Research on the Extraction Method of Wetland River Network from Remote Sensing Image Based on MobileNetV2

Abstract

Estuarine wetlands serve as transitional zones between terrestrial river ecosystems and marine ecosystems, playing a significant role in ecological terms. Addressing the challenges in river network extraction from remote sensing images of estuarine wetlands, such as diverse target scales, high detail proportion, and substantial model computational load, this study proposes a multi-scale lightweight river network extraction method based on MobileNetV2. The method employs the MobileNetV2 backbone network to extract local features, incorporates a spectral attention mechanism to enhance the spectral response characteristics of water bodies, thereby improving the model's adaptability to complex spectral information. It also combines lightweight Depthwise Separable Atrous Spatial Pyramid Pooling to capture multi-scale contextual information, significantly reducing computational complexity while ensuring model accuracy. Finally, the method optimizes the feature fusion and upsampling process through a progressive feature fusion decoder, further enhancing the accuracy and detail restoration capability of river network segmentation. Experiments were conducted on a mixed dataset constructed from Sentinel-2 multispectral images and publicly available datasets. The results show that the method achieves a mean Intersection over Union (mIoU) of 94.5%, which is an improvement of 1.5% and 0.8% over DeepLabV3+ and DeepWaterMapV2, respectively. This method provides efficient and reliable technical support for ecological monitoring of estuarine wetlands and analysis of river network morphological evolution.