Bike lanes are a critical element of urban infrastructure that promote cycling and support sustainable transportation goals. Effective planning and evaluation require comprehensive inventory datasets that both identify the locations of bike lanes and classify their types. However, existing data collection is limited by inconsistent municipal documentation practices and resource constraints. This paper introduces a computer vision–based approach for the automated detection and classification of bike lanes using publicly available multimodal imagery. Each data sample integrates two street view images, captured from opposite directions, with a corresponding satellite image, enabling complementary perspectives. This approach allows the model to reliably detect bike lane presence and distinguish between designated (marked lanes without physical barriers) and protected (lanes separated from traffic by physical barriers) types. To optimize performance, we conduct ablation experiments across three architectural dimensions: stage of modality concatenation, fusion strategy, and label structure. We also construct a training dataset using Google Street View and satellite imagery from 28 major U.S. cities to ensure broad applicability. Applying the model to over 1000 road segments in Atlanta, Georgia, we demonstrate its scalability and accuracy in a real-world urban setting. By providing an automated, transferable method for developing bike lane inventories, this research addresses a critical gap in infrastructure documentation and supports more effective planning of bicycle networks.
{"title":"Automated detection and classification of bike lanes using multimodal imagery","authors":"Seung Jae Lieu , Bon Woo Koo , Uijeong Hwang , Subhrajit Guhathakurta","doi":"10.1016/j.rsase.2025.101817","DOIUrl":"10.1016/j.rsase.2025.101817","url":null,"abstract":"<div><div>Bike lanes are a critical element of urban infrastructure that promote cycling and support sustainable transportation goals. Effective planning and evaluation require comprehensive inventory datasets that both identify the locations of bike lanes and classify their types. However, existing data collection is limited by inconsistent municipal documentation practices and resource constraints. This paper introduces a computer vision–based approach for the automated detection and classification of bike lanes using publicly available multimodal imagery. Each data sample integrates two street view images, captured from opposite directions, with a corresponding satellite image, enabling complementary perspectives. This approach allows the model to reliably detect bike lane presence and distinguish between designated (marked lanes without physical barriers) and protected (lanes separated from traffic by physical barriers) types. To optimize performance, we conduct ablation experiments across three architectural dimensions: stage of modality concatenation, fusion strategy, and label structure. We also construct a training dataset using Google Street View and satellite imagery from 28 major U.S. cities to ensure broad applicability. Applying the model to over 1000 road segments in Atlanta, Georgia, we demonstrate its scalability and accuracy in a real-world urban setting. By providing an automated, transferable method for developing bike lane inventories, this research addresses a critical gap in infrastructure documentation and supports more effective planning of bicycle networks.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"41 ","pages":"Article 101817"},"PeriodicalIF":4.5,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145665528","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-02DOI: 10.1016/j.rsase.2025.101814
Anita Gautam, Bharath Haridas Aithal
Rapid urbanization fundamentally restructures metropolitan form through non-linear, scale-dependent processes that reorganize spatial hierarchy and land-use configuration. This study examines Bengaluru's morphological evolution from 2012 to 2023 using high-resolution satellite imagery, interpreted through deep learning–based classification, spatial metrics, and fractal geometry, to quantify structural and scaling transformations. The results reveal a decisive transition from fragmented, spatially spread-led expansion to a spatially integrated yet hierarchically differentiated urban system. The improvements in Patch cohesion, surface occupancy, and insignificants in landscape fragmentation and irregular edge patterns indicate urban growth in-fill (or redevelopment) and further development along corridors. The fractal dimension increased from 1.78 to 1.91, indicating an improvement in filling space through increased compactness and geometric order, whereas the multifractal spectrum increased from 1.41 to 1.92, demonstrating an increase in structural heterogeneity over a range of scales. A positive relationship (r = 0.68) between patch cohesion and fractal compactness quantitatively confirms the association of local aggregation/compactness with global geometric order. Overall, these findings illustrate the hierarchical scaling organization of urban growth where compactness and heterogeneity co-evolve through self-organizing spatial logic. By incorporating metric-based morphological analysis to fractal scaling, the framework enhances urban theory, proposing a scale-consistent account of spatial evolution. This account describes how urban systems transition from dispersed growth to geometrically ordered and hierarchically structured forms.
