Spatial Pattern Analysis of the Ecosystem Services in The

sustainability

Article Spatial Pattern Analysis of the Ecosystem Services in the Guangdong-Hong Kong-Macao Greater Bay Area Using Sentinel-1 and Sentinel-2 Imagery Based on Deep Learning Method

Dawei Wen 1 , Song Ma 2,*, Anlu Zhang 1 and Xinli Ke 1

1 College of Public Administration, Huazhong Agricultural University, Wuhan 430070, China; [email protected] (D.W.); [email protected] (A.Z.); [email protected] (X.K.) 2 Shenzhen Environmental Monitoring Centre, Shenzhen 518049, China * Correspondence: [email protected]

Abstract: Assessment of ecosystem services supply, demand, and budgets can help to achieve sustainable urban development. The Guangdong-Hong Kong-Macao Greater Bay Area, as one of the most developed megacities in China, sets up a goal of high-quality development while fostering ecosystem services. Therefore, assessing the ecosystem services in this study area is very important to guide further development. However, the spatial pattern of ecosystem services, especially at local scales, is not well understood. Using the available 2017 land cover product, Sentinel-1 SAR and Sentinel-2 optical images, a deep learning land cover mapping framework integrating deep change vector analysis and the ResUnet model was proposed. Based on the produced 10 m land cover map Citation: Wen, D.; Ma, S.; Zhang, A.; for the year 2020, recent spatial patterns of the ecosystem services at different scales (i.e., the GBA, Ke, X. Spatial Pattern Analysis of the 11 cities, urban–rural gradient, and pixel) were analyzed. The results showed that: (1) Forest was Ecosystem Services in the the primary land cover in Guangzhou, Huizhou, Shenzhen, Zhuhai, Jiangmen, Zhaoqing, and Hong Guangdong-Hong Kong-Macao Kong, and an impervious surface was the main land cover in the other four cities. (2) Although Greater Bay Area Using Sentinel-1 and Sentinel-2 Imagery Based on ecosystem services in the GBA were sufﬁcient to meet their demand, there was undersupply for Deep Learning Method. Sustainability all the three general services in Macao and for the provision services in Zhongshan, Dongguan, 2021, 13, 7044. https://doi.org/ Shenzhen, and Foshan. (3) Along the urban–rural gradient in the GBA, supply and demand capacity 10.3390/su13137044 showed an increasing and decreasing trend, respectively. As for the city-level analysis, Huizhou and Zhuhai showed a ﬂuctuation pattern while Jiangmen, Zhaoqing, and Hong Kong presented Academic Editor: a decreasing pattern along the gradient. (4) Inclusion of neighborhood landscape led to increased Georgios Koubouris demand scores in a small proportion of impervious areas and oversupply for a very large percent of bare land. Received: 6 May 2021 Accepted: 16 June 2021 Keywords: ecosystem service; spatial pattern; land cover; deep learning; Sentinel Published: 23 June 2021

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in 1. Introduction published maps and institutional affil- iations. Ecosystem services, defined as benefits that humans can obtain from the ecosystem, have been increasingly investigated and considered in both research and policy communi- ties due to raising awareness among decision-makers and the public [1–3]. The maintenance of ecosystem services for enhancing human well-being was included in a set of goals of the 2050 Vision of the Strategic Plan for Biodiversity of the Convention on Biological Diver- Copyright: © 2021 by the authors. Licensee MDPI, Basel, Switzerland. sity [4]. In addition, as summarized in [5], China has established several relevant policies, This article is an open access article for example, a national Forest Eco-Compensation Fund system [6] and ecological redline distributed under the terms and policy [7]. Simultaneously, due to urbanization and population growth, ecosystem services conditions of the Creative Commons are continuously in decline. As indicated in [8], the global loss of ecosystem services due to Attribution (CC BY) license (https:// land use changes from 1997 to 2011 was estimated at $US 4.3–20.2 trillion/year. Therefore, creativecommons.org/licenses/by/ the assessment of ecosystem services is very important for decision-makers to achieve 4.0/). sustainable urban development.

Sustainability 2021, 13, 7044. https://doi.org/10.3390/su13137044 https://www.mdpi.com/journal/sustainability Sustainability 2021, 13, 7044 2 of 16

The ecosystem services can be quantified in semi-quantitatively (e.g., capacity scores) [9,10], monetary units (e.g., yuan and dollar) [11,12], and biophysical units (e.g., biomass) [13,14]. Among the different available methods, expert-based ecosystem services scores are considered to be flexible and adaptable [1]. One such method is the “capacity matrix approach” linking supply, demand, and budget capacity to land use/land cover types [15], which was referred to as “one of the most popular ecosystem services assessment techniques today” in [16]. It has been widely applied in many studies. For example, ecosystem services capacity scores for each land use type and their total supply, demand, and budgets in Shenzhen were mapped and monitored in [9]. In [17], the time-varying characteristics of different ecosystem services of China on three scales: the total quantity change on a national scale, the time-varying trends on a provincial scale, and the change rates on a city scale were analyzed. An assessment of the spatial pattern of ecosystem services across Europe on a 1 km2 grid was presented [18]. Although extensive studies at different spatial scales ranging from regional [9] to national [17], and even to continental [18,19] scales have been conducted, the spatial pattern of ecosystem services at higher spatial resolution (10 m or less) is rarely concerned. Unfortunately, relevant information for local-scale decision-making is still lacking. Since remote sensing can provide data at a low cost, it is appropriate for ecosystem services assessment at higher spatial resolution. The free availability of medium-high spatial resolution imagery, such as Sentinel-1 SAR and Sentinel-2 optical data, can facilitate the mapping of finer scale (10 m) ecosystem services. Compared with coarse or medium spatial resolution data, they have the advantage of improved spatial details for the identification of micro-scale land cover/land use (e.g., small vegetation patch in urban area). In addition, the fusion of Sentinel-1 and Sentinel-2 has been proved to be effective in improving land use/land cover mapping accuracies [20,21] since complemented backscatter and spectral information are utilized [22]. In this way, the reliability of ecosystem services analysis is guaranteed. For the land use/land cover mapping techniques, advanced artificial neural networks, especially deep learning models, have gained increased attention in the remote sensing domain owing to the end-to-end nature (i.e., learn to discriminate features designed for image classification and associated classifier jointly [23]). One of the obstacles affecting the performance of deep learning networks is the limited training samples problem [24], while training samples collection work is expensive and time-consuming. Therefore, a deep learning-based land use/land cover mapping framework that can fully integrate Sentinel SAR-optical information and handle the training samples lacking problem should be established. Based on the two aforementioned aspects, the overarching goal of this paper is to develop a land cover mapping architecture using Sentinel-1 and Sentinel-2 imagery based on the deep learning method and investigate the spatial pattern of ecosystem services at a finer spatial scale. To this end, we take the Guangdong-Hong Kong-Macao Greater Bay Area (GBA), one of the most important economic regions in China, as our study area. Firstly, a deep learning method combing deep change vector analysis and model fine-tuning was proposed for land cover mapping. Then, the ecosystem service values were quantified by linking land cover types to supply and demand matrixes. Finally, ecosystem services scores at different spatial scales (i.e., the GBA, different cities, buffer zones, and pixels) were investigated and analyzed.

