摘要
Abstract
Surface water plays an important role in the global ecological environment and human life.Dynamically capturing the distribution and extent of surface water on Earth is necessary.However,due to the high complexity of land surface environments,existing surface water body detection methods have limitations in applicability and accuracy,especially in highly heterogeneous regions such as urban areas,moun-tains,and cloud-covered areas.To improve the recognition accuracy of different types of water bodies in different land surface environments,this study proposes a water body detection method for remote sensing images based on multi-scale feature fusion(MFWD).The proposed method first extracts multi-level features of water bodies and land surfaces based on a deep residual network model.Then,an Atrous Spatial Pyramid Pooling(ASPP)module and a Channel-Spatial Attention Mechanism(CSAM)module are designed to fully exploit advanced seman-tic information and capture advanced features of water bodies.Finally,cross-scale connections are utilized to fuse multi-scale low-level spa-tial details and high-level semantic information,obtaining comprehensive feature representations for effective water body recognition.Experi-ments on Sentinel-2 data demonstrate that the proposed MFWD method achieves an overall recognition accuracy of 95.6%,exhibiting im-proved accuracy in identifying different types of water bodies.Moreover,the detection of small-scale water bodies as well as water bodies in highly heterogeneous regions is enhanced.关键词
空洞空间金字塔池化/注意力机制/多尺度特征融合/遥感影像/水体检测Key words
atrous spatial pyramid pooling/attention mechanism/multi-scale feature fusion/remote sensing images/water bodies detection分类
信息技术与安全科学
