ISPRS Journal of Photogrammetry and Remote Sensing最新文献

{"title":"VectorLLM: Human-like extraction of structured building contours via multimodal LLMs","authors":"Tao Zhang ,&nbsp;Shiqing Wei ,&nbsp;Shihao Chen ,&nbsp;Wenling Yu ,&nbsp;Muying Luo ,&nbsp;Shunping Ji","doi":"10.1016/j.isprsjprs.2026.01.025","DOIUrl":"10.1016/j.isprsjprs.2026.01.025","url":null,"abstract":"<div><div>Automatically extracting vectorized building contours from remote sensing imagery is crucial for urban planning, population estimation, and disaster assessment. Current state-of-the-art methods rely on complex multi-stage pipelines involving pixel segmentation, vectorization, and polygon refinement, which limits their scalability and real-world applicability. Inspired by the remarkable reasoning capabilities of Large Language Models (LLMs), we introduce VectorLLM, the first Multi-modal Large Language Model (MLLM) designed for regular building contour extraction from remote sensing images. Unlike existing approaches, VectorLLM performs corner-point by corner-point regression of building contours directly, mimicking human annotators’ labeling process. Our architecture consists of a vision foundation backbone, an MLP connector, and an LLM, enhanced with learnable position embeddings to improve spatial understanding capability. Through comprehensive exploration of training strategies including pretraining, supervised fine-tuning, and direct preference optimization across WHU, WHU-Mix, and CrowdAI datasets, VectorLLM outperforms the previous SOTA methods. Remarkably, VectorLLM exhibits strong zero-shot performance on unseen objects including aircraft, water bodies, and oil tanks, highlighting its potential for unified modeling of diverse remote sensing object contour extraction tasks. Overall, this work establishes a new paradigm for vector extraction in remote sensing, leveraging the topological reasoning capabilities of LLMs to achieve both high accuracy and exceptional generalization. All code and weights will be available at <span><span>https://github.com/zhang-tao-whu/VectorLLM</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 55-68"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146001038","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":"Multi-object tracking of vehicles and anomalous states in remote sensing videos: Joint learning of historical trajectory guidance and ID prediction","authors":"Bin Wang ,&nbsp;Yuan Zhou ,&nbsp;Haigang Sui ,&nbsp;Guorui Ma ,&nbsp;Peng Cheng ,&nbsp;Di Wang","doi":"10.1016/j.isprsjprs.2026.01.038","DOIUrl":"10.1016/j.isprsjprs.2026.01.038","url":null,"abstract":"<div><div>Research on multi-object tracking (MOT) of vehicles based on remote sensing video data has achieved breakthrough progress. However, MOT of vehicles in complex scenarios and their anomalous states after being subjected to strong deformation interference remains a huge challenge. This is of great significance for military defense, traffic flow management, vehicle damage assessment, etc. To address this problem, this study proposes an end-to-end MOT method that integrates a joint learning paradigm of historical trajectory guidance and identity (ID) prediction, aiming to bridge the gap between vehicle detection and continuous tracking after anomalous states occurrence. The proposed network framework primarily consists of a Frame Feature Aggregation Module (FFAM) that enhances spatial consistency of objects across consecutive video frames, a Historical Tracklets Flow Encoder (HTFE) that employs Mamba blocks to guide object embedding within potential motion flows based on historical frames, and a Semantic-Consistent Clustering Module (SCM) constructed via sparse attention computation to capture global semantic information. The discriminative features extracted by these modules are fused by a Dual-branch Modulation Fusion Unit (DMFU) to maximize the performance of the model. This study also constructs a new dataset for MOT of vehicles and anomalous states in videos, termed the VAS-MOT dataset. Extensive validation experiments conducted on this dataset demonstrate that the method achieves the highest level of performance, with HOTA and MOTA reaching 68.2% and 71.5%, respectively. Additional validation on the open-source dataset IRTS-AG confirms the strong robustness of the proposed method, showing excellent performance in long-term tracking of small vehicles in infrared videos under complex scenarios, where HOTA and MOTA reached 70.9% and 91.6%, respectively. The proposed method provides valuable insights for capturing moving objects and their anomalous states, laying a foundation for further damage assessment.