{"title":"Urban structural complexity in transition: Fractal analysis of deep learning-derived morphological patterns","authors":"Anita Gautam, Bharath Haridas Aithal","doi":"10.1016/j.rsase.2025.101814","DOIUrl":"10.1016/j.rsase.2025.101814","url":null,"abstract":"<div><div>Rapid urbanization fundamentally restructures metropolitan form through non-linear, scale-dependent processes that reorganize spatial hierarchy and land-use configuration. This study examines Bengaluru's morphological evolution from 2012 to 2023 using high-resolution satellite imagery, interpreted through deep learning–based classification, spatial metrics, and fractal geometry, to quantify structural and scaling transformations. The results reveal a decisive transition from fragmented, spatially spread-led expansion to a spatially integrated yet hierarchically differentiated urban system. The improvements in Patch cohesion, surface occupancy, and insignificants in landscape fragmentation and irregular edge patterns indicate urban growth in-fill (or redevelopment) and further development along corridors. The fractal dimension increased from 1.78 to 1.91, indicating an improvement in filling space through increased compactness and geometric order, whereas the multifractal spectrum increased from 1.41 to 1.92, demonstrating an increase in structural heterogeneity over a range of scales. A positive relationship (r = 0.68) between patch cohesion and fractal compactness quantitatively confirms the association of local aggregation/compactness with global geometric order. Overall, these findings illustrate the hierarchical scaling organization of urban growth where compactness and heterogeneity co-evolve through self-organizing spatial logic. By incorporating metric-based morphological analysis to fractal scaling, the framework enhances urban theory, proposing a scale-consistent account of spatial evolution. This account describes how urban systems transition from dispersed growth to geometrically ordered and hierarchically structured forms.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"41 ","pages":"Article 101814"},"PeriodicalIF":4.5,"publicationDate":"2025-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145685600","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-22DOI: 10.1016/j.rsase.2025.101804
Sahar Khoshnoud , S. Mohammad Mirmazloumi , Arsalan Ghorbanian , Hossein Mohammad Asgari , Meisam Amani
Investigating the potential for wind-induced soil erosion in arid and semi-arid regions is essential for understanding soil degradation and its associated impacts, such as agricultural productivity reduction, infrastructure damage, air quality decline, and adverse health effects. This study pioneers the integration of remote sensing data and Artificial Neural Networks (ANN) for wind erosion mapping, offering a novel approach to analyzing soil surface dynamics. ANN models were implemented to estimate aerodynamic roughness (z0) and friction velocity () using Sentinel-1 Synthetic Aperture Radar (SAR) data. These estimates were further integrated with meteorological datasets to identify areas prone to wind erosion and, subsequently, dust storms. The results indicated that wetlands, with the highest (6.98 cm) and (0.81 m/s) values have a negligible potential for wind erosion. Conversely, clay flats showed the lowest values ( = 0.89 cm, = 0.42 m/s), suggesting a higher susceptibility to wind erosion. Finally, the developed model was applied to generate wind erosion potential maps of the study area, serving as a practical asset for the identification of high-risk zones prone to erosion. This study emphasizes the importance of soil surface parameters to identify potential areas of wind erosion for developing more accurate dust emission models, which support effective management of wind erosion and mitigate the adverse effects of this environmental phenomenon. Although regionally focused, the methodology is transferable to other arid and semi-arid environments, offering valuable insights for soil conservation and land management.
{"title":"Mapping wind erosion potential using remote sensing and artificial neural networks: Insights for soil conservation in arid regions","authors":"Sahar Khoshnoud , S. Mohammad Mirmazloumi , Arsalan Ghorbanian , Hossein Mohammad Asgari , Meisam Amani","doi":"10.1016/j.rsase.2025.101804","DOIUrl":"10.1016/j.rsase.2025.101804","url":null,"abstract":"<div><div>Investigating the potential for wind-induced soil erosion in arid and semi-arid regions is essential for understanding soil degradation and its associated impacts, such as agricultural productivity reduction, infrastructure damage, air quality decline, and adverse health effects. This study pioneers the integration of remote sensing data and Artificial Neural Networks (ANN) for wind erosion mapping, offering a novel approach to analyzing soil surface dynamics. ANN models were implemented to estimate aerodynamic roughness (z<sub>0</sub>) and friction velocity (<span><math><mrow><msub><mi>u</mi><mo>∗</mo></msub></mrow></math></span>) using Sentinel-1 Synthetic Aperture Radar (SAR) data. These estimates were further integrated with meteorological datasets to identify areas prone to wind erosion and, subsequently, dust storms. The results indicated that wetlands, with the highest <span><math><mrow><msub><mi>z</mi><mn>0</mn></msub></mrow></math></span> (6.98 cm) and <span><math><mrow><msub><mi>u</mi><mo>∗</mo></msub></mrow></math></span> (0.81 m/s) values have a negligible potential for wind erosion. Conversely, clay flats showed the lowest values (<span><math><mrow><msub><mi>z</mi><mn>0</mn></msub></mrow></math></span> = 0.89 cm, <span><math><mrow><msub><mi>u</mi><mo>∗</mo></msub></mrow></math></span> = 0.42 m/s), suggesting a higher susceptibility to wind erosion. Finally, the developed model was applied to generate wind erosion potential maps of the study area, serving as a practical asset for the identification of high-risk zones prone to erosion. This study emphasizes the importance of soil surface parameters to identify potential areas of wind erosion for developing more accurate dust emission models, which support effective management of wind erosion and mitigate the adverse effects of this environmental phenomenon. Although regionally focused, the methodology is transferable to other arid and semi-arid environments, offering valuable insights for soil conservation and land management.