Pub Date : 2025-10-30DOI: 10.1016/j.srs.2025.100312
Yongcai Wang , Huawei Wan , Dongpo Wang , Jixi Gao , Zhuowei Hu , Junjie Wang , Fengming Wan , Bin Yang , Zhiru Zhang , Ma Bian , Jiqian Zhou
Grassland canopy cover acts as an essential metric for gauging the vitality and ecological functions of grassland. Unmanned aerial vehicles (UAVs) provide stable and reliable data for estimating grassland canopy cover. However, conventional approaches primarily rely on samples from ground surveys and visual assessments, where data consistency is often affected by variations in survey techniques and personnel expertise. By contrast, UAVs provide consistent multi-scale grassland canopy data. Thus, effectively harnessing the strengths of multi-scale UAV imagery can markedly improve the efficiency and precision of canopy cover estimation. This study uses high-resolution UAV imagery for semantic segmentation to derive precise quadrat-scale canopy cover as ground truth. Subsequently, a deep regression network is developed using UAV orthophotos to estimate canopy cover at the plot scale. The findings indicate that semantic segmentation models leveraging deep learning techniques provide accurate vegetation segmentation and canopy cover estimation at the quadrat level, with UNet++ delivering the highest performance, marked by a mean intersection over union (MIoU) of 0.81 and an F1-score of 0.88. The canopy cover results derived from UNet++ segmentation exhibit a coefficient of determination (R2) of 0.98 and a root mean square error (RMSE) under 3.6 %, surpassing conventional methods like Canopeo and Random Forest (RF). At plot scale, models based on convolutional neural networks (CNNs) and vision transformer (ViT) architecture show enhanced capabilities in predicting canopy cover, with the Swin transformer-based model achieving the greatest accuracy (R2 = 0.90, RMSE = 5.48 %). In meadow, typical, and desert steppe, the Swin tansformer-based model consistently delivers high-precision canopy cover estimates. This study highlights the potential of integrating multi-scale UAV imagery with advanced deep learning techniques for efficient and accurate grassland vegetation monitoring. Future research should focus on optimizing model performance, extending applications to diverse ecosystems, and incorporating additional data sources to enhance robustness and precision.
{"title":"Estimation of grassland canopy cover at quadrat and plot scales using multi-scale UAV imagery","authors":"Yongcai Wang , Huawei Wan , Dongpo Wang , Jixi Gao , Zhuowei Hu , Junjie Wang , Fengming Wan , Bin Yang , Zhiru Zhang , Ma Bian , Jiqian Zhou","doi":"10.1016/j.srs.2025.100312","DOIUrl":"10.1016/j.srs.2025.100312","url":null,"abstract":"<div><div>Grassland canopy cover acts as an essential metric for gauging the vitality and ecological functions of grassland. Unmanned aerial vehicles (UAVs) provide stable and reliable data for estimating grassland canopy cover. However, conventional approaches primarily rely on samples from ground surveys and visual assessments, where data consistency is often affected by variations in survey techniques and personnel expertise. By contrast, UAVs provide consistent multi-scale grassland canopy data. Thus, effectively harnessing the strengths of multi-scale UAV imagery can markedly improve the efficiency and precision of canopy cover estimation. This study uses high-resolution UAV imagery for semantic segmentation to derive precise quadrat-scale canopy cover as ground truth. Subsequently, a deep regression network is developed using UAV orthophotos to estimate canopy cover at the plot scale. The findings indicate that semantic segmentation models leveraging deep learning techniques provide accurate vegetation segmentation and canopy cover estimation at the quadrat level, with UNet++ delivering the highest performance, marked by a mean intersection over union (MIoU) of 0.81 and an F1-score of 0.88. The canopy cover results derived from UNet++ segmentation exhibit a coefficient of determination (<em>R</em><sup>2</sup>) of 0.98 and a root mean square error (RMSE) under 3.6 %, surpassing conventional methods like Canopeo and Random Forest (RF). At plot scale, models based on convolutional neural networks (CNNs) and vision transformer (ViT) architecture show enhanced capabilities in predicting canopy cover, with the Swin transformer-based model achieving the greatest accuracy (<em>R</em><sup>2</sup> = 0.90, RMSE = 5.48 %). In meadow, typical, and desert steppe, the Swin tansformer-based model consistently delivers high-precision canopy cover estimates. This study highlights the potential of integrating multi-scale UAV imagery with advanced deep learning techniques for efficient and accurate grassland vegetation monitoring. Future research should focus on optimizing model performance, extending applications to diverse ecosystems, and incorporating additional data sources to enhance robustness and precision.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100312"},"PeriodicalIF":5.2,"publicationDate":"2025-10-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145528076","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 : 2025-10-29DOI: 10.1016/j.srs.2025.100322
Mari Cullerton, Zhe Zhu, Shi Qiu, Chadwick D. Rittenhouse, Ji Won Suh
Globally, solar photovoltaic installations have been increasing at an exponential rate. This trend towards renewable energy is expected to continue for the foreseeable future to reduce greenhouse gas emissions. However, with the expansion of energy comes a greater land use footprint to accommodate this new infrastructure, i.e., energy sprawl. Given the effect of land conversion on habitat loss and fragmentation, it is imperative to better understand where large-scale solar installations are in our landscape. This study presents a new framework for combining deep learning and time series analysis to map and date large-scale solar installations in a complex temperate landscape. We first employed a U-Net model on a synthetic clear Sentinel-2 image to accurately classify solar facilities, achieving a user's accuracy of 97.33 % and a producer's accuracy of 100.00 %. Leveraging the extensive temporal coverage of Landsat data, we then implemented a reverse application of the Continuous monitoring of Land Disturbance (COLD) algorithm. By running COLD backward in time (2022–1985) over classified solar areas, we exploited the temporal stability of these features to reconstruct installation dates. This method proved to be highly accurate, with a mean difference of 0.125 years between the reference year and COLD-detected installation year. This study provides a high-quality solar dataset for Connecticut as of June 2021 and validates a novel framework with demonstrated potential for application in other regions. These results provide valuable information for land use planning and environmental impact assessment, as well as a powerful proof-of-concept for a methodology that can be used in future solar documentation work.