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 383-406"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079956","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":"A novel transformer-based CO2 retrieval framework incorporating prior constraint and hierarchical features injection: assessment of transferability for Tansat-2","authors":"Lingfeng Zhang ,&nbsp;Lu Zhang ,&nbsp;Xingying Zhang ,&nbsp;Tiantao Cheng ,&nbsp;Xifeng Cao ,&nbsp;Tongwen Li ,&nbsp;Dongdong Liu ,&nbsp;Yang Zhang ,&nbsp;Yuhan Jiang ,&nbsp;Ruohua Hu ,&nbsp;Haiyang Dou ,&nbsp;Lin Chen","doi":"10.1016/j.isprsjprs.2026.01.039","DOIUrl":"10.1016/j.isprsjprs.2026.01.039","url":null,"abstract":"<div><div>Carbon dioxide (CO₂), the primary contributor to global warming, significantly impacts global climate change. Remote sensing is an effective approach for monitoring atmospheric CO₂ concentrations. However, the commonly used satellite’s full-physics Optimal Estimation (OE) method is time-consuming and requires advanced equipment. Additionally, traditional deep learning algorithms for satellite CO₂ retrieval suffer from limitations in accuracy and an inability to extrapolate effectively to unseen high values, caused by the gradually increasing concentrations over time. Balancing both efficiency and extrapolation capabilities is a critical task, especially for the next generation of large-swath carbon satellite with a significant increase in data volumes, such as Tansat-2. In this study, we first employed the OCO-2 data from 2020 to construct a Transformer-based structure and integrate prior constraint and hierarchical features injection mechanism for high precision CO₂ retrieval, and achieved an outstanding result with the R, RMSE, and MAPE of 0.939, 0.746 ppm, and 0.132 %. Based on the model, we evaluated its extrapolation capability using OCO-2 data from 2021 to 2024, demonstrating a robust performance and strong generalization ability (R = 0.938–0.951, RMSE = 1.083–1.310 ppm, MAPE = 0.208–0.256 %). Finally, we assessed the transferability of this model using simulated Tansat-2 data (August 18, 2020), achieving metrics of R = 0.657, RMSE = 1.299 ppm, and MAPE = 0.239 %, indicating the model’s effective transfer capabilities. The proposed model has the potential to provide a feasible solution for rapidly retrieving high-precision CO₂, especially for the next generation of large-swath carbon satellites.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 423-436"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146110789","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":"Monitoring global power outages induced by tropical cyclones using nighttime light data","authors":"Liujun Zhu ,&nbsp;Yaqian Li ,&nbsp;Shanshui Yuan ,&nbsp;Shi Shi ,&nbsp;Fang Ji","doi":"10.1016/j.isprsjprs.2026.01.042","DOIUrl":"10.1016/j.isprsjprs.2026.01.042","url":null,"abstract":"<div><div>Tropical cyclones (TCs) are among the most destructive natural hazards, frequently causing widespread power outages (POs) in coastal urban areas that disrupt economic activity and social stability. Quantifying TC-induced POs remains challenging due to limited outage data availability. This study made the first global detection and quantification of TC-induced POs using NASA’s Black Marble nighttime lights (NTL) data. The proposed method analyzed changes in NTL brightness within urban agglomerations by establishing pre-TC baselines and applying statistical outlier detection to identify outages. A total of 1,239 POs were detected by the algorithm from 19,999 agglomeration-TC events between 2012 and 2023, with the corresponding outage duration and severity being also estimated. Validation against media reports showed an overall accuracy of 0.78, with the accuracy being improved with TC intensity. Case studies demonstrated robust performance in regions with vulnerable infrastructure and high-quality NTL observations, such as North America, while performance declined in areas affected by frequent data gaps or rapid restoration, notably East Asia and India. While only 50% of agglomeration–TC events can be evaluated due to the missing NTL data, this work offers a scalable, near-real-time approach to global TC-induced PO monitoring, providing critical insights for urban resilience planning, disaster response, and power system management.