2. Materials and Methods 2.1. Study Areas The GBA, situated at the Pearl River Delta, is composed of 9 mainland cities and 2 spe- cial administrative regions, including Guangzhou, Shenzhen, Zhuhai, Foshan, Huizhou, Dongguan, Zhongshan, Jiangmen, Zhaoqing, Hong Kong, and Macao (Figure1). With 56,000 km2 coverage (about 0.6% of China’s land area), the GBA created a gross domestic product (GDP) of more than $11 trillion in 2019, contributing about 1/7 of the national GDP. In February 2019, the Chinese government released the Outline Development Plan Sustainability 2021, 13, x FOR PEER REVIEW 3 of 16

With 56,000 km2 coverage (about 0.6% of China’s land area), the GBA created a gross do- Sustainability 2021, 13, 7044 3 of 16 mestic product (GDP) of more than $11 trillion in 2019, contributing about 1/7 of the national GDP. In February 2019, the Chinese government released the Outline Development Plan for the Guangdong-Hong Kong-Macao Greater Bay Area. It clarified the strategic positioningfor the Guangdong-Hong of a vibrant world-class Kong-Macao agglomeration, Greater Bay Area. an international It clariﬁed the technology strategic position- and inno- vationing of center, a vibrant an world-class important support agglomeration, pillar for an the international Belt and Road technology Initiative, and a innovationshowcase for in-depthcenter, an cooperation important support between pillar the for mainland, the Beltand Hong Road Kong, Initiative, and Macao, a showcase and fora high-quality in-depth livingcooperation circle [25]. between Under the the mainland, collaborative Hong development Kong, and Macao, of the and economy a high-quality and the livingenviron- ment,circle balancing [25]. Under urban the collaborativedevelopment development and ecological of conservation the economy will and become the environment, increasingly prominent.balancing urban Therefore, development analysis and of ecologicalecosystem conservation service value will in become the GBA increasingly is fundamental promi- to nent. Therefore, analysis of ecosystem service value in the GBA is fundamental to guide guide sustainable urban development. sustainable urban development.

FigureFigure 1. 1. GeographicalGeographical information ofof thethe Guangdong-Hong Guangdong-Hong Kong-Macao Kong-Macao Greater Greater Bay Bay Area. Area.

2.2.2.2. Datasets Datasets InIn this this study, study, thethe 1010 mm landland covercover product product FROM-GLC10 FROM-GLC10 [26 [26]] and and Sentinel-1 Sentinel-1 and and Sentinel-2Sentinel-2 remote remote sensing datasetsdatasets werewere used. used. FROM-GLC10 FROM-GLC10 is is a 10a 10 m m global global land land cover cover productproduct of thethe yearyear 2017 2017 and and has has 10 10 land land cover cover types, types, including including cropland, cropland, forest, forest, grassland, grass- shrubland, wetland, water, tundra, impervious surface, bare land, and snow/ice. Tundra land, shrubland, wetland, water, tundra, impervious surface, bare land, and snow/ice. and snow/ice are not included in our study area. Wetland is merged into grassland since Tundra and snow/ice are not included in our study area. Wetland is merged into grass- they show very similar spectral signature, and some classified wetland pixels are found landto be since grassland. they show Therefore, very 7similar land cover spectral types signature, (i.e., cropland, and some forest, classified grassland, wetland shrubland, pixels arewater, found impervious to be grassland. surface, Therefore, and bare land)7 land were cover employed types (i.e., as ourcropland, classification forest, grassland, system. shrubland,The remote water, sensing impervious datasets composed surface, and of Sentinel-1 bare land) SAR were ground employed range as detected our classification (GRD) system.images The and remote Sentinel-2 sensing optical datasets images compos were obtaineded of Sentinel-1 from Sentinels SAR ground Scientific range Data detected Hub (GRD)(https://scihub.copernicus.eu/ images and Sentinel-2 optical; accessed images on 18 we Februaryre obtained 2021) from. The Sentinels pre-processing Scientific steps Data Hubof Sentinel-1 (https://scihub.copernicus.eu/; GRD data, including calibration, accessed speckle-filter,on 18 February and 2021). terrain The correction, pre-processing were stepsimplemented of Sentinel-1 by the GRD European data, Spaceincluding Agency’s calibration, (ESA) Sentinel speckle-filter, Application and Platform terrain correction, (SNAP), weregenerating implemented geocoded by VVthe andEuropean VH backscattering Space Agency’s coefficients (ESA) Sentinel with the Application spatial resolution Platform (SNAP),of 10 m. generating Sentinel-2 geocoded satellites canVV provideand VH fourbackscattering 10 m bands, coefficients six 20 m with bands, the and spatial three res- olution60 m bands. of 10 Sentinel-2m. Sentinel-2 datasets satellites were processedcan provide into four atmospherically 10 m bands, correctedsix 20 m productsbands, and threeby the 60 ESA’s m bands. Sen2Cor Sentinel-2 algorithm datasets [27]. were Only processed four 10 m into bands atmospherically (i.e., blue, green, corrected red, and prod- uctsnear-infrared by the ESA’s bands) Sen2Cor of Sentinel-2 algorithm data [27]. were Only used. four Multi-temporal 10 m bands (i.e., Sentinel blue, green, optical-SAR red, and near-infraredimages were acquiredbands) of almost Sentinel-2 simultaneously data were around used. 26Multi-temporal October 2017 and Sentinel 29 October optical-SAR 2020 to cover the entire study area, referred to as the ‘GBA 2017’ and ‘GBA 2020’ images. Sustainability 2021, 13, x FOR PEER REVIEW 4 of 16

images were acquired almost simultaneously around 26 October 2017 and 29 October 2020 to cover the entire study area, referred to as the ‘GBA 2017′ and ‘GBA 2020′ images.

2.3. Land Cover Mapping The methodological framework of the deep learning-based land cover mapping is presented in Figure 2, which has three major components: (1) Residual Unet (ResUnet) training with pre-temporal data, (2) sample generation with deep change vector analysis (CVA), and (3) fine-tuning ResUnet with post-temporal data. ResUnet training with pre- temporal data refers to model training with 2017 labeled data, serving as feature extraction for the following CVA. Samples generation with deep CVA is to produce training samples for the classification of the GBA 2020 images. Moreover, fine-tuning ResUnet with post- temporal data is to fine-tune the network pre-trained on the GBA 2017 images, making it more adaptable to the 2020 image classification tasks. Sentinel-1 optical/Sentinel-2 SAR images with the spatial resolution of 10 m are rich in spectral/backscatter and spatial information. Inspired by the effective capturing of this information by ResUnet in optical [28] and SAR [29]images, it was used for effective feature extraction and following classification. In this study, ResUnet was adopted for both deep feature extraction in CVA and image classification. The architecture of the ResUnet is shown in Figure 3. ResUnet is an integrated architecture incorporating Unet and deep residual learning, which has three parts: encoding, decoding, and bridge [28]. On the one

Sustainability 2021, 13, 7044 hand, Unet, a pixel-wise semantic segmentation structure, can combine low-level4 of 16 detail information and high-level semantic information [30], thus enabling promising performance in feature extraction and classification [31]. On the other hand, residual blocks can ease 2.3.deep Land network Cover Mapping training by preserving input features through identity skip connections [28]. Therefore, ResUnet was used in this study, considering that it is appropriate for The methodological framework of the deep learning-based land cover mapping is accuratepresented pixel-wise in Figure classification2, which has threeand majorfine-scale components: spatial (1)pattern Residual analysis. Unet (ResUnet) Furthermore, to ensuretraining that information with pre-temporal of small data, targets (2) sample in generationthe 10 m spatial with deep resolution change vector images analysis can be better preserved(CVA), after and (3) convolution fine-tuning ResUnet and pooling with post-temporal layers, while data. higher ResUnet abstract training features with pre- can be captured,temporal a 7-level data architecture refers to model including training with the 2017 total labeled of 15 data, convolution serving as featurelayers extractionwas utilized. As shownfor in the Figure following 3, CVA.image Samples patches generation with the with size deep of 512 CVA × is 512 to produce × 6 (i.e., training four spectral samples and two for the classification of the GBA 2020 images. Moreover, fine-tuning ResUnet with post- backscattertemporal intensity data is to fine-tunebands) are the networkfed to the pre-trained network, on theand GBA the 2017 512 images, × 512 making× 7 classification it probabilitymore adaptable map is toproduced the 2020 image at the classification output layer. tasks.

