Remote Sensing Image Segmentation of Around Plateau Lakes Based on Multi-Attention Fusion

Land category monitoring in lake region around plateau provides decision-making basis for lake ecological protection and land resource planning. Aiming at the problem of low segmentation accuracy caused by scattered distribution and uneven scale of rivers, buildings and vegetation in remote sensing images of this region, a remote sensing semantic segmentation network integrating category and multi-scale attention is designed. The network adopts encoding-decoding end-to-end structure and constructs class and multi-scale attention modules based on depth residual neural network. Category attention makes a preliminary classification and spatial information filtering for the network feature layer, which is beneficial for the network to pay attention to category information and reduce pixel classification error; multi-scale attention combines mixed domain attention with multi-scale features, establishes context connection for different scale features, and improves the inherent segmentation and elimination problem of scattered small-scale targets. Experiments are performed on the semantic segmentation data set of remote sensing images around Dianchi Lake, and the test accuracy of the attention fusion semantic segmentation network designed in this paper reaches 77.4% and 86.3% under the average intersection ratio and average pixel accuracy index, respectively. From the overall segmentation effect, the fusion category and multi-scale attention segmentation network solve the segmentation and elimination problem of scattered small-scale target areas to a certain extent, and provide an effective basis for accurate monitoring and scientific planning of lakes around plateau. © 2023 Chinese Institute of Electronics. All rights reserved.