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"41 ","pages":"Article 101804"},"PeriodicalIF":4.5,"publicationDate":"2025-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145738347","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-01DOI: 10.1016/j.rsase.2025.101784
Jiaxin Wang , Xinyi Li , Haoran Liu , Yani Wang , Xin Zhang , Donghui Song
Coastal wetlands are essential for mitigating climate change but face significant challenges in carbon storage assessment due to spatial-scale constraints and the application of oversimplified models that fail to capture complex spatiotemporal dynamics and driving mechanisms. This study addresses two key gaps in understanding carbon sink degradation in coastal land-sea interface systems: (1) the insufficient analysis of unidirectional natural factors and (2) the inability of models to capture spatial heterogeneity in carbon sink degradation. Taking Bohai Bay (China) as a case study, we developed an integrated InVEST-PLUS-GeoDetector framework to reconstruct and project the coastal wetland carbon storage evolution since 1980. Key findings include: (1) A 15.7 % net decline in carbon storage (from 58.11 ± 2.87 Tg in 1980 to 48.98 ± 2.14 Tg in 2020), driven primarily by constructed wetlands expansion encroaching on natural wetlands; (2) GeoDetector analysis identified vegetation coverage (q = 0.38), soil type (q = 0.25), distance to coastline (q = 0.24), and GDP (q = 0.18) as dominant drivers of carbon storage variation, with vegetation-soil interactions being the most influential; (3) Multi-scenario simulations revealed that wetland conservation policies could significantly increase carbon storage by 0.25 Tg by 2050 (exceeding model uncertainty), a 1.48-fold enhancement compared to the economic development scenario, attributable to the preservation of high-carbon-density natural wetlands despite their slower sequestration rates. The proposed framework effectively addresses the two key gaps by capturing key driver couplings (natural-socioeconomic) and spatial heterogeneity in carbon dynamics. Our findings advance the understanding of human-environment interactions in intensely developed coastal zones and provide practical pathways for synergizing wetland conservation and carbon sink enhancement in semi-enclosed marine systems.
{"title":"Balancing development and carbon storage: Spatiotemporal heterogeneity of Bohai Bay's coastal wetlands under socio-ecological drivers","authors":"Jiaxin Wang , Xinyi Li , Haoran Liu , Yani Wang , Xin Zhang , Donghui Song","doi":"10.1016/j.rsase.2025.101784","DOIUrl":"10.1016/j.rsase.2025.101784","url":null,"abstract":"<div><div>Coastal wetlands are essential for mitigating climate change but face significant challenges in carbon storage assessment due to spatial-scale constraints and the application of oversimplified models that fail to capture complex spatiotemporal dynamics and driving mechanisms. This study addresses two key gaps in understanding carbon sink degradation in coastal land-sea interface systems: (1) the insufficient analysis of unidirectional natural factors and (2) the inability of models to capture spatial heterogeneity in carbon sink degradation. Taking Bohai Bay (China) as a case study, we developed an integrated InVEST-PLUS-GeoDetector framework to reconstruct and project the coastal wetland carbon storage evolution since 1980. Key findings include: (1) A 15.7 % net decline in carbon storage (from 58.11 ± 2.87 Tg in 1980 to 48.98 ± 2.14 Tg in 2020), driven primarily by constructed wetlands expansion encroaching on natural wetlands; (2) GeoDetector analysis identified vegetation coverage (q = 0.38), soil type (q = 0.25), distance to coastline (q = 0.24), and GDP (q = 0.18) as dominant drivers of carbon storage variation, with vegetation-soil interactions being the most influential; (3) Multi-scenario simulations revealed that wetland conservation policies could significantly increase carbon storage by 0.25 Tg by 2050 (exceeding model uncertainty), a 1.48-fold enhancement compared to the economic development scenario, attributable to the preservation of high-carbon-density natural wetlands despite their slower sequestration rates. The proposed framework effectively addresses the two key gaps by capturing key driver couplings (natural-socioeconomic) and spatial heterogeneity in carbon dynamics. Our findings advance the understanding of human-environment interactions in intensely developed coastal zones and provide practical pathways for synergizing wetland conservation and carbon sink enhancement in semi-enclosed marine systems.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"40 ","pages":"Article 101784"},"PeriodicalIF":4.5,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145474525","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-01DOI: 10.1016/j.rsase.2025.101810