Figure 2. Methodological framework of the deep learning-based land cover mapping. Figure 2. Methodological framework of the deep learning-based land cover mapping. Sentinel-1 optical/Sentinel-2 SAR images with the spatial resolution of 10 m are rich in spectral/backscatter and spatial information. Inspired by the effective capturing of this information by ResUnet in optical [28] and SAR [29] images, it was used for effective feature extraction and following classification. In this study, ResUnet was adopted for both deep feature extraction in CVA and image classification. The architecture of the ResUnet is shown in Figure3. ResUnet is an integrated architecture incorporating Unet and deep residual learning, which has three parts: encoding, decoding, and bridge [28]. On the one hand, Unet, a pixel-wise semantic segmentation structure, can combine low-level detail information and high-level semantic information [30], thus enabling promising performance in feature extraction and classification [31]. On the other hand, residual blocks can ease deep network training by preserving input features through identity skip connections [28]. Therefore, ResUnet was used in this study, considering that it is appropriate for accurate pixel-wise classification and fine-scale spatial pattern analysis. Furthermore, to ensure that information of small targets in the 10 m spatial resolution images can be better preserved after convolution and pooling layers, while higher abstract features can be captured, a 7-level architecture including the total of 15 convolution layers was utilized. As shown in Figure3, image patches with the size of 512 × 512 × 6 (i.e., four spectral and two backscatter intensity bands) are fed to the network, and the 512 × 512 × 7 classification probability map is produced at the output layer. For sample generation, unchanged locations between 2017 and 2020 were identified through deep CVA [32], and the corresponding unchanged 2020 images were labeled according to land cover types in the FROM-GLC10 map. Deep CVA starts from the extraction of deep features that can model low-level spatial context and high-level semantic information. Deep CVA proposed by Saha et al. [31] employed a pre-trained CNN model trained on aerial RGB image bands for deep feature extraction. Since it was trained on different data sets and tasks, extracted deep features are not optimal. Therefore, in our study, firstly, the ResUnet model was trained with stacked Sentinel 1 and 2 image bands in 2017 as inputs and FROM-GLC10 map as labels. With the trained ResUnet model, the multi-temporal deep features were obtained by feeding the GBA 2017 and 2020 images separately as inputs. Subsequently, deep change vectors were generated by subtracting Sustainability 2021, 13, 7044 5 of 16

the multi-temporal deep features. Locations with a magnitude of deep change vectors smaller than 10 were identiﬁed as unchanged. The threshold value of 10 was determined based on visual inspection of CVA magnitude results. It should be noted that a deliberately Sustainability 2021, 13, x FOR PEER REVIEW selected strict threshold was used to ensure that detected unchanged regions were highly 5 of 16

reliable. Finally, with the generated 2020 training samples, the trained ResUnet model can be ﬁne-tuned to suit the classiﬁcation of the GBA in the year 2020.

Figure 3. The architecture of ResUnet for land cover mapping (ResBlock = Residual Block, ReLU = Rectiﬁed Linear Unit). Figure 3. The architecture of ResUnet for land cover mapping (ResBlock = Residual Block, ReLU = Rectified Linear Unit). 2.4. Spatial Pattern Analysis of Ecosystem Services For sampleAssessment generation, of ecosystem unchanged services was location conducteds between based on the 2017 expert-based and 2020 supply were identified and demand matrixes proposed in [15]. The supply and demand capacity attributed throughto deep each classCVA are [32] expert, and evaluation the corresponding values. In the unchanged matrix, capacities 2020 of images different were land labeled ac- cordingcover to land types cover to supply/demand types in the 22 ecosystemFROM-GLC10 services (threemap. general Deep categoriesCVA starts including from the extraction of regulating,deep features provisioning, that can and model cultural low-leve services) werel spatial provided. context A relative and scalehigh-level ranging semantic in- from 0 to 5 was used to indicate no relevant capacity, low relevant capacity, relevant formation.capacity, Deep medium CVA relevant proposed capacity, by high Saha relevant et al. capacity, [31] employed and very high a relevant pre-trained capacity CNN model trainedof on supplying/demanding aerial RGB image a certainbands ecosystem for deep service feature within extraction. the particular Since land it cover was trained on differenttype. data There sets are and 11, 9, andtasks, 2 sub-services extracted in deep regulating, features provisioning, are not and optimal. cultural services, Therefore, in our respectively. Due to the clarity of the results, only the three general ecosystem categories study, firstly,were considered. the ResUnet With this model regard, was supply traine and demandd with capacity stacked scores Sentinel were calculated 1 and 2 by image bands in 2017multiplying as inputs the and score FROM-GLC10 of each individual map ecosystem as labels. service With for a certain the trained land cover ResUnet type in model, the multi-temporalthe matrix bydeep the areafeatures of this landwere cover obtained type in theby region feeding of interest the GBA and then 2017 summed, and 2020 images separatelyand subsequentlyas inputs. demandSubseq capacityuently, scores deep were change subtracted vectors from the were supply generated capacity scores by subtracting to derive budget scores (Table1). As can be seen in Table1, the main human-dominated the multi-temporalland cover type (i.e.,deep impervious features. surface) Locations has a higher with demand a magnitude capacity, andof thedeep natural change vectors smallerland than cover 10 typeswere including identified forest, as grassland,unchanged. shrubland, The threshold and water are value characterized of 10 was by determined based onhigher visual supply inspection capacity and of lower CVA demand magnitude capacity. results. It should be noted that a deliberately selected strict threshold was used to ensure that detected unchanged regions were highly reliable. Finally, with the generated 2020 training samples, the trained ResUnet model can be fine-tuned to suit the classification of the GBA in the year 2020.

2.4. Spatial Pattern Analysis of Ecosystem Services Assessment of ecosystem services was conducted based on the expert-based supply and demand matrixes proposed in [15]. The supply and demand capacity attributed to each class are expert evaluation values. In the matrix, capacities of different land cover types to supply/demand 22 ecosystem services (three general categories including regulating, provisioning, and cultural services) were provided. A relative scale ranging from 0 to 5 was used to indicate no relevant capacity, low relevant capacity, relevant capacity, medium relevant capacity, high relevant capacity, and very high relevant capacity of supplying/demanding a certain ecosystem service within the particular land cover type. There are 11, 9, and 2 sub-services in regulating, provisioning, and cultural services, respectively. Due to the clarity of the results, only the three general ecosystem categories were considered. With this regard, supply and demand capacity scores were calculated by multiplying the score of each individual ecosystem service for a certain land cover type in the matrix by the area of this land cover type in the region of interest and then summed, and subsequently demand capacity scores were subtracted from the supply capacity scores to derive budget scores (Table 1). As can be seen in Table 1, the main human-dominated land cover type (i.e., impervious surface) has a higher demand capacity, and the Sustainability 2021, 13, 7044 6 of 16

Table 1. Assessment matrix illustrating supply, demand, and budget capacity scores for regulating, provision, and cultural services within the different land cover types (Reg = regulating services, Pro = provisioning services, and Cul = cultural services).