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 437-451"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146135041","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":"SuperMapNet for long-range and high-accuracy vectorized HD map construction","authors":"Ruqin Zhou ,&nbsp;Chenguang Dai ,&nbsp;Wanshou Jiang ,&nbsp;Yongsheng Zhang ,&nbsp;Zhenchao Zhang ,&nbsp;San Jiang","doi":"10.1016/j.isprsjprs.2026.01.023","DOIUrl":"10.1016/j.isprsjprs.2026.01.023","url":null,"abstract":"<div><div>Vectorized high-definition (HD) map construction is formulated as the task of classifying and localizing typical map elements based on features in a bird’s-eye view (BEV). This is essential for autonomous driving systems, providing interpretable environmental structured representations for decision and planning. Remarkable work has been achieved in recent years, but several major issues remain: (1) in the generation of the BEV features, single modality methods suffer from limited perception capability and range, while existing multi-modal fusion approaches underutilize cross-modal synergies and fail to resolve spatial disparities between modalities, resulting in misaligned BEV features with holes; (2) in the classification and localization of map elements, existing methods heavily rely on point-level modeling information while neglecting the information between elements and between point and element, leading to low accuracy with erroneous shapes and element entanglement. To address these limitations, we propose SuperMapNet, a multi-modal framework designed for long-range and high-accuracy vectorized HD map construction. This framework uses both camera images and LiDAR point clouds as input. It first tightly couples semantic information from camera images and geometric information from LiDAR point clouds by a cross-attention based synergy enhancement module and a flow-based disparity alignment module for long-range BEV feature generation. Subsequently, local information acquired by point queries and global information acquired by element queries are tightly coupled by three-level interactions for high-accuracy classification and localization, where Point2Point interaction captures local geometric consistency between points of the same element, Element2Element interaction learns global semantic relationships between elements, and Point2Element interaction complement element information for its constituent points. Experiments on the nuScenes and Argoverse2 datasets demonstrate high accuracy, surpassing previous state-of-the-art methods (SOTAs) by 14.9%/8.8% and 18.5%/3.1% mAP under the hard/easy settings, respectively, even over the double perception ranges (up to 120 <span><math><mi>m</mi></math></span> in the X-axis and 60 <span><math><mi>m</mi></math></span> in the Y-axis). The code is made publicly available at <span><span>https://github.com/zhouruqin/SuperMapNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 89-103"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146015046","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":"Knowledge distillation with spatial semantic enhancement for remote sensing object detection","authors":"Kai Hu ,&nbsp;Jiaxin Li ,&nbsp;Nan Ji ,&nbsp;Xueshang Xiang ,&nbsp;Kai Jiang ,&nbsp;Xieping Gao","doi":"10.1016/j.isprsjprs.2026.01.017","DOIUrl":"10.1016/j.isprsjprs.2026.01.017","url":null,"abstract":"<div><div>Knowledge distillation is extensively utilized in remote sensing object detection within resource-constrained environments. Among knowledge distillation methods, prediction imitation has garnered significant attention due to its ease of deployment. However, prevailing prediction imitation paradigms, which rely on an isolated, point-wise alignment of prediction scores, neglect the crucial spatial semantic information. This oversight is particularly detrimental in remote sensing images due to the abundance of objects with weak feature responses. To this end, we propose a novel Spatial Semantic Enhanced Knowledge Distillation framework, called <span><math><msup><mrow><mi>S</mi></mrow><mrow><mn>2</mn></mrow></msup></math></span><em>EKD</em>, for remote sensing object detection. Through two complementary modules, <span><math><msup><mrow><mi>S</mi></mrow><mrow><mn>2</mn></mrow></msup></math></span><em>EKD</em> shifts the focus of prediction imitation from matching isolated values to learning structured spatial semantic information. First, for classification distillation, we introduce a Weak-feature Response Enhancement Module, which models the structured spatial relationships between objects and their background to establish an initial perception of objects with weak feature responses. Second, to further