Yuanrong He , Xiajing Meng , Liheng Zhang , Lefan Wang , Tianqi Yang , Guoliang Yun
Urban agglomerations (UAs) serve as key units for new-type urbanization, where spatial heterogeneity and development disparities significantly impact regional coordination efforts. Despite abundant literature on UAs, comparative analyses of long-term urbanization patterns across multiple UAs and their differentiated driving mechanisms remain insufficiently explored, especially regarding the driving mechanisms of development disparities within and between UAs. This study employs dynamic time warping (DTW), the Dagum Gini coefficient, and partial least squares (PLS) regression to analyze four UAs. Results show significant spatial-temporal heterogeneity: (1) High-urbanization areas cluster in established metropolises, while low-level regions concentrate in Beijing-Tianjin-Hebei (BTH) and Chengdu-Chongqing (CY). Most cities (72.15 %) exhibit “Recent Urban Growth,” with BTH dominated by “Constant Urban Growth” and metropolises showing “Early Urban Growth”. (2) Overall urbanization disparities declined with the development gap narrowing by 35.56 % over 30 years (from 0.503 to 0.324), driven by inter-regional unbalance but shifted to intra-regional density gaps recently. (3) Drivers vary regionally: resource endowment amplifies disparities, while population agglomeration mitigates them; technological innovation increases disparities in the Pearl River Delta (PRD, 1.038) but reduces them elsewhere (−0.329 to −0.208). The study emphasizes stage-specific and region-specific effects of factors, advocating tailored sustainable urbanization strategies to address each UA's developmental characteristics.
{"title":"Assessing urbanization differentiation and socioeconomic drivers in China's four major urban agglomerations based on nighttime light data (1992–2021)","authors":"Yuanrong He , Xiajing Meng , Liheng Zhang , Lefan Wang , Tianqi Yang , Guoliang Yun","doi":"10.1016/j.rsase.2025.101810","DOIUrl":"10.1016/j.rsase.2025.101810","url":null,"abstract":"<div><div>Urban agglomerations (UAs) serve as key units for new-type urbanization, where spatial heterogeneity and development disparities significantly impact regional coordination efforts. Despite abundant literature on UAs, comparative analyses of long-term urbanization patterns across multiple UAs and their differentiated driving mechanisms remain insufficiently explored, especially regarding the driving mechanisms of development disparities within and between UAs. This study employs dynamic time warping (DTW), the Dagum Gini coefficient, and partial least squares (PLS) regression to analyze four UAs. Results show significant spatial-temporal heterogeneity: (1) High-urbanization areas cluster in established metropolises, while low-level regions concentrate in Beijing-Tianjin-Hebei (BTH) and Chengdu-Chongqing (CY). Most cities (72.15 %) exhibit “Recent Urban Growth,” with BTH dominated by “Constant Urban Growth” and metropolises showing “Early Urban Growth”. (2) Overall urbanization disparities declined with the development gap narrowing by 35.56 % over 30 years (from 0.503 to 0.324), driven by inter-regional unbalance but shifted to intra-regional density gaps recently. (3) Drivers vary regionally: resource endowment amplifies disparities, while population agglomeration mitigates them; technological innovation increases disparities in the Pearl River Delta (PRD, 1.038) but reduces them elsewhere (−0.329 to −0.208). The study emphasizes stage-specific and region-specific effects of factors, advocating tailored sustainable urbanization strategies to address each UA's developmental characteristics.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"40 ","pages":"Article 101810"},"PeriodicalIF":4.5,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145623635","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
With its complex hydrological dynamics and high diversity of habitats, the Pantanal, a large South American wetland, is increasingly threatened by anthropogenic activities and climate change. In such vulnerable ecosystems, radar altimetry is a key source of Water Surface Elevation (WSE) measurements and, thus, fundamental for monitoring these often remote and poorly gauged environments. This study assesses the potential of the Sentinel-3 and Sentinel-6 missions to monitor hydrological dynamics in the Brazilian Pantanal (May 2016 to March 2024). We first evaluated the agreement between radar altimetry and in situ WSE and then demonstrated the contribution of these data to characterizing three distinct hydrological features across this large tropical wetland. Our results showed strong agreement between altimetry and in situ water levels, with correlation coefficients (R) greater than 0.85 and Root Mean Square Errors (RMSE) below 0.4 m in most cases. Using this extensive radar altimetry network, we demonstrated backwater flooding on the Miranda River, which experiences two annual flood cycles driven by local precipitation and the Paraguay River’s flood pulse. This dynamic was disrupted by recent megadroughts. We also detected significant WSE declines in the shallow lakes of the Nhecolândia region, directly linked to the megadroughts, and revealed along-channel variations in seasonal water level patterns across the ungauged Taquari megafan, a distributive fluvial system likely subject to the combined pressures of upland agriculture and climatic extremes. These findings underscore the high potential of radar altimetry for monitoring and understanding complex hydrologic dynamics in vulnerable ecosystems like the Pantanal.
