ISPRS Journal of Photogrammetry and Remote Sensing最新文献

{"title":"Two-stage offline knowledge distillation for onboard registration of multispectral satellite images","authors":"Darshana Priyasad ,&nbsp;Tharindu Fernando ,&nbsp;Maryam Haghighat ,&nbsp;Harshala Gammulle ,&nbsp;Roberto Del Prete ,&nbsp;Clinton Fookes","doi":"10.1016/j.isprsjprs.2025.12.009","DOIUrl":"10.1016/j.isprsjprs.2025.12.009","url":null,"abstract":"<div><div>Multi-band optical sensors onboard modern Earth observation satellites capture complementary spectral responses across varying spatial and spectral resolutions. To effectively fuse this information for downstream applications, accurate band co-registration is critical. However, for real-time processing, such registration must be performed onboard, where sensor distortions, platform-induced motion, and spectral disparities introduce significant challenges. Traditional feature matching algorithms struggle to cope with these variations or are often too computationally intensive for the constrained hardware typically found on small satellites. As a result, real-time onboard multimodal fusion has remained largely impractical in operational settings. With the emergence of next-generation satellites equipped with AI-enabled onboard processing, such as Australia’s Kanyini mission, there is now an opportunity to overcome these limitations. In this work, we introduce a deep learning-based, lightweight band registration framework specifically designed for real-time onboard deployment. Our approach features a band-independent teacher network that jointly leverages adversarial learning and supervised regression to estimate affine registration parameters across spectral bands. To meet hardware constraints, we employ a two-stage knowledge distillation strategy that produces a compact yet accurate student model. Experimental results demonstrate that our method delivers robust and efficient registration performance, enabling real-time spectral alignment and significantly enhancing the potential for onboard multimodal data fusion in Earth observation missions.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 319-335"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145840693","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Perspectives on spatio-temporal intelligence","authors":"Deren Li ,&nbsp;Mi Wang ,&nbsp;Jing Xiao ,&nbsp;Bo Yang","doi":"10.1016/j.isprsjprs.2025.12.011","DOIUrl":"10.1016/j.isprsjprs.2025.12.011","url":null,"abstract":"<div><div>The physical world, governed by the perpetual dynamics of matter, is inherently structured by spatio-temporal information. In the current Intelligence Era, rapid advances in data collection and Artificial Intelligence (AI) technologies have enabled large-scale acquisition and analysis of dynamic spatio-temporal information. These developments have led to the emergence of Spatio-Temporal Intelligence (STI), an interdisciplinary field that integrates spatio-temporal data with AI-driven computational methodologies to model, interpret, and manage complex physical, environmental, and social processes.</div><div>In this paper, we offer a perspective on the mission and evolving scope of STI and identify its critical challenges. A general STI framework comprising five interconnected components is proposed to support adaptive observation, multi-modal modeling, causal reasoning, and knowledge-driven service delivery. Through a case study in national park ecological monitoring, we demonstrate how STI enables large-scale, precise, and real-time environmental understanding. Distinct from approaches that simulate symbolic or linguistic cognition, STI is grounded in the physical world and leverages high-dimensional sensor data to enable machine perception, foster new cognitive paradigms, and enhance decision-making across domains.