Detection of Forest Cover Change in Carrascal, Surigao Del Sur Using Multi-Temporal Satellite Imagery: Comparing Pixel-Based and Object Based Approach
Total Page:16
File Type:pdf, Size:1020Kb
DETECTION OF FOREST COVER CHANGE IN CARRASCAL, SURIGAO DEL SUR USING MULTI-TEMPORAL SATELLITE IMAGERY: COMPARING PIXEL-BASED AND OBJECT BASED APPROACH Amira Janine T. Mag-usara, Michelle V. Japitana, and James Earl D. Cubillas Phil-LiDAR 2.B.14 Project, Caraga State University, Ampayon, Butuan City, Philippines E-mail addresses: [email protected] and [email protected] ABSTRACT: Together with the constant population increase is the widespread technological and industrial development which contributes to the environmental degradation and deterioration. Needless to say, Philippines is one of these nations that is facing serious biological disturbance and environmental damage. Thus, continuous monitoring of land cover change is important to establish links between policy decision making and subsequent land use activities. Carrascal is a municipality that belongs to Surigao del Sur, Philippines with rich nickel deposits, making its vegetative regions vulnerable to depletion. Therefore, the objective of this study is to identify the alterations on its forest areas from the year 2010 to 2014 using Object Based Image Analysis (OBIA) through eCognition with support vector machine (SVM) on Matlab to optimize the separating parameters for each class, to compare the methods’ accuracy against pixel-based approach using Envi 4.3 and to detect the changes made by mining activities. Landsat 7 ETM+ and Landsat 8 OLI for the year 2010 and 2014 were used respectively. Accuracy assessment was done using groundtruth points and an existing 2010 land cover map. From the derived land covers, results showed that forest areas had decreased by 14.46% where 1,514.62 hectares of forest areas on 2010 were converted into sparse vegetation and 881.82 hectares were eventually bared. Barren lands had increased to 125.09% or 1,228.48 hectares where 622.35 hectares of it or 50.66% was due to mining. Moreover, it proved that object based image analysis with SVM was more reliable than pixel-based approach using maximum likelihood and support vector machine algorithms with an accuracy of 68.75%, 81.25%, and 91.3% respectively for the September 2014 masked scene and 52.94%, 64.71%, and 85.29% respectively for the September 2009 image. KEY WORDS: Land cover, Change Detection, Object-Based Image Analysis, Pixel-Based Image Analysis 1. INTRODUCTION In remote sensing, land cover refers to data that is a result based on the classified return values of the satellite image. Forest cover is a land under natural or planted stands of These classified images can be subjected to change trees whether productive or not (World Bank, 2016). It detection using different remote sensing techniques. usually denotes a densely vegetated region. Moreover, change detection is a process of identifying changes in the state of an object or phenomenon by Historically, Philippines was known to harbor one of the observing multi-temporal datasets (Singh, 1989). world’s lush tropical rainforests. Although between 1934 and 1941, forest cover declined to around 57 percent During past decades, land cover change detection (DENR, 2014). This alarming status calls for immediate techniques had undergone substantial development. mitigation to lengthen the forest’s lifespan which is one Typical method of classification on remote sensing of the countries primary source of raw materials. Enable imagery has been pixel based. It is based on conventional to visualize the current situation of the forest areas land statistical techniques, such as supervised and cover change monitoring should be performed. unsupervised classification. Recently, a new method called object oriented image analysis was introduced. In This continuous deforestation is predominantly identified this approach the image is segmented into multi-pixel using satellite images. These images are processed and object primitives according to both spatial and spectral classified to produce land covers for visual and statistical features of the group of pixels. These segments are assessment. classified according to thresholds that are introduced to the classifier. Furthermore, monitoring the locations and distributions Preliminary processing of these scenes includes of land-cover change is important for establishing links calibration of land satellite images and sub-setting these between policy decisions, regulatory actions and datasets to the geographic boundary of the study area. subsequent land-use activities. Thus, this study aims to Pre-processing was performed using Envi 4.3®. identify the forest cover change of Carrascal, Surigao del Sur, Philippines from 2010 to 2014 and to assess the Thematic layers such as barangay, municipality, and current classification techniques with its ability to provincial boundaries were downloaded at Philippine GIS accurately classify land satellite images. This research Data (http://www.philgis.org/freegisdata.htm). Ground was specifically employed to gauge object based image truths were collected on January 24, 2015 and an existing analysis using eCognition with support vector machine as land cover map of Caraga for 2010 from Responsible its classifier and Pixel Based Image Analysis using Mining Project (ICT for RM) was used as a map maximum likelihood and support vector machine. reference during the assessment. An existing tenement map and mining site locations were gathered from ICT for RM and Google Earth to identify the areas bared by 2. METHODOLOGY mining. 