Recognition in Remote Sensing (PRRS), Beijing, China, 19–20 August 2018; IEEE: New York, NY, USA, 2018; pp. 1–4. 27. Croft, T.A. Nighttime images of the earth from space. Sci. Am. 1978, 239, 86–101. [CrossRef] 28. Stroppiana, D.; Pinnock, S.; Gregoire, J.M. The Global Fire Product: Daily fire occurrence from April 1992 to December1993 derived from NOAA AVHRR data. Int. J. Remote Sens. 2010, 21, 1279–1288. [CrossRef] 29. Chowdhury, S.; Shipman, T.; Chao, D.; Elvidge, C.D.; Zhizhin, M.; Hsu, F.C. Daytime Gas Flare Detection Using Landsat-8 Multispectral Data. In Proceedings of the IEEE Joint International Geoscience and Remote Sensing Symposium (IGARSS), Quebec City, QC, Canada, 13–18 July 2014. 30. Casadio, S.; Arino, O.; Minchella, A. Use of ATSR and SAR measurements for the monitoring and characterisation of night-time gas flaring from offshore platforms: The North Sea test case. Remote Sens. Environ. 2012, 123, 175–186. [CrossRef] 31. Anejionu, O.C.D.; Blackburn, G.A.; Whyatt, J.D. Detecting gas flares and estimating flaring volumes at individual flow stations using MODIS data. Remote Sens. Environ. 2015, 158, 81–94. [CrossRef] 32. Elvidge, C.D.; Ziskin, D.; Baugh, K.E.; Tuttle, B.T.; Ghosh, T.; Pack, D.W.; Erwin, E.H.; Zhizhin, M. A Fifteen Year Record of Global Natural Gas Flaring Derived from Satellite Data. Energies 2009, 2, 595–622. [CrossRef] 33. Elvidge, C.D.; Zhizhin, M.; Hsu, F.C.; Baugh, K.E. VIIRS Nightfire: Satellite Pyrometry at Night. Remote Sens. 2013, 5, 4423–4449. [CrossRef] 34. Elvidge, C.D.; Zhizhin, M.; Baugh, K.; Hsu, F.C.; Ghosh, T. Methods for Global Survey of Natural Gas Flaring from Visible Infrared Imaging Radiometer Suite Data. Energies 2016, 9, 14. [CrossRef] 35. Hu, R.; Jiang, F.; Wang, Y.; Li, J.; Li, Y.; Abdimijit, A.; Luo, G.; Zhang, J. Arid Ecological and Geographical Conditions in Five Countries of Central Asia. Arid Zone Res. 2014, 31, 1–12. 36. EIA. Caspian Sea Region. Available online: https://www.eia.gov/international/content/analysis/regions_of_interest/Caspian_ Sea/caspian_sea.pdf (accessed on 23 August 2013). 37. Liu, Y.X.; Hu, C.M.; Sun, C.; Zhan, W.F.; Sun, S.J.; Xu, B.H.; Dong, Y.Z. Assessment of offshore oil/gas platform status in the northern Gulf of Mexico using multi-source satellite time-series images. Remote Sens. Environ. 2018, 208, 63–81. [CrossRef] 38. Lee, D.S.; Storey, J.C.; Choate, M.J.; Hayes, R.W. Four years of Landsat-7 on-orbit geometric calibration and performance. IEEE Trans. Geosci. Remote Sens. 2004, 42, 2786–2795. [CrossRef] 39. Morel, A.C.; Fisher, J.B.; Malhi, Y. Evaluating the potential to monitor aboveground biomass in forest and oil palm in Sabah, Malaysia, for 2000–2008 with Landsat ETM+ and ALOS-PALSAR. Int. J. Remote Sens. 2012, 33, 3614–3639. [CrossRef] 40. Nooni, I.K.; Duker, A.A.; Van Duren, I.; Addae-Wireko, L.; Osei Jnr, E.M. Support vector machine to map oil palm in a heterogeneous environment. Int. J. Remote Sens. 2014, 35, 4778–4794. [CrossRef] 41. Zhang, Q.L.; Li, B.; Thau, D.; Moore, R. Building a Better Urban Picture: Combining Day and Night Remote Sensing Imagery. Remote Sens. 2015, 7, 11887–11913. [CrossRef] Remote Sens. 2021, 13, 1576 26 of 26

42. GEE. Landsat Collections. Available online: https://developers.google.com/earth-engine/datasets/catalog/landsat-7 (accessed on 27 December 2019). 43. National Oceanic and Atmospheric Administration (NOAA). Available online: https://www.ngdc.noaa.gov/eog/viirs/ download_dnb_composites.html (accessed on 2 February 2020). 44. GEE. Sentinel Collections. Available online: https://developers.google.com/earth-engine/datasets/catalog/sentinel-2 (accessed on 23 November 2019). 45. Markham, B.L.; Storey, J.C.; Williams, D.L.; Irons, J.R. Landsat sensor performance: History and current status. IEEE Trans. Geosci. Remote Sens. 2004, 42, 2691–2694. [CrossRef] 46. Zhu, Z.; Wang, S.X.; Woodcock, C.E. Improvement and expansion of the Fmask algorithm: Cloud, cloud shadow, and snow detection for Landsats 4-7, 8, and Sentinel 2 images. Remote Sens. Environ. 2015, 159, 269–277. [CrossRef] 47. Zhu, Z.; Woodcock, C.E. Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sens. Environ. 2012, 118, 83–94. [CrossRef] 48. Lu, S.L.; Wu, B.F.; Yan, N.N.; Wang, H. Water body mapping method with HJ-1 A/B satellite imagery. Int. J. Appl. Earth Obs. 2011, 13, 428–434. [CrossRef] 49. Xu, H.Q. A Study on Information Extraction of Water Body with the Modiﬁed Normalized Difference Water Index (MNDWI). J. Remote Sens. 2005, 9, 589–595. 50. Gao, B.C. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 1996, 58, 257–266. [CrossRef] 51. Ju, J.; Roy, D.P. The availability of cloud-free Landsat ETM plus data over the conterminous United States and globally. Remote Sens. Environ. 2008, 112, 1196–1211. [CrossRef] 52. Wei, J.; Huang, W.; Li, Z.; Sun, L.; Zhu, X.; Yuan, Q.; Liu, L.; Cribb, M. Cloud detection for Landsat imagery by combining the random forest and superpixels extracted via energy-driven sampling segmentation approaches. Remote Sens. Environ. 2020, 248, 112005. [CrossRef] 53. Candra, D.S.; Phinn, S.; Scarth, P. Cloud and cloud shadow masking for Sentinel-2 using multitemporal images in global area. Int. J. Remote Sens. 2020, 41, 2877–2904. [CrossRef] 54. Holben, B.N. Characteristics of Maximum-Value Composite Images from Temporal Avhrr Data. Int. J. Remote Sens. 1986, 7, 1417–1434. [CrossRef]