{"title":"Back in time: A novel time series and deep learning framework for mapping solar installations","authors":"Mari Cullerton, Zhe Zhu, Shi Qiu, Chadwick D. Rittenhouse, Ji Won Suh","doi":"10.1016/j.srs.2025.100322","DOIUrl":"10.1016/j.srs.2025.100322","url":null,"abstract":"<div><div>Globally, solar photovoltaic installations have been increasing at an exponential rate. This trend towards renewable energy is expected to continue for the foreseeable future to reduce greenhouse gas emissions. However, with the expansion of energy comes a greater land use footprint to accommodate this new infrastructure, i.e., energy sprawl. Given the effect of land conversion on habitat loss and fragmentation, it is imperative to better understand where large-scale solar installations are in our landscape. This study presents a new framework for combining deep learning and time series analysis to map and date large-scale solar installations in a complex temperate landscape. We first employed a U-Net model on a synthetic clear Sentinel-2 image to accurately classify solar facilities, achieving a user's accuracy of 97.33 % and a producer's accuracy of 100.00 %. Leveraging the extensive temporal coverage of Landsat data, we then implemented a reverse application of the Continuous monitoring of Land Disturbance (COLD) algorithm. By running COLD backward in time (2022–1985) over classified solar areas, we exploited the temporal stability of these features to reconstruct installation dates. This method proved to be highly accurate, with a mean difference of 0.125 years between the reference year and COLD-detected installation year. This study provides a high-quality solar dataset for Connecticut as of June 2021 and validates a novel framework with demonstrated potential for application in other regions. These results provide valuable information for land use planning and environmental impact assessment, as well as a powerful proof-of-concept for a methodology that can be used in future solar documentation work.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100322"},"PeriodicalIF":5.2,"publicationDate":"2025-10-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145528066","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 : 2025-10-28DOI: 10.1016/j.srs.2025.100321
Peiyuan Wang , Fang Cheng , Junqiang Han , Zhen Jiang , Yang Liu , Rui Tu , Xiaolei Wang , Weisheng Wang , Bayin Dalai , Gulayozov Majid Shonazarovich , Yaoming Li , Xiaochun Lu
Real-time water level monitoring is of critical significance in flood disaster mitigation and water resource management. This paper proposes a real-time Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) water level retrieval method based on the hybrid integration of sliding window and Long Short-Term Memory (LSTM). By dynamically updating input sequences through the sliding window mechanism, an LSTM model captures both temporal and nonlinear characteristics of water level variations, enabling high-precision real-time prediction. Experimental results demonstrate that during non-typhoon seasons, the predicted sea level achieves a correlation coefficient of 99.78 % and a root mean square error (RMSE) of 10.81 cm compared to tide gauge measurements. The system still formulates stable predictions for near-real-time sea level monitoring even with 1.31 % data gaps caused by missing values, which satisfies the requirements. During storm surge, the correlation coefficient between predicted and measured data reaches 96.18 %, with a RMSE of 16.55 cm. Notably, the method maintains robust real-time predictive capability even under extreme conditions where wind speeds exceed 30 m/s and retrieval values significantly decrease. These results demonstrate that the proposed method achieves high accuracy under both normal and extreme hydrological conditions, providing an efficient, cost-effective technical pathway for nearshore real-time water level monitoring and disaster early warning.
