基于改进VGG16-UNet的山区遥感影像地形阴影检测模型

Topographic Shadow Detection Model in Mountainous Remote Sensing Images Based on Improved VGG16-UNet

摘要:

山区地形阴影在遥感卫星影像中普遍存在形态不规则与边界复杂性，导致传统检测方法难以实现精准分割。为解决这一难题，本文提出一种融合可变形卷积与坐标注意力机制的增强型VGG16?UNet语义分割模型，以提升阴影区域的识别与定位能力。模型以VGG16作为编码器主干网络，通过引入可变形卷积动态调整采样位置，有效捕捉阴影的不规则邻域特征;同时嵌入坐标注意力模块，强化空间位置信息与通道特征的协同表达，优化细节恢复与结构一致性。利用自建数据集和国产高分七号卫星遥感影像开展验证实验，结果表明：本文方法平均交并比、平均召回率与总体精度分别达到94.77%、97.28%与97.52%，相较于VGG16?UNet基准模型分别提高0.62、0.41、0.30个百分点。在多种山区场景下测试进一步证实了本文方法具备稳定且可靠的阴影检测能力，具有良好的泛化性能与鲁棒性，为高精度地形阴影自动提取提供了可靠技术路径。

Abstract:

Mountainous terrain shadows in remote sensing satellite imagery typically exhibit irregular morphologies and complex boundaries, which pose significant challenges to accurate segmentation using conventional methods. To address this issue, an improved VGG16?UNet semantic segmentation model that integrated deformable convolution and a coordinate attention mechanism was proposed, aiming to improve the recognition and localization of shadow regions. The model employed VGG16 as the backbone encoder, where deformable convolution was introduced to dynamically adjust sampling locations, thereby effectively capturing features within irregular shadow neighborhoods. Simultaneously, a coordinate attention mechanism was embedded to enhance the synergistic representation of spatial positional information and channel?wise features, optimizing detail recovery and structural consistency. Validation experiments conducted on a self?built dataset and domestic GaoFen?7 satellite imagery showed that the proposed method achieved mean intersection over union (mIoU), mean recall (mRecall), and overall accuracy (OA) scores of 94.77%, 97.28%, and 97.52%, respectively. These results represented improvements of 0.62 percentage points, 0.41 percentage points, and 0.30 percentage points over the baseline VGG16?UNet model. Furthermore, tests across diverse mountainous scenarios confirmed that the method possessed stable and reliable shadow detection capabilities, along with strong generalization performance and robustness. This work can provide a reliable technical pathway for the automated extraction of high?precision terrain shadows.