采用深度学习智能提取建设用地,对定量评价、监测和预测长江流域水土流失具有重要作用.本文基于开源数据,采用半自动标注方式标注建设用地,构建了面向长江流域水土流失监测的建设用地遥感数据集.该数据集在多个深度学习语义分割模型(FPN、PSPNet、DeepLabV3+、UN-et++、Swin-Transformer)测试中的总体精度(overall accu-racy,OA)均优于93.00％,均交并比(mean intersection over union,MIoU)优于70％,具有较高有效性,可推动遥感智能解译在水土流失监测中的应用.
Abstract
The intelligent extraction of construction land using deep learning plays an important role in quantitative evalua-tion,monitoring and prediction of soil erosion in the Yangtze River Basin. Based on open source data,this paper adopts a semi-automatic labeling method to annotate samples,and con-structs remote sensing dataset of construction land for monitor-ing soil and water loss in the Yangtze River Basin. The over-all accuracy of multiple deep learning semantic segmentation networks (FPN,PSPNet,DeepLabV3+,UNet++,Swin-Transformer) with the testing dataset is higher than 93.00％,and MIoU is higher than 70％,which has high ef-fectiveness and can promote the application of remote sensing intelligent interpretation in soil erosion monitoring.
关键词
深度学习/水土流失监测/语义分割/智能解译
Key words
deep learning/soil erosion monitoring/semantic segmentation/intelligent interpretation
维普
万方数据