{"title":"GNSS-IR real-time water level retrieval method based on hybrid sliding window and LSTM","authors":"Peiyuan Wang , Fang Cheng , Junqiang Han , Zhen Jiang , Yang Liu , Rui Tu , Xiaolei Wang , Weisheng Wang , Bayin Dalai , Gulayozov Majid Shonazarovich , Yaoming Li , Xiaochun Lu","doi":"10.1016/j.srs.2025.100321","DOIUrl":"10.1016/j.srs.2025.100321","url":null,"abstract":"<div><div>Real-time water level monitoring is of critical significance in flood disaster mitigation and water resource management. This paper proposes a real-time Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) water level retrieval method based on the hybrid integration of sliding window and Long Short-Term Memory (LSTM). By dynamically updating input sequences through the sliding window mechanism, an LSTM model captures both temporal and nonlinear characteristics of water level variations, enabling high-precision real-time prediction. Experimental results demonstrate that during non-typhoon seasons, the predicted sea level achieves a correlation coefficient of 99.78 % and a root mean square error (RMSE) of 10.81 cm compared to tide gauge measurements. The system still formulates stable predictions for near-real-time sea level monitoring even with 1.31 % data gaps caused by missing values, which satisfies the requirements. During storm surge, the correlation coefficient between predicted and measured data reaches 96.18 %, with a RMSE of 16.55 cm. Notably, the method maintains robust real-time predictive capability even under extreme conditions where wind speeds exceed 30 m/s and retrieval values significantly decrease. These results demonstrate that the proposed method achieves high accuracy under both normal and extreme hydrological conditions, providing an efficient, cost-effective technical pathway for nearshore real-time water level monitoring and disaster early warning.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100321"},"PeriodicalIF":5.2,"publicationDate":"2025-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145416016","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}
<div><div>Recent droughts from 2017 to 2020 induced significant stress on woodland canopies across eastern Australia, resulting in widespread tree browning and mortality. However, the trajectory of post-drought recovery remains unclear, with uncertainty about whether canopy conditions are improving or continuing to decline. Identifying the key local environmental and climatic factors influencing drought-induced tree mortality and recovery is therefore critical for understanding these processes. In this study, we employed a data-driven deep learning framework that integrates CNNs and LSTM algorithms techniques that excel at capturing spatial dependencies and long-term temporal dynamics, respectively. We analyzed a seasonal time-series dataset spanning 2010–2022, which combined satellite-derived canopy stress anomalies (z-scores of the Normalized Burn Ratio, NBR) with environmental predictors including rainfall, temperature, and potential evapotranspiration (PET), soil texture (sand and clay fractions), vegetation type, and topographic variables (slope, aspect, and topographic wetness index, TWI). All predictors were at 30 m resolution to ensure spatial consistency. Among the compared models, the hybrid CNN-LSTM model performed the best, underscoring the superiority of hybrid architectures like that synergistically capture spatial patterns and temporal dependencies. Additionally, advanced sequential models, whether utilizing attention mechanisms, such as Selective Attention and TFT, or leveraging state-space formulations, such as TCN-Mamba and RWKV-TS, also outperformed traditional recurrent approaches. Beyond predictive performance, our aim was to interpret the ecological drivers of canopy stress and recovery by linking model sensitivities to physiological processes and landscape variability. Our phase-based analysis (wet, transition, drought, and post-drought conditions) revealed that in dry to mid-humid bioregions, canopy resilience during drought is shaped by the interplay of dynamic climatic stressors (e.g., evapotranspiration, rainfall variability) and static landscape features (soil texture, topography), highlighting how ecosystem vulnerability arises from synergistic abiotic thresholds under climatic extremes. Regions that followed a recovery pathway were primarily driven by rainfall and topographic variables, whereas non-recovery was largely associated with dominant climatic factors. In wet-humid bioregions, climatic variables were the primary triggers of drought impacts; recovery in these areas was influenced by an interaction between climatic and topographic variations, with topographic factors such as sand content and Topographic Wetness Index (TWI) playing a decisive role in non-recovered regions. Insights obtained through explainable algorithms can inform process-based models and facilitate a more robust mechanistic understanding of drought-impacted eucalypt woodlands. Ultimately, this remote sensing based approach holds promise for o
{"title":"Understanding drought related tree responses using deep learning approaches and satellite based proxy","authors":"Rachid Oucheikh , Nuwanthi Arampola , Pengxiang Zhao , Ali Mansourian","doi":"10.1016/j.srs.2025.100317","DOIUrl":"10.1016/j.srs.2025.100317","url":null,"abstract":"<div><div>Recent droughts from 2017 to 2020 induced significant stress on woodland canopies across eastern Australia, resulting in widespread tree browning and mortality. However, the trajectory of post-drought recovery remains unclear, with uncertainty about whether canopy conditions are improving or continuing to decline. Identifying the key local environmental and climatic factors influencing drought-induced tree mortality and recovery is therefore critical for understanding these processes. In this study, we employed a data-driven deep learning framework that integrates CNNs and LSTM algorithms techniques that excel at capturing spatial dependencies and long-term temporal dynamics, respectively. We analyzed a seasonal time-series dataset spanning 2010–2022, which combined satellite-derived canopy stress anomalies (z-scores of the Normalized Burn Ratio, NBR) with environmental predictors including rainfall, temperature, and potential evapotranspiration (PET), soil texture (sand and clay fractions), vegetation type, and topographic variables (slope, aspect, and topographic wetness index, TWI). All predictors were at 30 m resolution to ensure spatial consistency. Among the compared models, the hybrid CNN-LSTM model performed the best, underscoring the superiority of hybrid architectures like that synergistically capture spatial patterns and temporal dependencies. Additionally, advanced sequential models, whether utilizing attention mechanisms, such as Selective