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 305-318"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145840694","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Complex convolutional sparse coding InSAR phase filtering Incorporating directional gradients and second-order difference regularization","authors":"Pengcheng Hu ,&nbsp;Xu Li ,&nbsp;Junhuan Peng ,&nbsp;Xu Ma ,&nbsp;Yuhan Su ,&nbsp;Xiaoman Qi ,&nbsp;Xinwei Jiang ,&nbsp;Wenwen Wang","doi":"10.1016/j.isprsjprs.2025.12.016","DOIUrl":"10.1016/j.isprsjprs.2025.12.016","url":null,"abstract":"<div><div>Interferometric Synthetic Aperture Radar (InSAR) is a technology that can effectively obtain ground information, conduct large-scale topography mapping, and monitor surface deformation. However, InSAR data is interfered by speckle noise caused by radar echo signal fading, ground background clutter, and decoherence, which affects the InSAR interferometric phase quality and thus reduces the accuracy of InSAR results. The existing Complex Convolutional Sparse Coding Gradient Regularization (ComCSC-GR) method incorporates gradient regularization by considering the sparse coefficient matrix’s gradients in both row (azimuth) and column (range) directions. It is an advanced and effective interferogram phase filtering method that can improve the interferogram quality. However, this method does not take into account the variation characteristics of the diagonal gradient and the second-order difference information (caused by edge mutations). As a result, the interferogram still exhibits problems such as staircase artifacts in high-noise and low-coherence areas, uneven interferograms (caused by a large number of residual points), and unclear phase edge structure. This article introduces multiple directional gradients and second-order differential Laplacian operator information, and construct two models: “Complex Convolutional Sparse Coding Model with <em>L</em>₂-norm Regularization of Directional Gradients and Laplacian Operator (ComCSC-RCDL) ” and “Complex Convolutional Sparse Coding Model Coupled with <em>L</em>₁-norm Total Variation Regularization (ComCSC-RCDL-TV)”. These methods enhance the fidelity of phase texture and edge structure, and improve the quality of InSAR interferogram filtering phase in low-coherence scenarios. Comparative experiments were conducted using simulated data, real data from Sentinel-1 and LuTan-1 (LT-1), and advanced methods including ComCSC-GR and InSAR-BM3D (real data experiments included comparison experiments before and after removing the interferogram orbit error). The results show that the proposed model method performs better than the comparative model, verifying the effectiveness of the proposed model.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 740-765"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145957296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Towards High spatial resolution and fine-grained fidelity depth reconstruction of single-photon LiDAR with context-aware spatiotemporal modeling","authors":"Zhenyu Zhang ,&nbsp;Yuan Li ,&nbsp;Feihu Zhu ,&nbsp;Yuechao Ma ,&nbsp;Junying Lv ,&nbsp;Qian Sun ,&nbsp;Lin Li ,&nbsp;Wuming Zhang","doi":"10.1016/j.isprsjprs.2025.11.034","DOIUrl":"10.1016/j.isprsjprs.2025.11.034","url":null,"abstract":"<div><div>While single-photon LiDAR promises revolutionary depth sensing capabilities, existing deep learning frameworks fundamentally fail to overcome the challenge of high spatial resolution (SR) data processing. To address the amplification of fine geometric details and complex spatiotemporal dependencies in high-SR single-photon data, we adopt a U-Net++ backbone with dense skip connections to preserve high-frequency features. Our encoder cascades two novel modules, integrating attention-driven modulation and convolution to adaptively model intricate patterns without sacrificing detail. We propose a 3D triple local-attention fusion module (3D-TriLAF) to suppress incoherent responses across temporal, spatial, and channel axes. In parallel, an opposite continuous dilation spatial–temporal convolution module (OCDSConv) is designed to extract structured context while preserving transient cues. To alleviate the misalignment and semantic drift between low and high-level features—problems exacerbated by increased resolution—we design a multi-scale fusion mechanism that facilitates consistent geometric modeling across scales. Finally, we propose a hybrid loss combining ordinal regression (OR) loss, structural similarity index measure (SSIM) loss, and bilateral total variation (BTV) loss to jointly enhances peak localization, structural fidelity, and edge-aware smoothness. Extensive experiments on two 128 × 128 SR simulated datasets show that, compared with the best baseline, our framework reduces RMSE and Abs Rel by up to 60.00 % and 31.58 %. On two (200 + )×(200 + ) SR real-world datasets, RMSE and Abs Rel drop by 42.31 % and 39.44 %. These quantitative gains and visual improvements in geometric continuity under complex lighting confirm its suitability for fine-grained high-SR single-photon depth reconstruction.