building footprint extraction

This study addresses the scarcity of high-resolution remote sensing data hindering rapid post-disaster building damage assessment. To this end, we propose a lightweight deep learning framework that fuses spatially and temporally aligned features from medium-resolution Sentinel-1 (SAR) and Sentinel-2 (optical) imagery, enabling robust cross-scene damage detection at 10 m resolution. We introduce xBD-S12—the first publicly available, co-registered, multi-temporal, multimodal damage detection dataset—specifically designed for this task. Experimental results demonstrate that medium-resolution SAR-optical fusion achieves effective damage mapping, challenging the prevailing assumption that architectural complexity inherently improves generalization across diverse disaster events; in fact, more complex models yield no significant gains in cross-disaster robustness. The released dataset, source code, and pre-trained models substantially advance open science and operational emergency remote sensing applications.

This work addresses the challenge of building footprint extraction from high-resolution remote sensing imagery, where complex structures and varying imaging conditions often degrade performance. Conventional approaches typically rely on multi-stage post-processing pipelines, leading to low efficiency and error propagation. To overcome these limitations, the authors propose PolyBuild, an end-to-end model that directly generates vectorized building polygons without any post-processing—a first in the field. PolyBuild integrates a CNN-Transformer hybrid architecture: an initial contour generation module jointly performs detection and coarse outline extraction, while a Transformer-based decoder refines the contour by effectively fusing local details with global contextual information. Extensive experiments on three benchmark building datasets demonstrate that PolyBuild significantly outperforms state-of-the-art mask-based and contour-based methods, confirming its superior accuracy and robustness.

This study addresses the challenge of unreliable building damage assessment from low-resolution remote sensing imagery following natural disasters. To this end, the authors propose a multi-stage AI framework that first employs a Video Restoration Transformer for super-resolution reconstruction, then utilizes YOLOv11 for building localization, and finally integrates multiple vision-language models (VLMs) to perform four-level semantic damage classification. The work introduces an innovative VLM-as-a-Jury mechanism to mitigate model bias and leverages CLIPScore for reference-free semantic alignment, thereby enhancing robustness in safety-critical scenarios. Evaluated on the Moore tornado and Matthew hurricane subsets of the xBD dataset, the method demonstrates superior performance in semantic interpretability and generation of actionable emergency response recommendations.

This study addresses the critical challenge of monitoring brick kilns in South Asia—a major source of air pollution and forced labor—where effective, real-time surveillance has long been lacking. The authors construct a dataset comprising 1.3 million high-resolution (0.149 m/pixel) satellite image tiles and propose ClimateGraph, a region-adaptive graph neural network that effectively captures the spatial layout and directional structure of brick kilns. For the first time, they systematically evaluate and compare the performance and complementarity of graph neural networks, remote sensing object detection pipelines, and foundational satellite vision models for this task. Experiments across five regions in South and Central Asia demonstrate the method’s effectiveness, establishing a scalable technical framework and performance benchmark for large-scale brick kiln monitoring.

Existing methods for automatic building footprint reconstruction from single satellite imagery suffer from low accuracy, geometric irregularities, and heavy reliance on manual post-processing. To address these limitations, this paper proposes an end-to-end framework integrating attraction field map guidance, multi-scale feature aggregation, and graph convolutional networks (GCNs). Specifically, we embed an attraction field map—encoding structural priors—into a Transformer backbone to explicitly model building topology; employ GCNs to refine boundary connectivity and enforce topological consistency; and fuse multi-resolution features to enhance robustness in complex scenes. Evaluated on standard benchmarks, our method achieves a 6% improvement in average precision (AP) and a 10% gain in average recall (AR) over state-of-the-art approaches. The reconstructed footprints exhibit superior geometric regularity and spatial coherence, enabling direct application to large-scale geospatial analytics, including urban planning and post-disaster assessment.

This work addresses the limited geographic coverage of existing building change detection datasets, which hinders the evaluation of model generalization across regions. To this end, we construct a large-scale benchmark dataset spanning 28 French departments, comprising approximately 28,000 pairs of 0.2-meter-resolution bitemporal remote sensing images with corresponding pixel-level change masks. Geographic isolation is enforced between training, validation, and test sets to enable rigorous assessment of domain transfer. Derived from IGN orthophotos and building footprints, the dataset includes both pixel- and instance-level annotations, enhanced by manual verification to ensure high quality and complete spatial metadata. Experiments demonstrate that models trained on this geographically diverse benchmark significantly outperform those trained on existing datasets such as LEVIR-CD+ and WHU-CD in cross-regional scenarios, underscoring the critical role of geographic diversity in improving model generalization.

This work addresses the critical delay in post-disaster or conflict-zone building damage assessment caused by satellite data downlink and ground processing bottlenecks. The authors propose an on-orbit AI system that compresses pre-event high-resolution optical imagery into compact latent representations uploaded to the satellite, enabling onboard comparison with post-event imagery for damage localization and classification. The method introduces, for the first time, a spaceborne-oriented latent-space compression scheme coupled with a robust comparison mechanism leveraging siamese networks, cross-attention modules, and targeted data augmentation to maintain high accuracy under extreme compression ratios and image misalignment. Experimental results demonstrate that the approach drastically reduces downlinked data volume while preserving task performance, thereby significantly enhancing the timeliness of emergency response.

This work addresses the critical challenge of rapid building damage mapping in the immediate aftermath of disasters, where paired pre- and post-event imagery and labeled data are typically unavailable. To enable non-AI experts to perform building-level damage assessment using only post-disaster satellite imagery, the authors develop a no-code web platform that innovatively integrates a pretrained vision foundation model, footprint-level feature pooling, and lightweight browser-based semantic segmentation. Leveraging few-shot interactive classification and active learning, the system produces area-wide damage assessments in real time with minimal user annotations. Evaluated on the xBD dataset, the approach achieves performance comparable to fully supervised ResNet-50 with only 1/20 of the annotation effort. Since 2023, the platform has supported over 30 real-world disaster responses, delivering high-temporal-resolution damage maps to humanitarian organizations within hours.

This study addresses the scarcity of post-disaster street-level imagery, which hinders accurate ground-level structural damage assessment, noting that while satellite imagery is readily available, it lacks a ground-level perspective. The work presents the first systematic exploration of generating post-disaster street views from satellite images, introducing two novel strategies: a vision-language model (VLM)-guided generation approach and a damage-sensitive Mixture-of-Experts (MoE) architecture, alongside a structure-aware evaluation framework. Experiments across 300 disaster scenarios show that ControlNet achieves the highest semantic accuracy (0.71), whereas the VLM and MoE methods produce more photorealistic textures but suffer from reduced semantic clarity, revealing a critical trade-off between visual realism and structural fidelity in synthetic street-view generation for disaster assessment.

This study addresses key challenges in post-disaster satellite imagery-based building damage assessment, including severe class imbalance, background clutter, and limited cross-disaster generalization. To overcome these issues, the authors introduce three modular enhancements to the ChangeMamba architecture: Focal Loss to mitigate class imbalance, a lightweight attention gating mechanism to suppress irrelevant background features, and a compact feature alignment module to enforce consistency between pre- and post-disaster feature spaces. The proposed approach significantly improves model robustness and cross-domain generalization, achieving performance gains of 0.8%–5% on in-domain evaluations and up to 27% improvement in cross-domain settings across multiple datasets, including xBD, Pakistan floods, Turkey earthquakes, and Hurricane Ida.