Attention and TFT, or leveraging state-space formulations, such as TCN-Mamba and RWKV-TS, also outperformed traditional recurrent approaches. Beyond predictive performance, our aim was to interpret the ecological drivers of canopy stress and recovery by linking model sensitivities to physiological processes and landscape variability. Our phase-based analysis (wet, transition, drought, and post-drought conditions) revealed that in dry to mid-humid bioregions, canopy resilience during drought is shaped by the interplay of dynamic climatic stressors (e.g., evapotranspiration, rainfall variability) and static landscape features (soil texture, topography), highlighting how ecosystem vulnerability arises from synergistic abiotic thresholds under climatic extremes. Regions that followed a recovery pathway were primarily driven by rainfall and topographic variables, whereas non-recovery was largely associated with dominant climatic factors. In wet-humid bioregions, climatic variables were the primary triggers of drought impacts; recovery in these areas was influenced by an interaction between climatic and topographic variations, with topographic factors such as sand content and Topographic Wetness Index (TWI) playing a decisive role in non-recovered regions. Insights obtained through explainable algorithms can inform process-based models and facilitate a more robust mechanistic understanding of drought-impacted eucalypt woodlands. Ultimately, this remote sensing based approach holds promise for o","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100317"},"PeriodicalIF":5.2,"publicationDate":"2025-10-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145415892","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 : 2025-10-23DOI: 10.1016/j.srs.2025.100318
Saverio Francini , Costanza Borghi , Giovanni D'Amico , Lars T. Waser , Maciej Lisiewicz , Krzysztof Stereńczak , Mart-Jan Schelhaas , Cameron Pellett , Terje Gobakken , Erik Næsset , Federico Magnani , Sergio de-Miguel , Gert-Jan Nabuurs , Ruben Valbuena , Gherardo Chirici
European forests contribute to climate change mitigation by sequestering carbon, conserving biodiversity, and enhancing water retention. However, climate-induced disturbances such as fires, windthrows, droughts, and pest outbreaks underscore the need for stronger forest monitoring systems. National Forest Inventories (NFIs) serve as the primary source of forest data and information in Europe. Yet, inconsistencies in timing, coverage, methodologies, and data quality highlight the need for a more harmonized and spatially detailed approach. Critically, predicting forest variables directly from satellite data remains challenging, mainly due to the difficulties in aligning remote sensing with ground data. Meanwhile, the operational use of airborne laser scanning (ALS) data is limited by high costs, infrequent updates, and inconsistent coverage from different sensors and flight conditions. This study presents a novel approach relying on fully connected neural networks to integrate Landsat satellite time series and forest disturbance and recovery metrics with ALS data to predict forest height metrics, which can then be used to accurately predict critical forest variables, such as growing stock volume (GSV) and stand basal area (BA). The method was tested across five ecologically and geographically diverse European forest regions: Tuscany (Italy), the Netherlands, the Canton of Grisons (Switzerland), Białowieża Forest (Poland), and the Vindelälven-Juhttátahkka Biosphere Reserve (Sweden). ALS forest height metrics were predicted with R2 values ranging from 0.47 to 0.68. Then, based on field data, forest height metrics were used to predict GSV (R2 = 0.78) and BA (R2 = 0.69). Our method addresses the issue of limited spatial and temporal availability of ALS data by predicting ALS-derived height metrics using Landsat time series. This study examines the challenges of combining satellite and NFI data, building on the premise that satellite data can be effectively used to predict forest height metrics derived from ALS, which in turn can be used to accurately quantify several forest variables. The methods presented here support scalable and cost-effective forest monitoring by providing the spatially and temporally detailed information needed to implement climate-smart forestry.
{"title":"Bridging spatio-temporal gaps in ALS data using Landsat time series and forest disturbance-recovery metrics via multi-task neural networks","authors":"Saverio Francini , Costanza Borghi , Giovanni D'Amico , Lars T. Waser , Maciej Lisiewicz , Krzysztof Stereńczak , Mart-Jan Schelhaas , Cameron Pellett , Terje Gobakken , Erik Næsset , Federico Magnani , Sergio de-Miguel , Gert-Jan Nabuurs , Ruben Valbuena , Gherardo Chirici","doi":"10.1016/j.srs.2025.100318","DOIUrl":"10.1016/j.srs.2025.100318","url":null,"abstract":"<div><div>European forests contribute to climate change mitigation by sequestering carbon, conserving biodiversity, and enhancing water retention. However, climate-induced disturbances such as fires, windthrows, droughts, and pest outbreaks underscore the need for stronger forest monitoring systems. National Forest Inventories (NFIs) serve as the primary source of forest data and information in Europe. Yet, inconsistencies in timing, coverage, methodologies, and data quality highlight the need for a more harmonized and spatially detailed approach. Critically, predicting forest variables directly from satellite data remains challenging, mainly due to the difficulties in aligning remote sensing with ground data. Meanwhile, the operational use of airborne laser scanning (ALS) data is limited by high costs, infrequent updates, and inconsistent coverage from different sensors and flight conditions. This study presents a novel approach relying on fully connected neural networks to integrate Landsat satellite time series and forest disturbance and recovery metrics with ALS data to predict forest height metrics, which can then be used to accurately predict critical forest variables, such as growing stock volume (GSV) and stand basal area (BA). The method was tested across five ecologically and geographically diverse European forest regions: Tuscany (Italy), the Netherlands, the Canton of Grisons (Switzerland), Białowieża Forest (Poland), and the Vindelälven-Juhttátahkka Biosphere Reserve (Sweden). ALS forest height metrics were predicted with R<sup>2</sup> values ranging from 0.47 to 0.68. Then, based on field data, forest height metrics were used to predict GSV (R<sup>2</sup> = 0.78) and BA (R<sup>2</sup> = 0.69). Our method addresses the issue of limited spatial and temporal availability of ALS data by predicting ALS-derived height metrics using Landsat time series. This study examines the challenges of combining satellite and NFI data, building on the premise that satellite data can be effectively used to predict forest height metrics derived from ALS, which in turn can be used to accurately quantify several forest variables. The methods presented here support scalable and cost-effective forest monitoring by providing the spatially and temporally detailed information needed to implement climate-smart forestry.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100318"},"PeriodicalIF":5.2,"publicationDate":"2025-10-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145415890","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 : 2025-10-21DOI: 10.1016/j.srs.2025.100305