{"title":"Monitoring water surface elevation dynamics in the Brazilian Pantanal wetland using radar altimetry","authors":"Uelison Mateus Ribeiro , Samuel Corgne , Manuela Grippa , Félix Girard , Sly Wongchuig , Carolina Joana da Silva , Vitor Matheus Bacani , Mauro Henrique Soares da Silva , Frederico Gradella , Damien Arvor","doi":"10.1016/j.rsase.2025.101805","DOIUrl":"10.1016/j.rsase.2025.101805","url":null,"abstract":"<div><div>With its complex hydrological dynamics and high diversity of habitats, the Pantanal, a large South American wetland, is increasingly threatened by anthropogenic activities and climate change. In such vulnerable ecosystems, radar altimetry is a key source of Water Surface Elevation (WSE) measurements and, thus, fundamental for monitoring these often remote and poorly gauged environments. This study assesses the potential of the Sentinel-3 and Sentinel-6 missions to monitor hydrological dynamics in the Brazilian Pantanal (May 2016 to March 2024). We first evaluated the agreement between radar altimetry and in situ WSE and then demonstrated the contribution of these data to characterizing three distinct hydrological features across this large tropical wetland. Our results showed strong agreement between altimetry and in situ water levels, with correlation coefficients (R) greater than 0.85 and Root Mean Square Errors (RMSE) below 0.4 m in most cases. Using this extensive radar altimetry network, we demonstrated backwater flooding on the Miranda River, which experiences two annual flood cycles driven by local precipitation and the Paraguay River’s flood pulse. This dynamic was disrupted by recent megadroughts. We also detected significant WSE declines in the shallow lakes of the Nhecolândia region, directly linked to the megadroughts, and revealed along-channel variations in seasonal water level patterns across the ungauged Taquari megafan, a distributive fluvial system likely subject to the combined pressures of upland agriculture and climatic extremes. These findings underscore the high potential of radar altimetry for monitoring and understanding complex hydrologic dynamics in vulnerable ecosystems like the Pantanal.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"40 ","pages":"Article 101805"},"PeriodicalIF":4.5,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145623738","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Predicting winter wheat yields is necessary for sustainable global farming and ensuring food supply. This study introduced a multi–task transfer learning framework based on a hybrid 3D–ResNet–BiLSTM architecture to predict county–level winter wheat yields in the U.S. using multi–source remote sensing (RS) data. The primary goal was to investigate how the choice of source crop (i.e., corn, soybean, or their combination) influences transfer learning performance for winter wheat yield prediction. To achieve this, two–stage modeling framework was used. First, a multi–task 3D–ResNet–BiLSTM model (3D–ResNet–BiLSTM–MT) was trained on corn and soybean yield data from 2016 to 2020, leveraging their overlapping growing seasons to capture shared spatio–temporal representations. Second, a fine–tuned transfer model (Transfer–3D–ResNet–BiLSTM) was developed using limited winter wheat data (2018-2020). During fine–tuning, the feature extraction layers were frozen, reducing trainable parameters by ∼50 % and enhancing robustness under data–scarce conditions. The models integrated multi–source inputs from Sentinel–1/2 imagery, Daymet weather variables, and SoilGrids, and were evaluated on independent test data (2021- 2022). The multi–task model efficiently predicted corn and soybean yields, achieving an R2 of 0.78. For winter wheat, the corn–based transfer model achieved the highest performance (RMSE = 9.63, MAE = 7.61, MAPE = 13.23, R2 = 0.75), followed by the soybean–based model (R2 = 0.69). In contrast, the shared corn–soybean model (the best–performing model trained specifically on both crops using 3D–ResNet–BiLSTM–MT) underperformed (R2 = 0.63), while the baseline wheat–only model without transfer learning showed the weakest performance (RMSE = 12.24, R2 = 0.60). Overall, the source-specific transfer models (corn- and soybean-based) outperformed both the wheat-only deep learning baseline and conventional machine learning models (RF, SVM, XGBoost, and LightGBM), demonstrating the strong generalization ability and data efficiency of deep transfer learning for yield prediction. These findings highlight the importance of source crop selection and show that cross–crop transfer learning is a practical, data–efficient, and generalizable approach for yield prediction, especially valuable where labeled data are scarce.