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 238-261"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145785309","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Urban–rural gradients in cooling efficiency trends of tree covers across global cities","authors":"Chunli Wang ,&nbsp;Kaiyuan Li ,&nbsp;Wenfeng Zhan ,&nbsp;Long Li ,&nbsp;Chenguang Wang ,&nbsp;Shasha Wang ,&nbsp;Sida Jiang ,&nbsp;Shuang Ge ,&nbsp;Zihan Liu","doi":"10.1016/j.isprsjprs.2025.11.035","DOIUrl":"10.1016/j.isprsjprs.2025.11.035","url":null,"abstract":"<div><div>Understanding how cooling efficiency (CE) trends of tree covers respond to background climate change and local urbanization is essential for optimizing urban greening strategies. However, previous studies have largely focused on city-wide CE trends, often with a limited number of cities or with limited geographic scope. Consequently, the spatial heterogeneity in CE trends across urban–rural gradients, and their underlying drivers remain poorly understood, especially across cities worldwide. Here we quantified summer daytime CE trends from 2003 to 2020 across six urban–rural gradients (i.e., urban cores, new towns, urban fringes, suburbs, rural fringes, and rural backgrounds) in more than 5,000 global cities, utilizing MODIS-derived land surface temperature and tree cover data. We observed an inverse-V shape in CE trends along urban–rural gradient, with peak values in urban fringes (0.10 ± 0.01 °C/%/century, mean ± one standard error), followed by new towns, cores, and three rural gradients. The trends in CE were generally positive across most climate zones, yet arid regions exhibited a decline (−0.06 ± 0.06 °C/%/century). CE trends strengthened with increasing city size in urban fringes, yet they decreased in cores and new towns. Using a LightGBM-SHAP algorithm, we found that the macro-scale background climate dominated the CE trends in urban areas (43%), whereas micro-scale local surface properties emerged as the primary contributors in rural areas (48%). Our findings provide critical insights into the spatial heterogeneity of CE trends of urban tree covers at a global scale.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 210-222"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145732136","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"RSMT: Robust stereo matching training with geometric correction, clean pixel selection and loss weighting","authors":"Haoxuan Sun ,&nbsp;Taoyang Wang ,&nbsp;Qian Cheng ,&nbsp;Jiaxuan Huang","doi":"10.1016/j.isprsjprs.2025.12.014","DOIUrl":"10.1016/j.isprsjprs.2025.12.014","url":null,"abstract":"<div><div>Inaccurate or noisy labels have a huge impact on the training of deep learning models. To date, few studies have focused on the label error problem in satellite image stereo matching. In this paper, we analyzed and found the two open datasets US3D and WHU-Stereo contain label errors that cannot be overlooked. A new task is extremely necessary: learning from inaccurate labels with neural networks. Our motivation is to deal with label errors at the training level. A robust stereo matching training framework (RSMT) with geometric correction, clean pixel selection, and loss weighting modules is proposed. In addition, we also propose a dataset correcting method and provide two inaccurate-label stereo matching datasets US3D(E) and WHU(E) based on raw datasets. The framework can be applied to common stereo methods like IGEV-Stereo and ACVNet to achieve SOTA performance on the corrected datasets. To the best of our knowledge, the study is the first systemic inaccurate-label learning framework dedicated to stereo matching. Datasets are available at <span><span>https://github.com/endu111/robust-satellite-image-stereo-matching</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 421-436"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145884095","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"EnBoT-SORT: Hierarchical fusion-association tracking with pseudo-sample generation for dense thermal infrared UAVs","authors":"Jinxin Guo ,&nbsp;Weida Zhan ,&nbsp;Yu Chen ,&nbsp;Depeng Zhu ,&nbsp;Yichun Jiang ,&nbsp;Xiaoyu Xu ,&nbsp;Deng Han","doi":"10.1016/j.isprsjprs.2025.12.001","DOIUrl":"10.1016/j.isprsjprs.2025.12.001","url":null,"abstract":"<div><div>Thermal infrared dense UAV target detection and tracking present significant challenges at both data and algorithmic levels. At the data level, there exists a scarcity of accurately annotated real-world samples coupled with high acquisition costs. At