2.1 Study Area All data were projected to the Universal Transverse Carrascal, Surigao del Sur, Philippines covers an area of Mercator (UTM) projection system Zone 51N and datum 26,580 hectares. It is geographically located in the of World Geodetic System 1984 (WGS84) to achieve province of Surigao Del Sur Region XIII Caraga which is consistency. a part of the Mindanao group of islands (Alojado, 2015). 2.3 Methodological Framework Regions where forest cover change had clearly occurred Figure 2 shows the classification procedure of the study were chosen to be the area of interest of the study as wherein satellite images were calibrated through Envi shown in Figure 1. The entire vicinity has a total area of ® 4.3 . Images for change detection were classified using 14,246.64 hectares. the object-based approach. Forest cover change was then derived from the resulting land cover maps. Figure 1. Study Area Carrascal, Surigao del Sur, Philippines is characterized by large areas of forest land wherein the Municipality had undergone vegetation change throughout the decade. In 2007 its doors were opened to mining. It houses three Figure 2. Object Based Image Classification and mining corporations namely: VTP Construction and Change Detection Mining Corporation; CTP Construction and Mining Corporation; and Carrascal Nickel Corporation (DOLE, For the comparison of land cover maps from the 2012). September 2014 and September 2009 satellite images, the extracted classes were accuracy assessed using the 2.2 Data Collection and Data Pre-processing gathered groundtruth points and an existing land cover map as shown in Figure 3. Land Satellite images of 2010 and 2014 with a spatial resolution of 30 meters were collected through Earth Explorer (www.earthexplorer.usgs.gov). Landsat 8 OLI of September 23, June 3, June 19, and May 2, 2014 was used in generating a land cover map for 2014. February 8, 2010 and September 17, 2009 Landsat 7 ETM+ was utilized for the 2010 land cover. 2.4.2 Pixel Based Image Analysis Satellite images were classified using Envi 4.3®’s maximum likelihood and support vector machine. Since the sample objects in eCognition have the same idea with the region of interest in Envi 4.3®, these were positioned as close as possible to eradicate biases. The resulting land cover maps were not subjected to contextual editing. 2.5 Accuracy Assessment An error matrix based on the groundtruth points and 2010 land cover map were used to evaluate the quality of the land cover maps. 2.6 Forest Cover Change Figure 3. Evaluation of Different Classification Techniques Maps from the Object Based Image Approach were contextually edited. These were used to determine the 2.4 Image Classification forest cover alterations. It was calculated using the formula: 2.4.1 Object Based Image Analysis Segmentation algorithm was used to subdivide the entire % Change= [(Af – Ai)/Ai] * 100 (1) image into smaller image objects according to its homogeneity. The segmentation was based on the spectral differences. Blue, Green, Red, Shortwave 3. RESULTS AND DISCUSSION Infrared 1, Panchromatic, Cirrus, and Normalized Difference Vegetation Index were used for 2014 3.1 Land Cover Classification Results segmentation while 2010 and 2009 images used the Blue, Green, Red, NIR, Shortwave Infrared 1, Shortwave From the Object Based Image Analysis land cover maps, Infrared 2, Panchromatic, and Normalized Difference it obtained an accuracy of 99.6% for the year 2010 and Vegetation Index. 99.8% for 2014. Multi-resolution segmentation was done for the 2014 In 2010, dense vegetation was dominant taking up images. These were divided into homogenous objects 64.10% of the study landscape followed by sparse using the parameter scale 0.5, shape 0.1, and compactness vegetation (26.80%) and bare lands 6.89%, while water 0.5. During the ruleset development process for the and built-up shared small proportion of 1.23%, and Landsat 7 ETM+ images, it was observed that the 0.97%, respectively. In 2014, dense vegetation was also segmentation parameters used on Landsat 8 OLI accounted for having the largest area with 54.83% and subdivides the image into bigger objects from which it sparse vegetation, bare lands, built-up areas, and water enclose features that belongs to different classes. Since have 27.28%, 15.52%, 1.47%, and 0.90% respectively. the segmentation parameters are interdependent to the given images, a new ruleset was developed. To further optimize the classification accuracy, Chessboard segmentation algorithm was employed for the 2010 and 2009 images. Sample objects were extracted from the segmented images. These samples were classified into Barren, Built- up, Dense Vegetation, Sparse Vegetation, and Water.