Supply Demand Budgets Land Cover Reg Pro Cul Reg Pro Cul Reg Pro Cul Cropland 7 17 1 15 5 0 −8 12 1 Forest 39 22 10 0 3 0 39 19 10 Grassland 22 5 6 6 5 5 16 0 1 Shrubland 39 22 10 0 3 0 39 19 10 Water 7 14 9 0 1 0 7 13 9 Impervious surface 0 1 0 28 46 6 −28 −45 −6 Bare land 0 0 0 11 10 0 −11 −10 0

By linking the land cover types in the classification map to corresponding capacity scores, fine-scale ecosystem service supply, demand, and budget maps were generated. Three aspects were considered to analyze the spatial pattern of ecosystem services in the GBA: total ecosystem services values, ecosystem services values along the urban–rural gradient, and ecosystem services balances at the local scale. The total ecosystem services values were investigated for the GBA and the 11 cities. For the ecosystem services values along the urban–rural gradient, buffer zones emanating outward from the urban center (i.e., the centroid of impervious surface area) were constructed. The radiation distances between each buffer zone for the GBA and the 11 cities were set as 10 km and 5 km, respectively. The constructed buffer zones were supposed to cover the GBA and each city. In particular, the buffer zones emanating outward from cities’ urban centers were constructed to cover each city, and zone areas that overlapped with neighboring cities were kept. In this way, spatial neighboring effects can be considered and analyzed. The mean values of ecosystem services in each buffer zone were calculated to ensure the evaluation metric independent of the buffer zone size and the shape of cities. Focusing on distance from urban center versus mean values of ecosystem services, their spatial patterns along the urban–rural gradient can be presented. Finally, ecosystem services balances between the demand capacity of a certain pixel and the supply capacity of its neighboring pixels were considered, which was designed to measure the influence of neighborhood landscape on ecosystem services supply and demand balances at a fine spatial scale. In specific, a local neighborhood with a certain circle radius R was defined. Based on the underlying assumption that neighboring pixels that are closer to the center one can provide greater supply capacity, the inverse distance weights were used. On the other hand, since each pixel can provide supply capacity for pixels within the neighborhood, the inverse distance was multiplied by a constant coefficient K to ensure the sum of weight values equal to 1. Ecosystem services balance for each pixel was calculated as follows:

1 1 ESB = ∑ K × × Si − Dcenter, ∑ K × = 1, Distcenter = 0.5 (1) i∈pixels Disti i∈pixels Disti

where pixels and center denote all pixels and center one in a local neighborhood, Disti is the distance between pixel i in a local neighborhood and a center one, S and D are the supply and demand capacity scores, respectively. In this study, R was set as 200 m, 500 m, and 1000 m to exploit the effect of multi-scale landscape on ecosystem services balance.

3. Results 3.1. Land Cover Mapping Results The accuracy of the land cover mapping network around 90% was achieved, which indicated that the proposed framework was able to generate reliable classification results. Sustainability 2021, 13, 7044 7 of 16 The land cover mapping result is presented in Figure 4. The impervious surface located in the middle of the GBA and the impervious surface of some cities were spatially con- nected, for example, Foshan and Guangzhou. Forestland was mainly distributed in the upper-leftfor example, and Foshanupper-right and Guangzhou. regions of Forestland the GBA was(i.e., mainly Zhaoqing distributed and Huizhou). in the upper-left Cropland, and upper-right regions of the GBA (i.e., Zhaoqing and Huizhou). Cropland, taking up taking up a small proportion, was mainly distributed in the rural area of Zhaoqing, Hui- a small proportion, was mainly distributed in the rural area of Zhaoqing, Huizhou, and zhou,Jiangmen. and Jiangmen. The GBA The was GBA rich was in natural rich in water natural resources. water resources. A dense riverA dense network river was network wasstaggered staggered in in the the impervious impervious surface surface area, area, and and ponds ponds and and lakes lakes were were intertwined intertwined with with riversrivers or ordistributed distributed in in the the forest area.area.

FigureFigure 4. 4. LandLand cover cover mapping result result of of the the GBA. GBA.

TheThe percentages percentages of of different different land land cover cover type typess in in the the GBA GBA and different citiescitiesare are pre- sentedpresented in Figure in Figure 5 5to to reveal reveal land covercover composition. composition. Results Results indicated indicated that forest,that forest, cropland, and impervious surface were the three major land cover types in the GBA. As for cropland, and impervious surface were the three major land cover types in the GBA. As the eleven cities, forests were the primary land cover in Guangzhou, Huizhou, Shenzhen, forZhuhai, the eleven Jiangmen, cities, Zhaoqing,forests were and the Hong primary Kong; land and the cover other in four Guangzhou, cities (i.e., Huizhou, Dongguan, Shen- zhen,Foshan, Zhuhai, Zhongshan, Jiangmen, and Macao)Zhaoqing, had theand largest Hong proportion Kong; and in impervious the other surface. four cities Land (i.e., Dongguan,urbanization Foshan, that can Zhongshan, be approximated and byMacao) a percentage had the of imperviouslargest proportion surface was in greatest impervious surface.in Macao, Land followed urbanization by Dongguan, that can Shenzhen, be approximated Zhongshan, by Foshan,a percentage Zhuhai, of Guangzhou,impervious sur- faceHong was Kong, greatest Jiangmen, in Macao, Huizhou, followed and Zhaoqing. by Dongguan, The last four Shenzhen, cities had theZhongshan, percentage Foshan, of Zhuhai,impervious Guangzhou, surface smallerHong Kong, than that Jiangmen, in the GBA, Huiz suggestinghou, and that Zhaoqing. they had The more last natural four cities hadland the sources. percentage of impervious surface smaller than that in the GBA, suggesting that Overall, the GBA was a megacity region with a high urbanization level and abundant they had more natural land sources. natural landscape. In addition, the composition and distribution of the different land cover types were varied across the GBA and among different cities. Therefore, it is important to conduct an analysis focusing on the spatial pattern of the ecosystem services. Sustainability 2021, 13, x FOR PEER REVIEW 8 of 16 Sustainability 2021, 13, 7044 8 of 16