Mohammad Esmaeili , Dariush Abbasi-Moghadam , Alireza Sharifi , Nizom Farmonov
Mapping wildfire burned areas using satellite imagery is essential for immediate response measures as well as for long-term recovery planning. These maps provide critical information to response teams, allowing them to effectively allocate resources and prioritize affected areas. This study focuses on the aim of providing accurate maps of areas affected by wildfires in the Guzli region near Bukhara province in Uzbekistan. The benchmark dataset from the study area, named UZB-WF2022, which indicates the country name and year of occurrence, and includes Sentinel-2 and Plant-Scope multispectral multi-resolution images. This study uses a mixture of unsupervised deep learning with the k-means algorithm to accurately identify and map burned areas. The core of the proposed method is an autoencoder model designed with 3-dimensional convolutional layers. This autoencoder is mixed with the k-means algorithm in the latent space of the model and uses the k-means cost function to improve the training process. In addition, the proposed method has an attention mechanism based on morphological operations called Inject-Multiply. This mechanism integrates morphological features obtained from post-wildfire vegetation index data and moisture changes captured in Sentinel-2 images, focusing on enriching features related to the shape and boundaries of burned areas. This study evaluates the effectiveness of the proposed method in labeling, identifying, and mapping burned areas using benchmark datasets, using different evaluation criteria. The model achieves an accuracy of over 93 % on the UZB-WF2022 dataset. This approach increases the accuracy of burned area detection in similar datasets and facilitates more informed decision-making for post-wildfire recovery and land management.
{"title":"Hybrid unsupervised methods and inject-multiply morphological features for mapping wildfire burned areas with multi-spectral satellite data","authors":"Mohammad Esmaeili , Dariush Abbasi-Moghadam , Alireza Sharifi , Nizom Farmonov","doi":"10.1016/j.srs.2025.100305","DOIUrl":"10.1016/j.srs.2025.100305","url":null,"abstract":"<div><div>Mapping wildfire burned areas using satellite imagery is essential for immediate response measures as well as for long-term recovery planning. These maps provide critical information to response teams, allowing them to effectively allocate resources and prioritize affected areas. This study focuses on the aim of providing accurate maps of areas affected by wildfires in the Guzli region near Bukhara province in Uzbekistan. The benchmark dataset from the study area, named UZB-WF2022, which indicates the country name and year of occurrence, and includes Sentinel-2 and Plant-Scope multispectral multi-resolution images. This study uses a mixture of unsupervised deep learning with the k-means algorithm to accurately identify and map burned areas. The core of the proposed method is an autoencoder model designed with 3-dimensional convolutional layers. This autoencoder is mixed with the k-means algorithm in the latent space of the model and uses the k-means cost function to improve the training process. In addition, the proposed method has an attention mechanism based on morphological operations called Inject-Multiply. This mechanism integrates morphological features obtained from post-wildfire vegetation index data and moisture changes captured in Sentinel-2 images, focusing on enriching features related to the shape and boundaries of burned areas. This study evaluates the effectiveness of the proposed method in labeling, identifying, and mapping burned areas using benchmark datasets, using different evaluation criteria. The model achieves an accuracy of over 93 % on the UZB-WF2022 dataset. This approach increases the accuracy of burned area detection in similar datasets and facilitates more informed decision-making for post-wildfire recovery and land management.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100305"},"PeriodicalIF":5.2,"publicationDate":"2025-10-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145361666","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 : 2025-10-17DOI: 10.1016/j.srs.2025.100316
Justin Holvoet , Nicolas Latte , Jérôme Perin , Jean-François Bastin , Hugo de Lame , Daniel Kükenbrink , Philippe Lejeune
In the context of a growing need to diversify forest information, national and regional forest inventories (NFI and RFI) could benefit from mobile Light Detection and Ranging (LiDAR) technologies. Ground-based mobile laser scanning (MLS) and unmanned aerial laser scanning (ULS) can potentially retrieve a large panel of forest attributes quickly, efficiently, and accurately. In this study, conducted in Wallonia (southern Belgium), we aimed to evaluate, in the context of an NFI, the accuracy of MLS at tree, plot, and inventory levels and the potential benefits of fusing ULS with MLS. In total, 60 circular forest plots of 0.1 ha containing 2497 trees were measured by traditional inventory means and scanned using MLS. Among them, 27 were additionally scanned by ULS, and ULS and MLS scans were fused to produce an enhanced point cloud. We then evaluated the accuracy of MLS considering, at tree level, the diameter at breast height, total height, merchantable wood volume, and crown projected area and volume; at plot level, the total merchantable wood volume, number of trees, and total basal area; and for the whole inventory, the total volume and number of trees. Tree, plot, and inventory metrics were accurately acquired with a strong correlation to field measurements (r2 ranging from 0.83 to 0.98). Out of all estimated metrics, height has a potential accurately estimated by MLS than by field measurements. The fusion of ULS and MLS allowed for a more accurate crown measurement, but height estimation was not significantly better than with MLS scan alone. The accuracy of soft- and hardwood forest plot estimations differed considering total plot wood volume, number of trees, and individual tree height. In this study, we explored the possibility and limitations of MLS in undertaking large-scale inventory in terms of accuracy, time, and reliability.