{"title":"A Multi–Task 3D–ResNet–BiLSTM transfer learning approach for winter wheat yield prediction using multi–source remote sensing data: Evaluating the impact of source crop selection (corn and soybean) in transfer learning","authors":"Mahdiyeh Fathi , Reza Shah–Hosseini , Hossein Arefi , Armin Moghimi","doi":"10.1016/j.rsase.2025.101766","DOIUrl":"10.1016/j.rsase.2025.101766","url":null,"abstract":"<div><div>Predicting winter wheat yields is necessary for sustainable global farming and ensuring food supply. This study introduced a multi–task transfer learning framework based on a hybrid 3D–ResNet–BiLSTM architecture to predict county–level winter wheat yields in the U.S. using multi–source remote sensing (RS) data. The primary goal was to investigate how the choice of source crop (i.e., corn, soybean, or their combination) influences transfer learning performance for winter wheat yield prediction. To achieve this, two–stage modeling framework was used. First, a multi–task 3D–ResNet–BiLSTM model (3D–ResNet–BiLSTM–MT) was trained on corn and soybean yield data from 2016 to 2020, leveraging their overlapping growing seasons to capture shared spatio–temporal representations. Second, a fine–tuned transfer model (Transfer–3D–ResNet–BiLSTM) was developed using limited winter wheat data (2018-2020). During fine–tuning, the feature extraction layers were frozen, reducing trainable parameters by ∼50 % and enhancing robustness under data–scarce conditions. The models integrated multi–source inputs from Sentinel–1/2 imagery, Daymet weather variables, and SoilGrids, and were evaluated on independent test data (2021- 2022). The multi–task model efficiently predicted corn and soybean yields, achieving an R<sup>2</sup> of 0.78. For winter wheat, the corn–based transfer model achieved the highest performance (RMSE = 9.63, MAE = 7.61, MAPE = 13.23, R<sup>2</sup> = 0.75), followed by the soybean–based model (R<sup>2</sup> = 0.69). In contrast, the shared corn–soybean model (the best–performing model trained specifically on both crops using 3D–ResNet–BiLSTM–MT) underperformed (R<sup>2</sup> = 0.63), while the baseline wheat–only model without transfer learning showed the weakest performance (RMSE = 12.24, R<sup>2</sup> = 0.60). Overall, the source-specific transfer models (corn- and soybean-based) outperformed both the wheat-only deep learning baseline and conventional machine learning models (RF, SVM, XGBoost, and LightGBM), demonstrating the strong generalization ability and data efficiency of deep transfer learning for yield prediction. These findings highlight the importance of source crop selection and show that cross–crop transfer learning is a practical, data–efficient, and generalizable approach for yield prediction, especially valuable where labeled data are scarce.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"40 ","pages":"Article 101766"},"PeriodicalIF":4.5,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145417690","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-01DOI: 10.1016/j.rsase.2025.101776
Konstantinos Chatzopoulos-Vouzoglanis , Karin J. Reinke , Mariela Soto-Berelov , Simon D. Jones
Current wildfire impact assessments at the landscape scale often overlook the complexity of active fire behaviour, focusing only on pre- and post-fire spectral differencing, despite remotely sensed active fire data being readily available. This study integrates high temporal resolution active fire intensity measures from geostationary satellite sensors and high spatial resolution normalised spectral differencing index products from polar-orbiting satellite sensors to produce a new approach for describing wildfire impact. Himawari-8 BRIGHT/AHI Fire Radiative Power (FRP) estimates are combined with Normalised Burn Ratio (NBR) metrics from Sentinel-2, to derive wildfire impact categories over Australia for one year of data, using a clustering approach. The wildfire impact categories summarise fire hotspot commonalities based on their maximum and total FRP, duration, differenced NBR (dNBR), burned area patchiness, and pre-fire NBR, and reveal expected 2019–2020 Australian fire season patterns. Furthermore, land cover emerges as an important factor, with forests and woodlands reflecting higher impact fires compared to grasslands and shrublands. Our wildfire impact categories show a moderate agreement with burn severity assessments conducted by state governments, further stressing the need for more diverse information inclusion in such assessments. The proposed composite wildfire impact rating combines diverse remotely sensed wildfire behaviour information and can assist in a better understanding of wildfire effects on a continental scale. More research, leveraging longer temporal and spatial baselines and fire ecology expertise, is needed to refine the used nomenclature for the improvement of wildfire impact assessments.