Figure 5.Figure The percentages 5. The percentages of the of thedifferent different land land covercover types types in thein GBAthe GBA and different and different cities. cities. 3.2. Ecosystem Services of the GBA and Different Cities Overall,The capacitythe GBA scores was of a ecosystem megacity services region supply, with demand, a high urbanization and budget in the level GBA and are abundant naturalpresented landscape. in Table In2. Theaddition, total budget the in composit the GBA wasion 184,467,815, and distribution indicating thatof suppliedthe different land cover ecosystemtypes were services varied were across sufﬁcient the to GBA meet demandand am capacity.ong different The GBA cities. had the Therefore, highest it is im- portantsupply to conduct capacity an in regulatinganalysis focusing services, followed on the byspatial provisioning pattern and of culturalthe ecosystem services. services. There was the highest demand capacity in provisioning services for the GBA, followed by regulating and cultural ones. 3.2. Ecosystem Services of the GBA and Different Cities TheTable capacity 2. Capacity scores of ecosystem of ecosystem services supply, services demand, andsupply, budget demand, in the GBA. and budget in the GBA are presentedServices in Table Regulating2. The total budget Provisioning in the GBA Cultural was 184,467,815, Total indicating that supplied ecosystemSupply services 148,159,321 were sufficient 91,634,000 to meet 40,677,589demand capacity. 280,470,910 The GBA had the Demand 33,558,265 55,486,016 6,958,814 96,003,095 highest supplyBudget capacity 114,601,056 in regulating services, 36,147,984 followed 33,718,775 by provisioning 184,467,815 and cultural services. There was the highest demand capacity in provisioning services for the GBA, followed by Greatregulating differences and incultural ecosystem ones. services among different cities were observed, as shown in Figure6. Although budgets were positive in most cities, there are several Table exceptions,2. Capacity for of ecosystem example, all services the services supply, categories demand, in Macao, and budget provisioning in the GBA. services in Zhongshan, Dongguan, Shenzhen, and Foshan. The total budgets for Macao, Zhongshan, ServicesDongguan, and FoshanRegulating were also negative.Provisioning Among the 11 cities,Cultural the supply capacity Total Supplyof regulating, provisioning,148,159,321 and cultural91,634,000 services was the largest40,677,589 in Zhaoqing, followed280,470,910 by Huizhou, Jiangmen, and Dongguan. The top four cities that had higher demand Demandcapacity were Guangzhou,33,558,265 Huizhou, Jiangmen,55,486,016 and Foshan. Focusing6,958,814 on the three services96,003,095 Budgetcategories, the supply114,601,056 capacity for the different36,147,984 cities tended to33,718,775 share a similar pattern:184,467,815 regulating > provisioning > cultural services. The demand capacity also had the same descending order as provisioning > regulating > cultural services for all the 11 cities. Great differences in ecosystem services among different cities were observed, as shown3.3. in EcosystemFigure 6. Services Although along the budgets Urban–Rural were Gradient positive in most cities, there are several exceptions, forThe example, mean value all of differentthe services ecosystem categories services capacity in Macao, along the urban–ruralprovisioning gradi- services in Zhongshan,ent in the Dongguan, GBA is shown Shenzhen, in Figure7. and Supply Foshan. and demand The total capacity budgets showed for an increasingMacao, Zhongshan, and decreasing trend from urban center to rural areas, respectively, which is attributed Dongguan,to the fact and that Foshan the urban were core also in the negative. GBA is agglomerated Among the by the11 imperviouscities, the surfacesupply of capacity of regulating,several provisioning, cities and surrounded and cultural by natural services landscape was such the as forests. largest The in peak Zhaoqing, and valley followed by Huizhou,points Jiangmen, for supply capacity and Dongguan. of the three services The top categories four cities turnto that be the had 22nd higher and 3rd demand zones, capacity were Guangzhou, Huizhou, Jiangmen, and Foshan. Focusing on the three services categories, the supply capacity for the different cities tended to share a similar pattern: regulating > provisioning > cultural services. The demand capacity also had the same descending order as provisioning > regulating > cultural services for all the 11 cities. Sustainability 2021, 13, x FOR PEER REVIEW 9 of 16

Sustainability 2021, 13, 7044 9 of 16

Sustainability 2021, 13, x FOR PEER REVIEW 9 of 16 which are the opposite for demand capacity. Therefore, the 22nd and 3rd zones have the largest and smallest total budgets, which correspond to regions with the highest percentage of the natural landscape and impervious surface, respectively.

(a) (b)

Figure 6. Capacity value in the different cities: (a) regulating, provisioning, and cultural services and (b) total services. S_Reg/S_Pro/S_Cul/S_Tot, D_Reg/D_Pro/D_Cul/D_Tot, and B_Reg/B_Pro/B_Cul/B_Tot, are the capacity of supply, demand, and budgets for regulating/provisioning/cultural/total services.

3.3. Ecosystem Services along the Urban–Rural Gradient The mean value of different ecosystem services capacity along the urban–rural gradient in the GBA is shown in Figure 7. Supply and demand capacity showed an increasing and decreasing trend from urban center to rural areas, respectively, which is attributed to the fact that the urban core in the GBA is agglomerated by the impervious surface of several cities and surrounded by natural landscape such as forests. The peak and valley points for supply(a) capacity of the three services categories turn(b) to be the 22nd and 3rd zones, which are the opposite for demand capacity. Therefore, the 22nd and 3rd zones Figure 6.Figure Capacity 6. Capacity value value in the in thedifferent different cities: cities: ( a)) regulating,regulating, provisioning, provisioning, and cultural and cultural services services and (b) total and services. (b) total services. have the largest and smallest total budgets, which correspond to regions with the highest S_Reg/S_Pro/S_Cul/S_Tot,S_Reg/S_Pro/S_Cul/S_Tot, D_Reg/D_Pro/D_Cul/D_Tot, D_Reg/D_Pro/D_Cul/D_Tot, and and B_ B_Reg/B_Pro/B_Cul/B_Tot,Reg/B_Pro/B_Cul/B_Tot, are are the capacitythe capacity of supply, of supply, demand, anddemand, budgets and budgetsfor regulating/provisioning/cultural/total forpercentage regulating/provisioning/cultural/total of the natural landscape services. services. and impervious surface, respectively.

Figure 7. Mean capacity values of ecosystem service along the urban–rural gradient in the GBA. Figure 7. Mean capacity values of ecosystem service along the urban–rural gradient in the GBA. Furthermore, the mean value of different ecosystem services capacity along the urban– ruralFurthermore, gradient for the the mean 11 cities value is presented of different in Figure ecosystem8. It should services be noted capacity that the along con- the urban–ruralstruction gradient of buffer for zones the for 11 each cities city is does present excludeed regionsin Figure of other 8. It cities should to assess be noted spatial that the constructionneighboring of effects.buffer The zones ﬁrst zonefor each that coverscity do otheres exclude cities is marked regions with of aother vertical cities dotted to assess

Figure 7. Mean capacity values of ecosystem service along the urban–rural gradient in the GBA.

Furthermore, the mean value of different ecosystem services capacity along the urban–rural gradient for the 11 cities is presented in Figure 8. It should be noted that the construction of buffer zones for each city does exclude regions of other cities to assess Sustainability 2021, 13, x FOR PEER REVIEW 10 of 16

Sustainability 2021, 13, 7044 spatial neighboring effects. The first zone that covers other cities is marked with a10 vertical of 16 dotted line, and results excluding other cities are presented by dotted lines as a comparison group (Supply_com, Demand_com, and Budgets_com in Figure 8). The vast majority of buffer zones around urban centers covered other cities, for example, 18 out of 21 for line, and results excluding other cities are presented by dotted lines as a comparison group Guangzhou and 19 out of 20 for Zhuhai, since the urban centers are close to each other (Supply_com, Demand_com, and Budgets_com in Figure8). The vast majority of buffer with the mean nearest neighbor distance about 60 km. In view of this, relatively greater zones around urban centers covered other cities, for example, 18 out of 21 for Guangzhou accessibilityand 19 out of of 20 ecosystem for Zhuhai, services since the for urban urba centersn centers are can close be to provided each other by with urban the meanareas in othernearest cities neighbor rather distancethan its own about rural 60 km. area. In Ther viewefore, of this, it is relatively vital to greaterexploit accessibilitythe spatial neigh- of boringecosystem effects. services In contrast for urban to comparison centers can begr providedoups, neighboring by urban areascitiesin generally other cities had rather a nega- tivethan effect its own on the rural ecosystem area. Therefore, service it accordin is vital tog to exploit the smaller the spatial budgets. neighboring However, effects. for Shen- In zhencontrast zones to within comparison 20–30 groups, km from neighboring an urban citiescenter, generally a positive had neighboring a negative effect effect on was the observed.ecosystem The servicegeneral according increasing to patterns the smaller for budgets. budgets However,from urban for centers Shenzhen to rural zones areas within were influenced20–30 km fromin some an urban cities. center, There awere positive two neighboring types of exceptions: effect was fluctuation observed. The(i.e., general Huizhou andincreasing Zhuhai) patterns and decrease for budgets (i.e., fromJiangmen, urban centersZhaoqing, to rural and areasHong were Kong), influenced which may in some be at- tributedcities. There to the were existence two types of multiple of exceptions: centers. fluctuation In addition, (i.e., the Huizhou three cities and Zhuhai)(i.e., Jiangmen, and Zhaoqing,decrease (i.e., and Jiangmen, Hong Kong) Zhaoqing, provided and supply Hong Kong), that exceeds which may demand be attributed in all the to zones, the exis- sug- gestingtence of the multiple requirement centers. of In ecological addition, the benefits three citiesfor humans (i.e., Jiangmen, in both Zhaoqing, urban and and rural Hong areas areKong) met. provided It is interesting supply thatto note exceeds that demandZhongshan in all and the Macao zones, suggestingwere balanced the requirement or undersup- of ecological benefits for humans in both urban and rural areas are met. It is interesting plied in all the zones. The other cities had an undersupply of ecosystem services in the to note that Zhongshan and Macao were balanced or undersupplied in all the zones. The urban core area, and other zones were characterized by ecosystem services’ supplies ex- other cities had an undersupply of ecosystem services in the urban core area, and other ceedingzones were their characterized demands. by ecosystem services’ supplies exceeding their demands.