{"title":"Mobile laser scanning in support of national and regional forest inventories","authors":"Justin Holvoet , Nicolas Latte , Jérôme Perin , Jean-François Bastin , Hugo de Lame , Daniel Kükenbrink , Philippe Lejeune","doi":"10.1016/j.srs.2025.100316","DOIUrl":"10.1016/j.srs.2025.100316","url":null,"abstract":"<div><div>In the context of a growing need to diversify forest information, national and regional forest inventories (NFI and RFI) could benefit from mobile Light Detection and Ranging (LiDAR) technologies. Ground-based mobile laser scanning (MLS) and unmanned aerial laser scanning (ULS) can potentially retrieve a large panel of forest attributes quickly, efficiently, and accurately. In this study, conducted in Wallonia (southern Belgium), we aimed to evaluate, in the context of an NFI, the accuracy of MLS at tree, plot, and inventory levels and the potential benefits of fusing ULS with MLS. In total, 60 circular forest plots of 0.1 ha containing 2497 trees were measured by traditional inventory means and scanned using MLS. Among them, 27 were additionally scanned by ULS, and ULS and MLS scans were fused to produce an enhanced point cloud. We then evaluated the accuracy of MLS considering, at tree level, the diameter at breast height, total height, merchantable wood volume, and crown projected area and volume; at plot level, the total merchantable wood volume, number of trees, and total basal area; and for the whole inventory, the total volume and number of trees. Tree, plot, and inventory metrics were accurately acquired with a strong correlation to field measurements (<em>r</em><sup>2</sup> ranging from 0.83 to 0.98). Out of all estimated metrics, height has a potential accurately estimated by MLS than by field measurements. The fusion of ULS and MLS allowed for a more accurate crown measurement, but height estimation was not significantly better than with MLS scan alone. The accuracy of soft- and hardwood forest plot estimations differed considering total plot wood volume, number of trees, and individual tree height. In this study, we explored the possibility and limitations of MLS in undertaking large-scale inventory in terms of accuracy, time, and reliability.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100316"},"PeriodicalIF":5.2,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145361667","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 : 2025-10-17DOI: 10.1016/j.srs.2025.100309
Chandrika Pinnepalli , Roujean Jean-Louis , Eswar Rajasekaran , Thomas Vidal , Zunjian Bian , Tian Hu , Mark Irvine , Biao Cao , Philippe Gamet
This study evaluates the influence of the hot spot effect, i.e. when the solar and viewing angles coincide, producing a radiance peak on the diurnal reflectance and temperature cycles (DRC and DTC, respectively) observed by the SEVIRI (Spinning Enhanced Visible and InfraRed Imager) sensor aboard the Meteosat Second Generation (MSG) satellite. Focusing on clear-sky conditions and multiple land cover types, we assess the directional impact on both spectral brightness temperature (Tb) and land surface temperature (LST). A four-parameter DTC model is coupled with a directional kernel-driven model (KDM), including a hot spot term, to create Time-Evolving KDMs. The models are applied to six diverse sites to evaluate whether optical BRDF characteristics can inform thermal BRDF (Bidirectional Reflectance Distribution Function) behavior, and to what extent directional effects distort DTC profiles. Findings indicate a clear hot spot signature in the DRC, while in the DTC, it subtly alters the bell-shaped curve, resulting in Tb deviations up to 3 K and LST differences up to 4 °C. The results underscore the need to correct for angular effects when comparing DTCs across sites or seasons. Moreover, visual inspections show that optical BRDF peaks align closely with cosine peaks for two satellites, whereas thermal peaks diverge—highlighting mismatches and the challenges of modeling mixed land cover. Present findings underscore the need for improved models and multi-sensor validation to support a full exploitation of thermal remote sensing.
