计算机科学
人工智能
计算机视觉
图像分割
分割
遥感
图像处理
尺度空间分割
特征提取
模式识别(心理学)
可视化
目标检测
作者
Jianzhang Che,Hongquan Sun,Lyuzhou Gao,Jue Wang,Yanfei Zhong,Liang Chen,Ailong Ma
标识
DOI:10.1109/tgrs.2026.3740012
摘要
Pixel-level semantic segmentation of high-resolution uncrewed aerial vehicle (UAV) imagery is critical for rapid postdisaster assessment. Although real-time models support flexible on-board edge deployment, lightweight networks still face two challenges in unstructured postdisaster scenes. First, structural damage to affected land-cover objects makes them irregular and fragmented, while different disaster-related targets, as well as targets and surrounding backgrounds (BGs), may share similar local appearance and texture, causing confusion between targets and BGs and among semantically distinct but visually similar classes. Second, the limited channel representation capacity of lightweight models causes spatial detail loss during hierarchical resolution recovery, degrading boundary localization and small-object detail. To address these issues, we propose the prior-guided disaster network (PGDNet). During encoding, its synergistic guidance module (SGM) uses an explicitly supervised three-domain categorical prior to calibrate shallow spatial-frequency responses and condition the spatial feature transform (SFT) modulation, thereby enhancing class discrimination and alleviating semantic confusion. During decoding, the adaptive lightweight inverted bottleneck (ALIB) unit follows an alignment-first, refinement-later strategy to preserve small-target boundaries with low computational overhead. A boundary-aware hybrid loss further constrains damaged-target boundaries. The PGDNet-ResNet18 variant achieves mean intersection over union (mIoU) values of 69.83% and 78.85% on FloodNet and RescueNet, respectively. Its MobileNetV2 variant uses 2.31 M parameters and 5.52 GFLOPs and reaches 137.02 frames per second (FPS). These results demonstrate the feasibility of PGDNet for UAV edge deployment in terms of segmentation accuracy and computational efficiency.
科研通智能强力驱动
Strongly Powered by AbleSci AI