Guangzhou

Dongguan Huizhou

Figure 8. Cont. SustainabilitySustainability2021 2021,,13 13,, 7044 x FOR PEER REVIEW 1111 ofof 1616

Shenzhen Foshan

Zhongshan Zhuhai

Jiangmen Zhaoqing

Hong Kong Macao

FigureFigure 8.8. TheThe meanmean capacitycapacity valuevalue forfor ecosystemecosystem servicesservices supply,supply, demand,demand, andand budgetsbudgets alongalong thethe urban–ruralurban–rural gradientgradient inin thethe 1111 cities.cities. Sustainability 2021, 13, x FOR PEER REVIEW 12 of 16

Sustainability 2021, 13, 7044 12 of 16

3.4. Ecosystem Services Balance at Local Scale 3.4. EcosystemEcosystem Services services Balance balance at Localat multiple Scale scales was analyzed by presenting the per- centageEcosystem of different services groups balance of balance at multiple (i.e., negative scales values was analyzed indicate bydemand presenting exceeds the sup- per- ply/undersupply,centage of different and groups positive of values balance indica (i.e.,te negative supply valuesexceeds indicate demand/oversupply) demand exceeds in Figuresupply/undersupply, 9. The following and groups positive were values considered: indicate <− supply4 = very exceeds high undersupply, demand/oversupply) (−4, −3) = highin Figure undersupply,9. The following (−3, −2) = groupsmedium were undersupply, considered: (−2, <−1)− 4= low = very undersupply, high undersupply, (−1, 0) = very(−4, low−3) =undersupply, high undersupply (0, 1) =,( very−3, − low2) = over mediumsupply, undersupply, (1, 2) = low ( −oversupply,2, −1) = low (2, undersup- 3) = me- diumply, (− oversupply,1, 0) = very low(3, 4) undersupply, = high oversupply, (0, 1) = very>4 = very low oversupply,high oversupply. (1, 2) = The low influence oversupply, of the(2, 3)neighborhood = medium oversupply, size on the percentages (3, 4) = high of oversupply, different balance >4 = verygroups high is variant oversupply. for differ- The entinﬂuence land cover of the types. neighborhood Several general size on observations the percentages can of be different drawn from balance Figure groups 9. Over is variant 90% offor the different impervious land surface cover types. had very Several high general undersupply, observations and over can 90% be drawn of forests from had Figure a very9. highOver oversupply. 90% of the imperviousThe proportion surface of oversuppl had veryy high was undersupply, over 80% on andthe scale over of 90% 1000 of forestsm and increasedhad a very over high 90% oversupply. on smaller The scales proportion (i.e., 500 of m oversupply and 200 m). was Cropland over 80% and on grassland the scale showedof 1000 ma anddescending increased trend over for 90% percentages on smaller from scales very (i.e., low 500 to m andvery 200high m). oversupply Cropland groups.and grassland Their undersupply showed a descending also took up trend about for 5% percentages and 20%, respectively. from very low It is to interesting very high tooversupply note that the groups. inclusion Their of undersupply neighborhood also increased took up demand about 5% scores and 20%, in a small respectively. proportion It is ofinteresting impervious to note areas that since the the inclusion impervious of neighborhood surface had increased a very low demand budget scores score in and a small its densityproportion was of high impervious in the urban areas area. since Another the impervious land cover surface type had (i.e., a verybare lowland) budget also had score a negativeand its density balance was score high and in showed the urban oversupply area. Another in a landvery coverlarge typepercentage (i.e., bare region land) when also neighborhoodhad a negative landscape balance score was included. and showed oversupply in a very large percentage region when neighborhood landscape was included.

(a) (b)

(c)

Figure 9. Comparison ofof pixel-levelpixel-level ecosystem ecosystem services services demand demand and and supply supply balance balance for for the the different different land land cover cover classes classes with withthe neighborhood the neighborhood being being (a) 200 (a) m, 200 (b m,) 500 (b) m, 500 and m, ( cand) 1000 (c) m.1000 m.

4.4. Discussion Discussion AA series series of policies havehave beenbeen launchedlaunched inin the the GBA GBA to to promote promote ecological ecological land land protec- pro- tectiontion (e.g., (e.g., National National Forest Forest City City Group Group Construction Construction in the in Pearl the Pearl River River Delta Delta (2016–2025)). (2016– 2025)).Beneﬁting Benefiting from the from impact the impact of National of National Forest Forest City Group City Group Construction, Construction, the forest the forest land landreached reached 62.5%, 62.5%, according according to our to land our coverland cover map inmap 2020. in 2020. As reported As reported by a recent by a recent study, study,ecological ecological land shifted land shifted from decline from decline and fragmentation and fragmentation during during 1990–2010 1990–2010 to growth to growth and in- tegration during 2010–2019, and 659 km2 of non-ecological land was converted to ecological Sustainability 2021, 13, 7044 13 of 16