{"title":"Composite Wildfire Impact (CWI) rating: Integrating fire intensity and burn severity earth observations","authors":"Konstantinos Chatzopoulos-Vouzoglanis , Karin J. Reinke , Mariela Soto-Berelov , Simon D. Jones","doi":"10.1016/j.rsase.2025.101776","DOIUrl":"10.1016/j.rsase.2025.101776","url":null,"abstract":"<div><div>Current wildfire impact assessments at the landscape scale often overlook the complexity of active fire behaviour, focusing only on pre- and post-fire spectral differencing, despite remotely sensed active fire data being readily available. This study integrates high temporal resolution active fire intensity measures from geostationary satellite sensors and high spatial resolution normalised spectral differencing index products from polar-orbiting satellite sensors to produce a new approach for describing wildfire impact. Himawari-8 BRIGHT/AHI Fire Radiative Power (FRP) estimates are combined with Normalised Burn Ratio (NBR) metrics from Sentinel-2, to derive wildfire impact categories over Australia for one year of data, using a clustering approach. The wildfire impact categories summarise fire hotspot commonalities based on their maximum and total FRP, duration, differenced NBR (dNBR), burned area patchiness, and pre-fire NBR, and reveal expected 2019–2020 Australian fire season patterns. Furthermore, land cover emerges as an important factor, with forests and woodlands reflecting higher impact fires compared to grasslands and shrublands. Our wildfire impact categories show a moderate agreement with burn severity assessments conducted by state governments, further stressing the need for more diverse information inclusion in such assessments. The proposed composite wildfire impact rating combines diverse remotely sensed wildfire behaviour information and can assist in a better understanding of wildfire effects on a continental scale. More research, leveraging longer temporal and spatial baselines and fire ecology expertise, is needed to refine the used nomenclature for the improvement of wildfire impact assessments.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"40 ","pages":"Article 101776"},"PeriodicalIF":4.5,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145424390","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-01DOI: 10.1016/j.rsase.2025.101764
Ewa Kopec , Agata M. Wijata , Jakub Nalepa
The recent advancements in satellite imaging bring various possibilities in Earth observation in numerous domains, including the analysis of the evolution of urban areas, precision agriculture, environmental monitoring, event detection and tracking, and many more. Change detection plays a key role in a multitude of applications, as it allows for precisely monitoring the changes within an area of interest. In this article, we tackle this issue and introduce deep learning ensembles for change detection in Sentinel-2 times series of multispectral images—the proposed ensembles benefit from different deep learning model architectures. The experimental study performed over the widely-adopted benchmark datasets showed that the ensembles combine the strengths of the individual models, thus they reduce false positives and false negatives of base learners. The ensembles compensated the under-performing models, ultimately obtaining the change detection accuracy that exceeds 95% over the unseen test scenes.