capture more refined spatial information, we propose a Teacher Boundary Refinement Module for localization distillation. It provides robust boundary guidance by constructing a regression target enriched with more comprehensive spatial information. Furthermore, we introduce a Feature Mapping mechanism to ensure this spatial semantic knowledge is effectively utilized. Through extensive experiments on the DIOR and DOTA-v1.0 datasets, our method’s superiority is consistently demonstrated across diverse architectures, including both single-stage and two-stage detectors. The results show that our <span><math><msup><mrow><mi>S</mi></mrow><mrow><mn>2</mn></mrow></msup></math></span><em>EKD</em> achieves state-of-the-art results and, in some cases, even surpasses the performance of its teacher model. The code will be available soon.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 144-157"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146033661","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":"Estimation of global riverine total phosphorus concentration based on multi-source data and stacked ensemble learning","authors":"Qi Li ,&nbsp;Lan Zhang ,&nbsp;Xi Chen ,&nbsp;Chen Zhang ,&nbsp;Jingyi Tian ,&nbsp;Xianghan Sun ,&nbsp;Liqiao Tian","doi":"10.1016/j.isprsjprs.2026.01.041","DOIUrl":"10.1016/j.isprsjprs.2026.01.041","url":null,"abstract":"<div><div>Quantifying riverine total phosphorus (TP) concentration at the global scale using remote sensing remains challenging because TP is not optically active and its spatial variability is strongly regulated by hydrological and environmental processes. In this study, a dataset at the global scale comprising 25,060 in situ TP measurements from 75 major river basins was used to examine how satellite-derived reflectance, river morphology, hydrological conditions, topography, and climate jointly constrain TP variability. The results demonstrate that integrating spectral and environmental predictors substantially improves the stability and transferability of TP estimation across heterogeneous river systems. Further improvements are achieved through stacked ensemble learning (R² = 0.80, RMSE = 0.5204, MAE = 0.3692), which effectively leverages the complementary strengths of different learning algorithms in processing both optical and environmental information. The resulting global riverine TP distribution patterns exhibit coherent latitudinal and regional gradients associated with river size, climatic regimes, and anthropogenic pressure, supporting the physical consistency of the estimates. Model explanation indicates that environmental factors such as elevation, river width, and discharge play key regulatory roles alongside spectral information. These findings demonstrate that integrating multi-source data and employing ensemble modeling approaches provides a viable pathway for large-scale estimation of non-optically active water quality parameters.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 588-608"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146153115","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":"Roadside lidar-based scene understanding toward intelligent traffic perception: A comprehensive review","authors":"Jiaxing Zhang ,&nbsp;Chengjun Ge ,&nbsp;Wen Xiao ,&nbsp;Miao Tang ,&nbsp;Jon Mills ,&nbsp;Benjamin Coifman ,&nbsp;Nengcheng Chen","doi":"10.1016/j.isprsjprs.2026.01.012","DOIUrl":"10.1016/j.isprsjprs.2026.01.012","url":null,"abstract":"<div><div>Urban transportation systems are undergoing a paradigm shift with the integration of high-precision sensing technologies and intelligent perception frameworks. Roadside lidar, as a key enabler of infrastructure-based sensing technology, offers robust and precise 3D spatial understanding of dynamic urban scenes. This paper presents a comprehensive review of roadside lidar-based traffic perception, structured around five key modules: sensor placement strategies; multi-lidar point cloud fusion; dynamic traffic information extraction;subsequent applications including trajectory prediction, collision risk assessment, and behavioral analysis; representative roadside perception benchmark datasets. Despite notable progress, challenges remain in deployment optimization, robust registration under occlusion and dynamic conditions, generalizable object detection and tracking, and effective utilization of heterogeneous multi-modal data. Emerging trends point toward perception-driven infrastructure design, edge-cloud-terminal collaboration, and generalizable models enabled by domain adaptation, self-supervised learning, and foundation-scale datasets. This review aims to serve as a technical reference for researchers and practitioners, providing insights into current advances, open problems, and future directions in roadside lidar-based traffic perception and digital twin applications.