{"title":"Joint analysis and modeling of the hot spot effect from the diurnal reflectance and temperature cycles observed by SEVIRI","authors":"Chandrika Pinnepalli , Roujean Jean-Louis , Eswar Rajasekaran , Thomas Vidal , Zunjian Bian , Tian Hu , Mark Irvine , Biao Cao , Philippe Gamet","doi":"10.1016/j.srs.2025.100309","DOIUrl":"10.1016/j.srs.2025.100309","url":null,"abstract":"<div><div>This study evaluates the influence of the hot spot effect, i.e. when the solar and viewing angles coincide, producing a radiance peak on the diurnal reflectance and temperature cycles (DRC and DTC, respectively) observed by the SEVIRI (<span><span>Spinning Enhanced Visible and InfraRed Imager</span><svg><path></path></svg></span>) sensor aboard the Meteosat Second Generation (MSG) satellite. Focusing on clear-sky conditions and multiple land cover types, we assess the directional impact on both spectral brightness temperature (<em>Tb</em>) and land surface temperature (LST). A four-parameter DTC model is coupled with a directional kernel-driven model (KDM), including a hot spot term, to create Time-Evolving KDMs. The models are applied to six diverse sites to evaluate whether optical BRDF characteristics can inform thermal BRDF (Bidirectional Reflectance Distribution Function) behavior, and to what extent directional effects distort DTC profiles. Findings indicate a clear hot spot signature in the DRC, while in the DTC, it subtly alters the bell-shaped curve, resulting in <em>Tb</em> deviations up to 3 K and LST differences up to 4 °C. The results underscore the need to correct for angular effects when comparing DTCs across sites or seasons. Moreover, visual inspections show that optical BRDF peaks align closely with cosine peaks for two satellites, whereas thermal peaks diverge—highlighting mismatches and the challenges of modeling mixed land cover. Present findings underscore the need for improved models and multi-sensor validation to support a full exploitation of thermal remote sensing.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100309"},"PeriodicalIF":5.2,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145415925","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 : 2025-10-17DOI: 10.1016/j.srs.2025.100306
Venkanna Babu Guthula , Stefan Oehmcke , Remigio Chilaule , Hui Zhang , Nico Lang , Ankit Kariryaa , Johan Mottelson , Christian Igel
Since low-quality housing and in particular certain roof characteristics are associated with an increased risk of malaria, classification of roof types based on remote sensing imagery can aid in assessing malaria risk and thereby help prevent the disease. To support research in this area, we release the Nacala-Roof-Material dataset,1 which contains high-resolution drone images from Mozambique with corresponding labels delineating houses and specifying their roof types. The dataset defines a multi-task computer vision problem, encompassing object detection, classification, and segmentation. Additionally, we benchmarked various state-of-the-art approaches on the dataset. Canonical U-Nets, YOLOv8, and a custom decoder on pretrained DINOv2 served as baselines. We demonstrate that while each method has its advantages, none is superior across all metrics, highlighting the potential of our dataset for future research in multi-task learning. Although the tasks are closely related, accurate segmentation of objects does not necessarily imply accurate instance separation, and vice versa. We address this general issue by introducing a variant of the deep ordinal watershed (DOW) approach, which additionally segments the interior of objects, allowing for improved object delineation and separation. We show that our DOW variant is a versatile approach that improves the performance of both U-Net and DINOv2 backbones, mitigating potential conflicts between semantic segmentation and instance segmentation.
{"title":"Drone imagery for roof detection, classification, and segmentation to support mosquito-borne disease risk assessment: The Nacala-Roof-Material dataset","authors":"Venkanna Babu Guthula , Stefan Oehmcke , Remigio Chilaule , Hui Zhang , Nico Lang , Ankit Kariryaa , Johan Mottelson , Christian Igel","doi":"10.1016/j.srs.2025.100306","DOIUrl":"10.1016/j.srs.2025.100306","url":null,"abstract":"<div><div>Since low-quality housing and in particular certain roof characteristics are associated with an increased risk of malaria, classification of roof types based on remote sensing imagery can aid in assessing malaria risk and thereby help prevent the disease. To support research in this area, we release the Nacala-Roof-Material dataset,<span><span><sup>1</sup></span></span> which contains high-resolution drone images from Mozambique with corresponding labels delineating houses and specifying their roof types. The dataset defines a multi-task computer vision problem, encompassing object detection, classification, and segmentation. Additionally, we benchmarked various state-of-the-art approaches on the dataset. Canonical U-Nets, YOLOv8, and a custom decoder on pretrained DINOv2 served as baselines. We demonstrate that while each method has its advantages, none is superior across all metrics, highlighting the potential of our dataset for future research in multi-task learning. Although the tasks are closely related, accurate segmentation of objects does not necessarily imply accurate instance separation, and vice versa. We address this general issue by introducing a variant of the deep ordinal watershed (DOW) approach, which additionally segments the interior of objects, allowing for improved object delineation and separation. We show that our DOW variant is a versatile approach that improves the performance of both U-Net and DINOv2 backbones, mitigating potential conflicts between semantic segmentation and instance segmentation.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100306"},"PeriodicalIF":5.2,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145361668","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 : 2025-10-16DOI: 10.1016/j.srs.2025.100315
Babak Ghassemi , Cassio F. Dantas , Raffaele Gaetano , Dino Ienco , Omid Ghorbanzadeh , Emma Izquierdo-Verdiguier , Francesco Vuolo
Land use and land cover mapping from Earth Observation (EO) data is a critical tool for sustainable land and resource management as, for instance, in domains like biodiversity and agricultural food production. While advanced machine learning and deep learning algorithms excel at analyzing EO imagery data, they often overlook crucial geospatial metadata information that could enhance scalability and accuracy across regional, continental, and global scales. To address this limitation, we propose BRIDGE-LC (Bi-level Representation Integration for Disentangled GEospatial Land Cover), a novel deep learning framework that explicitly integrates multi-scale geospatial information into the land cover classification process. By simultaneously leveraging fine-grained (latitude/longitude) and coarse-grained (biogeographical region) spatial information, our lightweight multi-layer perceptron architecture learns from both multi-scale information during training but only requires fine-grained information for inference, allowing it to disentangle region-specific from region-agnostic land cover features while maintaining computational efficiency comparable with standard machine learning approaches. To assess the quality of our framework, we use an open-access in-situ dataset spanning the 27 countries of the European Union and we adopt several competing classification approaches commonly considered for large-scale land cover mapping. We evaluated all the approaches through two scenarios: an extrapolation scenario in which training data encompasses samples coming from all the biogeographical regions and a leave-one-region-out scenario where samples from all the regions, except one, are employed for the training stage. Additionally, we also explore the spatial representation learned by the proposed deep learning model, highlighting a connection between its internal manifold and the geographical information used during the training stage. Our results demonstrate that integrating geospatial information improves land cover mapping performances, with the most substantial gains achieved by jointly leveraging both fine-grained and coarse-grained spatial information.