{"title":"Change detection in Sentinel-2 images using deep learning ensembles","authors":"Ewa Kopec , Agata M. Wijata , Jakub Nalepa","doi":"10.1016/j.rsase.2025.101764","DOIUrl":"10.1016/j.rsase.2025.101764","url":null,"abstract":"<div><div>The recent advancements in satellite imaging bring various possibilities in Earth observation in numerous domains, including the analysis of the evolution of urban areas, precision agriculture, environmental monitoring, event detection and tracking, and many more. Change detection plays a key role in a multitude of applications, as it allows for precisely monitoring the changes within an area of interest. In this article, we tackle this issue and introduce deep learning ensembles for change detection in Sentinel-2 times series of multispectral images—the proposed ensembles benefit from different deep learning model architectures. The experimental study performed over the widely-adopted benchmark datasets showed that the ensembles combine the strengths of the individual models, thus they reduce false positives and false negatives of base learners. The ensembles compensated the under-performing models, ultimately obtaining the change detection accuracy that exceeds 95% over the unseen test scenes.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"40 ","pages":"Article 101764"},"PeriodicalIF":4.5,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145424475","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-01DOI: 10.1016/j.rsase.2025.101798
Kirill Korznikov , Dmitriy Kislov , Jiří Doležal , Jan Altman
Wind- and insect-induced forest disturbances are becoming increasingly frequent and severe due to climate change, resulting in significant forest loss worldwide. Accurate detection of these disturbances is essential for understanding carbon storage, forest dynamics, ecosystem resilience, and for developing effective climate adaptation strategies. The Landsat-based Global Forest Change (GFC) product and the related Global Forest Watch web service are widely used for large-scale forest monitoring. However, its capacity to detect disturbances caused by wind and insect outbreaks remains uncertain. In this study, we assessed the accuracy of GFC forest loss detection by comparing it with U-Net neural network forest loss masks derived from very high-resolution satellite imagery. We analyzed several study areas in natural mixed-species southern boreal forests affected by windthrows and bark beetle outbreaks, evaluating true positive (TP), false negative (FN), and false positive (FP) detection rates. Our results show that GFC substantially underestimates forest loss, with TP detection rates ranging from 1.56 % to 62.18 % and FN errors reaching 37.82 %–98.44 %. In some cases, overestimation occurred due to high FP rates up to 65.55 %. The FPs happen when a small patch of forest loss within a 30 × 30 m Landsat pixel triggers the entire pixel to be classified as forest loss, even though most of the pixel remains undisturbed. The limitation stems from the small-scale nature of windthrows and insect-induced diebacks, which cannot be reliably captured by Landsat's spatial resolution. Our findings suggest that integrating higher-resolution satellite data is crucial for accurate area estimation and improved assessments of forest loss in the face of climate-driven disturbances such as windthrows and diebacks in natural mixed-species forests. Although GFC can be unsuitable for precisely mapping forest losses, it remains a valuable, globally consistent early warning tool due to its annual updates and broad coverage. Practitioners should treat GFC detections as indicative and conduct rapid visual or automated checks with higher-resolution imagery when assessing windthrow or insect-driven mortality, especially when disturbance patches are less than 450 m2.
{"title":"Hidden forest loss: challenges in detecting wind- and insect-driven forest disturbances with global forest change landsat-based products in mixed southern boreal forests","authors":"Kirill Korznikov , Dmitriy Kislov , Jiří Doležal , Jan Altman","doi":"10.1016/j.rsase.2025.101798","DOIUrl":"10.1016/j.rsase.2025.101798","url":null,"abstract":"<div><div>Wind- and insect-induced forest disturbances are becoming increasingly frequent and severe due to climate change, resulting in significant forest loss worldwide. Accurate detection of these disturbances is essential for understanding carbon storage, forest dynamics, ecosystem resilience, and for developing effective climate adaptation strategies. The Landsat-based Global Forest Change (GFC) product and the related Global Forest Watch web service are widely used for large-scale forest monitoring. However, its capacity to detect disturbances caused by wind and insect outbreaks remains uncertain. In this study, we assessed the accuracy of GFC forest loss detection by comparing it with U-Net neural network forest loss masks derived from very high-resolution satellite imagery. We analyzed several study areas in natural mixed-species southern boreal forests affected by windthrows and bark beetle outbreaks, evaluating true positive (TP), false negative (FN), and false positive (FP) detection rates. Our results show that GFC substantially underestimates forest loss, with TP detection rates ranging from 1.56 % to 62.18 % and FN errors reaching 37.82 %–98.44 %. In some cases, overestimation occurred due to high FP rates up to 65.55 %. The FPs happen when a small patch of forest loss within a 30 × 30 m Landsat pixel triggers the entire pixel to be classified as forest loss, even though most of the pixel remains undisturbed. The limitation stems from the small-scale nature of windthrows and insect-induced diebacks, which cannot be reliably captured by Landsat's spatial resolution. Our findings suggest that integrating higher-resolution satellite data is crucial for accurate area estimation and improved assessments of forest loss in the face of climate-driven disturbances such as windthrows and diebacks in natural mixed-species forests. Although GFC can be unsuitable for precisely mapping forest losses, it remains a valuable, globally consistent early warning tool due to its annual updates and broad coverage. Practitioners should treat GFC detections as indicative and conduct rapid visual or automated checks with higher-resolution imagery when assessing windthrow or insect-driven mortality, especially when disturbance patches are less than 450 m<sup>2</sup>.</div></div>","PeriodicalId":53227,"journal":{"name":"Remote Sensing Applications-Society and Environment","volume":"40 ","pages":"Article 101798"},"PeriodicalIF":4.5,"publicationDate":"2025-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145578726","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号