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 69-88"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146015045","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":"A geometric Cross-Propagation-Calibration method for SAR constellation based on the graph theory","authors":"Yitong Luo ,&nbsp;Xiaolan Qiu ,&nbsp;Bei Lin ,&nbsp;Zekun Jiao ,&nbsp;Wei Wang ,&nbsp;Chibiao Ding","doi":"10.1016/j.isprsjprs.2026.01.007","DOIUrl":"10.1016/j.isprsjprs.2026.01.007","url":null,"abstract":"<div><div>The networking capability of SAR constellations can effectively reduce the average revisit period, which has become a new trend in SAR Earth observation. However, the system electronic delay of several or even dozens of SAR satellites in a constellation must be calibrated and monitored for a long time to ensure high geometric accuracy of the product. In this paper, a geometric cross-propagation-calibration method for SAR constellations is proposed, which can calibrate the slant ranges of the SAR satellites in a constellation without any calibrators. The proposed method constructs a graph from all reference and uncalibrated SAR images involved in a cross-calibration task. For each uncalibrated image, the cumulative calibration error along paths originating from the reference images is estimated, enabling the identification of a path that minimizes this error. Cross-calibration is then performed sequentially along this optimal path. A closed-form expression is derived to estimate the cumulative calibration error along any path, which also reveals the underlying mechanism of error propagation in cross-calibration. Experiments based on real data show that the proposed method enables two China’s microsatellites, Qilu-1 and Xingrui-9, to achieve geometric accuracy of less than 5 m after calibration.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 346-359"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079954","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":"Adaptive image zoom-in with bounding box transformation for UAV object detection","authors":"Tao Wang ,&nbsp;Chenyu Lin ,&nbsp;Chenwei Tang ,&nbsp;Jizhe Zhou ,&nbsp;Deng Xiong ,&nbsp;Jianan Li ,&nbsp;Jian Zhao ,&nbsp;Jiancheng Lv","doi":"10.1016/j.isprsjprs.2026.01.036","DOIUrl":"10.1016/j.isprsjprs.2026.01.036","url":null,"abstract":"<div><div>Detecting objects from UAV-captured images is challenging due to the small object size. In this work, a simple and efficient adaptive zoom-in framework is explored for object detection on UAV images. The main motivation is that the foreground objects are generally smaller and sparser than those in common scene images, which hinders the optimization of effective object detectors. We thus aim to zoom in adaptively on the objects to better capture object features for the detection task. To achieve the goal, two core designs are required: (i) How to conduct non-uniform zooming on each image efficiently? (ii) How to enable object detection training and inference with the zoomed image space? Correspondingly, a lightweight offset prediction scheme coupled with a novel box-based zooming objective is introduced to learn non-uniform zooming on the input image. Based on the learned zooming transformation, a corner-aligned bounding box transformation method is proposed. The method warps the ground-truth bounding boxes to the zoomed space to learn object detection, and warps the predicted bounding boxes back to the original space during inference. We conduct extensive experiments on three representative UAV object detection datasets, including VisDrone, UAVDT, and SeaDronesSee. The proposed ZoomDet is architecture-independent and can be applied to an arbitrary object detection architecture. Remarkably, on the SeaDronesSee dataset, ZoomDet offers more than 8.4 absolute gain of mAP with a Faster R-CNN model, with only about 3 ms additional latency. The code is available at <span><span>https://github.com/twangnh/zoomdet_code</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50269,"journal":{"name":"ISPRS Journal of Photogrammetry and Remote Sensing","volume":"233 ","pages":"Pages 452-466"},"PeriodicalIF":12.2,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146135035","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}