land, which may be in response to the “ecological red line” policy [33]. Since forest provides the largest budget capacity among the seven land cover types, the GBA can obtain an oversupply of ecosystem services with rich forest resources. Although there were very large budgets for regulating, provisioning, and cultural services in the GBA, city-scale results indicated a quietly unbalanced supply–demand mismatch. Speciﬁcally, only Zhaoqing, Huizhou, and Jiangmen had relatively greater supply capacity than demand capacity, and the supply capacity of provisioning services in Zhongshan, Dongguan, Shenzhen, and Foshan could not meet the demand. Meanwhile, Macao faces serious ecological security as all three general services categories were negative. Since cropland, forest, shrubland, and water can supply more provisioning services, it is very necessary to protect or even increase these areas in the above-mentioned four cities. Corresponding city-scale policies have been considered. Shenzhen City issued a “Management stipulation of the basic ecological line” that forbids any construction projects within the basic ecological line [34]. Land policy in Hong Kong emphasizes ecological protection, and its Town Planning Ordinance constitutes the legal basis for the conservation or protection of country parks, green belts [35], making it a highly urbanized city with favorable ecological conditions. How to further improve the effectiveness of ecological land protection policies has become a major challenge. According to the Outline Development Plan for the Guangdong-Hong Kong-Macao Greater Bay Area, the development goal for Huizhou is to turn into a green eco-tourism city and to take ecological responsibility for the Greater Bay Area [36]. The ecosystem service supply capacity of Huizhou ranked second among the 11 cities, which followed behind Zhaoqing. However, Huizhou has a unique location (i.e., adjacent to Guangzhou, Dongguan, and Shenzhen (they are developed cities with dense population density and intensive industry)). As indicated by the urban–rural gradient analysis, interregional cooperation of Shenzhen and other neighboring cities lead to improved budgets in the urban zones. Therefore, interregional cooperation of the multiple cities in the GBA from the perspective of ecological conservation is encouraged. With the growing coordination development of various aspects like economic development, public services, infrastruc- ture construction, and environmental protection in the GBA [37,38], spatial fairness of ecosystem services at the city and local scale should be improved by putting forward multi-city ecological protection and coordinated development policies. Our spatial pattern analysis can provide essential information for further regional development plans without administrative constraints. The spatial pattern analysis in this study is only concerned about the ecosystem in each administrative district scale (i.e., the GBA and 11 cities), along the urban–rural gradient, and at a local scale. To deal with it, most manual spatial pattern analyses (e.g., zone buffers and local neighborhood analysis) were conducted. Recently, some more advanced spatial pattern mining techniques (e.g., spatial hotspot detection, colocation detection, and outlier detection [39], and even spatio-temporal pattern analysis [40]) have been utilized in other studies and show promising results. Therefore, more in-depth spatial and spatio-temporal pattern analysis work can be considered in further research.

5. Conclusions With the support of Sentinel-1 SAR and Sentinel-2 optical remote sensing data, we proposed a deep learning-based land cover mapping framework to generate a 10 m spatial resolution map. Based on the land cover map of the GBA in 2020, the ﬁne-scale spatial pattern of ecosystem services was analyzed. Speciﬁcally, total ecosystem services supply, demand, and budgets in the GBA and the 11 cities are presented. Ecosystem services along the urban–rural gradient were analyzed to reveal their trend from urban centers to rural areas. Finally, ecosystem services balance assessments at the local scale were performed to identify the pixel-level status of supply and demand balance. The major study results showed as follows: (1) Forest, cropland, and impervious surface were the three major land cover types in the GBA. Forest was the primary land cover in Guangzhou, Huizhou, Shenzhen, Zhuhai, Sustainability 2021, 13, 7044 14 of 16

Jiangmen, Zhaoqing, and Hong Kong, and the impervious surface was the main land cover in the other four cities. (2) Although ecosystem services in the GBA were sufﬁcient to meet their demand, there was undersupply for all the three general services in Macao, provision services in Zhongshan, Dongguan, Shenzhen, and Foshan. (3) Along the urban–rural gradient in the GBA, supply and demand capacity showed an increasing and decreasing trend since the urban core in the GBA is agglomerated by the impervious surface of several cities and is surrounded by natural landscape such as forests. As for the city-level urban–rural gradient analysis, except for the general increasing pattern for budgets from the urban centers to rural areas, Huizhou and Zhuhai showed a ﬂuctuation pattern, while Jiangmen, Zhaoqing, and Hong Kong presented a decreasing pattern, which may be attributed to the existence of multiple urban centers. (4) The inclusion of neighborhood landscape increased demand scores in a small proportion of impervious areas, and bare land with a negative balance score showed an oversupply in a very large percentage region when neighborhood landscape was included. To realize the sustainable development of the GBA, on the one hand, local governments should establish policies that strictly protect ecological land (i.e., forests, shrubland, water, and cropland); on the other hand, interregional cooperation of the multiple cities in the GBA from the perspective of ecological conservation is encouraged.

Author Contributions: Conceptualization, S.M. and A.Z.; methodology, D.W. and S.M.; writing— original draft preparation, D.W.; writing—review and editing, X.K.; supervision, A.Z. and X.K. All authors have read and agreed to the published version of the manuscript. Funding: This research was supported by the National Natural Science Foundation of China under Grant 41901279, Shenzhen Science and Technology Program under Grant JCYJ20180306170645080, Environmental research project in Shenzhen City (from Ecological Environment Bureau of Shenzhen), and Hubei Natural Resources Research Project under Grant ZRZY2020KJ11. Institutional Review Board Statement: Not Applicable. Informed Consent Statement: Not Applicable. Data Availability Statement: The data presented in this study are available on request from the corresponding author. Conﬂicts of Interest: The authors declare no conﬂict of interest.

References 1. Roche, P.K.; Campagne, C.S. Are expert-based ecosystem services scores related to biophysical quantitative estimates? Ecol. Indic. 2019, 106, 105421. [CrossRef] 2. Araujo Barbosa, C.C.; Atkinson, P.M.; Dearing, J.A. Remote sensing of ecosystem services: A systematic review. Ecol. Indic. 2015, 52, 430–443. [CrossRef] 3. Maes, J.; Egoh, B.; Willemen, L.; Liquete, C.; Vihervaara, P.; Schägner, J.P.; Grizzetti, B.; Drakou, E.G.; Notte, A.L.; Zulian, G.; et al. Mapping ecosystem services for policy support and decision making in the European Union. Ecosyst. Serv. 2012, 1, 31–39. [CrossRef] 4. Lucas, P.L.; Kok, M.T.J.; Nilsson, M.; Alkemade, R. Integrating Biodiversity and Ecosystem Services in the Post-2015 Development Agenda: Goal Structure, Target Areas and Means of Implementation. Sustainability 2014, 6, 193–216. [CrossRef] 5. Jiang, W. Ecosystem services research in China: A critical review. Ecosyst. Serv. 2017, 26, 10–16. [CrossRef] 6. Deng, H.; Zheng, P.; Liu, T.; Liu, X. Forest Ecosystem Services and Eco-Compensation Mechanisms in China. Environ. Manag. 2011, 48, 1079–1085. [CrossRef][PubMed] 7. Bai, Y.; Jiang, B.; Wang, M.; Li, H.; Alatalo, J.M.; Huang, S. New ecological redline policy (ERP) to secure ecosystem services in China. Land Use Policy 2016, 55, 348–351. [CrossRef] 8. Costanza, R.; Groot, R.; Sutton, P.; Ploeg, S.; Anderson, S.J.; Kubiszewski, I.; Farber, S.; Turner, R.K. Changes in the global value of ecosystem services. Global Environ. Chang. 2014, 26, 152–158. [CrossRef] 9. Huang, X.; Han, X.; Ma, S.; Lin, T.; Gong, J. Monitoring ecosystem service change in the City of Shenzhen by the use of high-resolution remotely sensed imagery and deep learning. Land Degrad. Dev. 2019, 30, 1490–1501. [CrossRef] Sustainability 2021, 13, 7044 15 of 16

