【摘要】 【目的】无人机（Unmanned Aerial Vehicle,UAV）低空遥感因机动性强、响应迅速和空间分辨率高，已成为洪涝等突发灾害现场信息获取的重要手段。建筑物直接承载人口与基础设施，其受淹状态能够直观反映灾害影响范围和严重程度，对灾情评估和救援决策具有关键意义。然而，现有受淹建筑检测多依赖机下高精度分割与空间叠置分析，需等待影像回传且对分割稳定性敏感，难以满足应急响应的时效需求；同时，水体反射、阴影干扰及训练域差异等因素易导致机载轻量化分割结果出现边界破碎和结构缺失，限制了单景影像直接判定受淹状态的能力。为此，本文提出FloodSAM-Duo(Dual-scale Prompt-guided SAM for Flooded Building Extraction)方法，用于在机载端实现受淹建筑的轻量化对象级检测。【方法】该方法利用金字塔场景解析网络（Pyramid Scene Parsing Network,PSPNet）在双尺度下提取建筑与水体的弱语义响应，推断受淹候选建筑，并在候选区域内自动生成语义稳定的提示点，引导轻量化实例分割模型FastSAM(Fast Segment Anything Model)在局部范围内完成精细分割，从而直接获得结构完整、边界一致的建筑实例掩膜，形成弱语义判定与提示引导精分割相结合的协同推理机制。在此基础上，本文进一步将实例分割结果转化为可下传、可聚合的结构化灾情信息集合，从而降低对高分辨率影像持续回传的依赖。【结果】基于近两年国家减灾中心提供的随州、新乡等地区的洪涝无人机影像数据开展实验，结果表明与ENet、YOLOv8n-seg和RT-DETR Tiny等轻量化模型相比，FloodSAM-Duo在Precision、Recall、F1和mIoU等指标上均取得更优表现，其中F1值提高约18%, mIoU提升约17%，建筑边界完整度指标BF1提高约18%，并可在Jetson Xavier NX平台上实现15～20 ms的单帧推理延迟，满足无人机机载端对洪涝建筑灾情信息准实时获取的应用需求。【结论】FloodSAM-Duo能够在单景无人机影像上直接生成对象级受淹建筑信息，减少对高精度分割和机下处理的依赖，为洪涝灾害应急监测提供一种轻量高效的建筑灾情评估方案。更多还原

【Abstract】 [Objectives] Low-altitude remote sensing using unmanned aerial vehicles(UAVs), characterized by high mobility, rapid response, and high spatial resolution, has become an important means for acquiring on-site information during flood and other emergency disasters. Buildings directly accommodate population and infrastructure, and their inundation status can intuitively reflect the spatial extent and severity of disasters, providing critical support for damage assessment and rescue decision-making. However, existing flooded-building detection methods typically rely on high-precision segmentation and spatial overlay analysis performed off-board, which requires image transmission and is sensitive to segmentation stability, making them unsuitable for timecritical emergency response. In addition, complex flood backgrounds—such as water reflections, shadows, and domain shifts—often lead to fragmented boundaries and missing structures in lightweight onboard segmentation results, limiting the feasibility of direct inundation assessment from single images. To address these challenges, this study proposes Flood SAM-Duo(Dual-scale Prompt-guided SAM for Flooded Building Extraction) for lightweight object-level detection of flooded buildings on UAV platforms. [Methods] The proposed method employs a Pyramid Scene Parsing Network(PSPNet) to extract weak semantic responses of buildings and water at dual scales to infer flooded building candidates. Semantic-stable prompt points are then automatically generated within candidate regions to guide a lightweight instance segmentation model, Fast SAM, to perform local fine segmentation, producing structurally complete and boundary-consistent building masks. This forms a collaborative inference mechanism that integrates weak semantic reasoning with prompt-guided fine segmentation. Furthermore, segmentation outputs are converted into structured disaster information that can be transmitted and aggregated, thereby reducing the reliance on continuous high-resolution image transmission. This design enables the system to directly generate object-level inundation results from single UAV images while maintaining a lightweight computational structure suitable for onboard deployment. [Results] Experiments conducted on UAV flood imagery from Suizhou and Xinxiang, China, collected over the past two years, demonstrate that the proposed method significantly outperforms comparative approaches in building contour completeness and inundation detection reliability. Specifically, compared with representative lightweight models such as ENet, YOLOv8n-seg and RT-DETR Tiny, Flood SAM-Duo improves the F1-score by approximately 18% and the mIoU by about 17%, while achieving higher boundary consistency measured by the BF1 metric. Moreover, it achieves a singleframe inference latency of 15～20 ms on a Jetson Xavier NX platform, meeting the requirements for near-realtime onboard disaster information acquisition. [Conclusions] Flood SAM-Duo enables direct generation of object-level flooded-building information from single UAV images while reducing dependence on high-precision segmentation and off-board processing. The method provides a lightweight and efficient solution for rapid building damage assessment in flood emergency monitoring. By integrating weak semantic reasoning with prompt-guided instance segmentation, the proposed framework offers a practical strategy for improving the timeliness and reliability of UAV-based disaster monitoring systems.更多还原