the algorithmic level, the key difficulty lies in addressing frequent identity switches caused by highly dense target clustering, frequent occlusions, and reappearances. To overcome these challenges, this paper proposes an innovative infrared pseudo-sample generation paradigm by designing a physically-driven Heterogeneous Interactive Degradation Model (HIDM). This model simulates real infrared imaging through background-target cooperative degradation mechanisms that account for multiple coupled degradation factors, combined with a random trajectory generation strategy to produce large-scale physically realistic pseudo-sample data, significantly enhancing the domain adaptability of the generated data. Building upon this foundation, we propose a hierarchical fusion-association tracking framework—EnBoT-SORT. This framework employs YOLOv12 as a powerful detector and innovatively incorporates a dynamic target density regulator, a hybrid feature association engine, and a trajectory continuity enhancement module into BoT-SORT, effectively maintaining the continuity and stability of target IDs. Experimental results demonstrate that EnBoT-SORT significantly outperforms existing trackers in intensive UAV motion scenarios, achieving state-of-the-art performance on the IRT-B and IRC-B datasets with HOTA scores of 68.7% and 67.3%, and MOTA scores of 76.2% and 74.6%, respectively. Furthermore, cross-modal experiments on real infrared and visible-light datasets indicate that EnBoT-SORT possesses strong generalization capabilities. This work provides a comprehensive solution for infrared-intensive UAV tracking, spanning from data generation to algorithmic optimization. Our code and datasets are available at <span><span>GitHub</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 138-154"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145704860","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Automated TLS multi-scan registration in forest environments: A novel solution based on hash table","authors":"Xiaochen Wang,&nbsp;Xinlian Liang","doi":"10.1016/j.isprsjprs.2025.11.021","DOIUrl":"10.1016/j.isprsjprs.2025.11.021","url":null,"abstract":"<div><div>Terrestrial laser scanning (TLS) has proven to be an effective tool for forest inventories due to its accurate, non-destructive capability to document 3D space structures. The multi-scan mode of TLS enables comprehensive data acquisition, but the point cloud of each scan must be aligned to a common coordinate frame. In practice, the most common solution involves manually placing artificial markers in the field, which is time-consuming and labor-intensive. Consequently, the automated multi-scan registration method is highly appreciated for subsequent applications. This study presents an automated TLS multi-scan registration algorithm for forest point clouds, HashReg, utilizing the high-efficiency operations of Hash Table. HashReg comprises four key procedures, including stem mapping, estimating coarse transformation parameters, factor graph optimization, and fine-tuned registration. Using optimized transformation parameters, the global poses of individual TLS scans are subsequently determined within a unified coordinate system through a depth-first strategy. Extensive experiments were performed on four datasets with diverse forest characteristics, such as dense and sparse stems, flat and undulating terrain, and natural and plantation forests. The experimental results demonstrate that HashReg achieves milliradian-level rotation accuracy and centimeter-level translation accuracy, i.e., 0–3 mrad and 0–3 cm for most of the plots, respectively. Another evaluation metric, the point-wise upper bound errors, is reported to show the variation of point discrepancy with increasing distance. For most plots, these errors remained within the centimeter range, i.e., 1–4 cm, 1–5 cm, and 2–7 cm for the distance at 5 m, 10 m, and 20 m respectively. Moreover, the efficiency of HashReg’s four key procedures was also assessed. The running time of coarse registration and global optimization is at the millisecond level, i.e., 4 ms and 6 ms, while the stem mapping and fine registration were at the second level, i.e., 3 s and 15 s. Comparison with four state-of-the-art (SOTA) point cloud registration approaches, including FMP + BnB, HL-MRF, GlobalMatch, and SGHR, was quantitatively conducted on three public datasets. HashReg achieves superior accuracy, i.e., ranking first or second across all plots, with 100 % successful