10. Haas, J.; Ban, Y. Mapping and monitoring urban ecosystem services using multitemporal high-resolution satellite data. IEEE J. Stars 2016, 10, 669–680. [CrossRef] 11. Xi, H.; Cui, W.; Cai, L.; Chen, M.; Xu, C. Evaluation and Prediction of Ecosystem Service Value in the Zhoushan Islands Based on LUCC. Sustainability 2021, 13, 2302. [CrossRef] 12. Fenta, A.A.; Tsunekawa, A.; Haregeweyn, N.; Tsubo, M.; Yasuda, H.; Shimizu, K.; Kawai, T.; Ebabu, K.; Berihun, M.L.; Sultan, D.; et al. Cropland expansion outweighs the monetary effect of declining natural vegetation on ecosystem services in sub-Saharan Africa. Ecosyst. Serv. 2020, 45, 101154. [CrossRef] 13. Hu, T.; Zhang, Y.; Su, Y.; Zheng, Y.; Lin, G.; Guo, Q. Mapping the global mangrove forest aboveground biomass using multisource remote sensing data. Remote Sens. 2020, 12, 1690. [CrossRef] 14. Trlica, A.; Hutyra, L.R.; Morreale, L.L.; Smith, I.A.; Reinmann, A.B. Current and future biomass carbon uptake in Boston’s urban forest. Sci. Total Environ. 2020, 709, 136196. [CrossRef][PubMed] 15. Burkhard, B.; Kroll, F.; Nedkov, S.; Müller, F. Mapping ecosystem service supply, demand and budgets. Ecol. Indic. 2012, 21, 17–29. [CrossRef] 16. Jacobs, S.; Burkhard, B.; Daele, T.V.; Staes, J.; Schneiders, A. The Matrix Reloaded: A review of expert knowledge use for mapping ecosystem services. Ecol. Model. 2015, 295, 21–30. [CrossRef] 17. Wu, X.; Liu, S.; Zhao, S.; Hou, X.; Xu, J.; Dong, S.; Liu, G. Quantification and driving force analysis of ecosystem services supply, demand and balance in China. Sci. Total Environ. 2019, 652, 1375–1386. [CrossRef] 18. Mouchet, M.A.; Paracchini, M.L.; Schulp, C.J.E.; Stürck, J.; Verkerk, P.J.; Verburg, P.H.; Lavorel, S. Bundles of ecosystem (dis)services and multifunctionality across European landscapes. Ecol. Indic. 2017, 73, 23–28. [CrossRef] 19. Stoll, S.; Frenzel, M.; Burkhard, B.; Adamescu, M.; Augustaitis, A.; Baeßler, C.; Bonet, F.J.; Carranza, M.L.; Cazacu, C.; Cosor, G.L.; et al. Assessment of ecosystem integrity and service gradients across Europe using the LTER Europe network. Ecol. Model. 2015, 295, 75–87. [CrossRef] 20. Van Tricht, K.; Gobin, A.; Gilliams, S.; Piccard, I. Synergistic use of radar Sentinel-1 and optical Sentinel-2 imagery for crop mapping: A case study for Belgium. Remote Sens. 2018, 10, 1642. [CrossRef] 21. Carrasco, L.; O’Neil, A.W.; Morton, R.D.; Rowland, C.S. Evaluating combinations of temporally aggregated Sentinel-1, Sentinel-2 and Landsat 8 for land cover mapping with Google Earth Engine. Remote Sens. 2019, 11, 288. [CrossRef] 22. Adrian, J.; Sagan, V.; Maimaitijiang, M. Sentinel SAR-optical fusion for crop type mapping using deep learning and Google Earth Engine. ISPRS J. Photogramm. 2021, 175, 215–235. [CrossRef] 23. Ienco, D.; Interdonato, R.; Gaetano, R.; Minh, D.H.T. Combining Sentinel-1 and Sentinel-2 Satellite Image Time Series for land cover mapping via a multi-source deep learning architecture. ISPRS J. Photogramm. 2019, 158, 11–22. [CrossRef] 24. Fang, B.; Li, Y.; Zhang, H.; Chan, J.C.-W. Collaborative learning of lightweight convolutional neural network and deep clustering for hyperspectral image semi-supervised classification with limited training samples. ISPRS J. Photogramm. 2020, 161, 164–178. [CrossRef] 25. Chong, Z.; Pan, S. Understanding the structure and determinants of city network through intra-firm service relationships: The case of Guangdong-Hong Kong-Macao Greater Bay Area. Cities 2020, 103, 102738. [CrossRef] 26. Gong, P.; Liu, H.; Zhang, M.; Li, C.; Wang, J.; Huang, H.; Clinton, N.; Ji, L.; Li, W.; Bai, Y.; et al. Stable classification with limited sample: Transferring a 30-m resolution sample set collected in 2015 to mapping 10 m resolution global land cover in 2017. Sci. Bull. 2019, 64, 370–373. [CrossRef] 27. Li, Y.; Chen, J.; Ma, Q.; Zhang, H.K.; Liu, J. Evaluation of Sentinel-2A surface reflectance derived using Sen2Cor in North America. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 1997–2021. [CrossRef] 28. Zhang, Z.; Liu, Q.; Wang, Y. Road Extraction by Deep Residual U-Net. IEEE Geosci. Remote Sens. 2018, 15, 749–753. [CrossRef] 29. Li, L.; Wang, C.; Zhang, H.; Zhang, B. Residual Unet for Urban Building Change Detection with Sentinel-1 SAR Data. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 28 July–2 August 2019. 30. Zhang, C.; Chen, Y.; Yang, X.; Gao, S.; Li, F.; Kong, A.; Zu, D.; Sun, L.; Kong, A. Improved remote sensing image classification based on multi-scale feature fusion. Remote Sens. 2020, 12, 213. [CrossRef] 31. Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the In- ternational Conference on Medical Image Computing and Computer-Assisted Intervention, Munich, Germany, 5–9 October 2015; Springer: Berlin/Heidelberg, Germany; pp. 234–241. 32. Saha, S.; Bovolo, F.; Bruzzone, L. Unsupervised Deep Change Vector Analysis for Multiple-Change Detection in VHR Images. IEEE T. Geosci. Remote 2019, 57, 3677–3693. [CrossRef] 33. Feng, R.; Wang, F.; Wang, K.; Xu, S. Quantifying influences of anthropogenic-natural factors on ecological land evolution in mega-urban agglomeration: A case study of Guangdong-Hong Kong-Macao greater Bay area. J. Clean. Prod. 2021, 283, 125304. [CrossRef] 34. Peng, J.; Zhao, M.; Guo, X.; Pan, Y.; Liu, Y. Spatial-temporal dynamics and associated driving forces of urban ecological land: A case study in Shenzhen City, China. Habitat Int. 2017, 60, 81–90. [CrossRef] 35. Wong, K.; Zhang, Y.; Tsou, J.Y.; Li, Y. Assessing impervious surface changes in sustainable coastal land use: A case study in Hong Kong. Sustainability 2017, 9, 1029. [CrossRef] 36. Zhou, Y.; Shan, Y.; Liu, G.; Guan, D. Emissions and low-carbon development in Guangdong-Hong Kong-Macao Greater Bay Area cities and their surroundings. Appl. Energy 2018, 228, 1683–1692. [CrossRef] Sustainability 2021, 13, 7044 16 of 16

37. Li, H.; Jin, R.; Ning, X.; Skitmore, M.; Zhang, T. Prioritizing the Sustainability Objectives of Major Public Projects in the Guangdong–Hong Kong–Macao Greater Bay Area. Sustainability 2018, 10, 4110. [CrossRef] 38. Zhou, W.; Mu, R. Exploring Coordinative Mechanisms for Environmental Governance in Guangdong-Hong Kong-Macao Greater Bay Area: An Ecology of Games Framework. Sustainability 2019, 11, 3119. [CrossRef] 39. Xie, Y.; Eftelioglu, E.; Ali, R.Y.; Tang, X.; Li, Y.; Doshi, R.; Shekhar, S. Transdisciplinary Foundations of Geospatial Data Science. ISPRS Int. J. Geo-Inf. 2017, 6, 395. [CrossRef] 40. Atluri, G.; Karpatne, A.; Kumar, V. Spatio-temporal data mining: A survey of problems and methods. ACM Comput. Surv. 2018, 51, 1–41. [CrossRef]