利用地球观测数据绘制土地利用和土地覆盖地图是可持续土地和资源管理的重要工具，例如在生物多样性和农业粮食生产等领域。虽然先进的机器学习和深度学习算法擅长分析EO图像数据，但它们往往忽略了关键的地理空间元数据信息，而这些信息可以提高区域、大陆和全球范围内的可扩展性和准确性。为了解决这一限制，我们提出了BRIDGE-LC （Disentangled GEospatial Land Cover的双层表示集成），这是一个新的深度学习框架，明确地将多尺度地理空间信息集成到土地覆盖分类过程中。通过同时利用细粒度（纬度/经度）和粗粒度（生物地理区域）空间信息，我们的轻量级多层感知器架构在训练期间从多尺度信息中学习，但只需要细粒度信息进行推理，从而使其能够将特定区域与不可知区域的土地覆盖特征区分开来，同时保持与标准机器学习方法相当的计算效率。为了评估我们的框架的质量，我们使用了一个覆盖欧盟27个国家的开放获取的现场数据集，我们采用了几种通常用于大规模土地覆盖制图的竞争性分类方法。我们通过两种场景对所有方法进行了评估：一种是外推场景，其中训练数据包含来自所有生物地理区域的样本；另一种是留一个区域的场景，其中除了一个区域外，所有区域的样本都被用于训练阶段。此外，我们还探讨了所提出的深度学习模型所学习的空间表示，强调了其内部流形与训练阶段使用的地理信息之间的联系。我们的研究结果表明，整合地理空间信息提高了土地覆盖制图的性能，其中最显著的收益是通过共同利用细粒度和粗粒度空间信息实现的。
{"title":"Geographical context matters: Bridging fine and coarse spatial information to enhance continental land cover mapping","authors":"Babak Ghassemi , Cassio F. Dantas , Raffaele Gaetano , Dino Ienco , Omid Ghorbanzadeh , Emma Izquierdo-Verdiguier , Francesco Vuolo","doi":"10.1016/j.srs.2025.100315","DOIUrl":"10.1016/j.srs.2025.100315","url":null,"abstract":"<div><div>Land use and land cover mapping from Earth Observation (EO) data is a critical tool for sustainable land and resource management as, for instance, in domains like biodiversity and agricultural food production. While advanced machine learning and deep learning algorithms excel at analyzing EO imagery data, they often overlook crucial geospatial metadata information that could enhance scalability and accuracy across regional, continental, and global scales. To address this limitation, we propose BRIDGE-LC (Bi-level Representation Integration for Disentangled GEospatial Land Cover), a novel deep learning framework that explicitly integrates multi-scale geospatial information into the land cover classification process. By simultaneously leveraging fine-grained (latitude/longitude) and coarse-grained (biogeographical region) spatial information, our lightweight multi-layer perceptron architecture learns from both multi-scale information during training but only requires fine-grained information for inference, allowing it to disentangle region-specific from region-agnostic land cover features while maintaining computational efficiency comparable with standard machine learning approaches. To assess the quality of our framework, we use an open-access in-situ dataset spanning the 27 countries of the European Union and we adopt several competing classification approaches commonly considered for large-scale land cover mapping. We evaluated all the approaches through two scenarios: an extrapolation scenario in which training data encompasses samples coming from all the biogeographical regions and a leave-one-region-out scenario where samples from all the regions, except one, are employed for the training stage. Additionally, we also explore the spatial representation learned by the proposed deep learning model, highlighting a connection between its internal manifold and the geographical information used during the training stage. Our results demonstrate that integrating geospatial information improves land cover mapping performances, with the most substantial gains achieved by jointly leveraging both fine-grained and coarse-grained spatial information.</div></div>","PeriodicalId":101147,"journal":{"name":"Science of Remote Sensing","volume":"12 ","pages":"Article 100315"},"PeriodicalIF":5.2,"publicationDate":"2025-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145320117","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