registrations. It also has substantially higher efficiency, with runtime improvements exceeding two-fold relative to the SOTA methods. All these advantages demonstrate that HashReg can bridge the gap between raw data and practical applications. The implementation of HashReg is open-sourced at <span><span>https://github.com/MSpace-WHU/Forest_TLS_Reg</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 280-304"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145785301","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"AEOS: Active Environment-aware Optimal Scanning Control for UAV LiDAR-Inertial Odometry in Complex Scenes","authors":"Jianping Li ,&nbsp;Xinhang Xu ,&nbsp;Zhongyuan Liu ,&nbsp;Shenghai Yuan ,&nbsp;Muqing Cao ,&nbsp;Lihua Xie","doi":"10.1016/j.isprsjprs.2026.01.006","DOIUrl":"10.1016/j.isprsjprs.2026.01.006","url":null,"abstract":"<div><div>LiDAR-based 3D perception and localization on unmanned aerial vehicles (UAVs) are fundamentally limited by the narrow field of view (FoV) of compact LiDAR sensors and the payload constraints that preclude multi-sensor configurations. Traditional motorized scanning systems with fixed-speed rotations lack scene awareness and task-level adaptability, leading to degraded odometry and mapping performance in complex, occluded environments. Inspired by the active sensing behavior of owls, we propose AEOS (Active Environment-aware Optimal Scanning), a biologically inspired and computationally efficient framework for adaptive LiDAR control in UAV-based LiDAR-Inertial Odometry (LIO). AEOS combines model predictive control (MPC) and reinforcement learning (RL) in a hybrid architecture: an analytical uncertainty model predicts future pose observability for exploitation, while a lightweight neural network learns an implicit cost map from panoramic depth representations to guide exploration. To support scalable training and generalization, we develop a point cloud-based simulation environment with real-world LiDAR maps across diverse scenes, enabling sim-to-real transfer. Extensive experiments in both simulation and real-world environments demonstrate that AEOS significantly improves odometry accuracy compared to fixed-rate, optimization-only, and fully learned baselines, while maintaining real-time performance under onboard computational constraints. The project page can be found at <span><span>https://kafeiyin00.github.io/AEOS/</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 476-491"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145902908","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Mapping land uses following tropical deforestation with location-aware deep learning","authors":"Jan Pišl ,&nbsp;Gencer Sumbul ,&nbsp;Gaston Lenczner ,&nbsp;Camilo Zamora ,&nbsp;Martin Herold ,&nbsp;Jan Dirk Wegner ,&nbsp;Devis Tuia","doi":"10.1016/j.isprsjprs.2025.12.007","DOIUrl":"10.1016/j.isprsjprs.2025.12.007","url":null,"abstract":"<div><div>The rates of tropical deforestation remain alarmingly high. To enable effective, targeted policy responses, detailed data on its driving forces is needed—each deforestation event needs to be attributed to an agricultural commodity or another land use. Remote sensing allows us to monitor land use conversion following deforestation, providing a proxy of drivers. However, recognizing individual commodities is challenging due to spectral similarities, the limited spatial resolution of free satellite imagery, and limited labeled data. To tackle these challenges, we propose a deep learning, multi-modal approach for the recognition of post-deforestation land uses from a time series of Sentinel-2 images, geographic coordinates, and country-level statistics of deforestation drivers. To integrate the modalities, we design a Transformer-based model with modality-specific encoders. The approach reaches 87% accuracy, an improvement of 10% over the image-only baseline, with little increase in data volume, computations, and model size. It works well in low-data regimes, and can be easily extended to include other modalities. Overall, this work contributes towards detailed, repeatable, and scalable mapping of deforestation landscapes, providing necessary data for the design and implementation of targeted interventions to protect tropical forests.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"232 ","pages":"Pages 578-593"},"PeriodicalIF":12.2,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145925544","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}