Sensitivity Analysis of Parameters of U-Net Model for Semantic Segmentation of Silt Storage Dams from Remote Sensing Images

Abstract

Building silt storage dams is an important measure to control soil erosion. Sensitivity analysis of the parameters in a deep learning model is the premise of extracting high-precision silt storage dams from high-resolution remote sensing (RS) images. In this study, watershed features of Hulu River and Lanni River in the Loess Plateau, China, are extracted using a geographic information system and digital elevation model. The detection of silt storage dams using the U-Net model considered three high-resolution RS image datasets to evaluate the effect of different input sizes, batch sizes, and sample sizes on accuracies of silt storage dams. The results show that a large input size, batch size, and sample size can improve the accuracy of silt storage dams extracted by U-Net. U-Net with Dataset 3, input size of 576 × 576, and batch size of 4 achieved an overall accuracy of 96.26%, F1 score of 70.61%, mean intersection over union of 75.33%, training time of 485 ms/step, minimum noises and shadow, and clear outlines of silt storage dams. This study provides theoretical and practical decision-making for the planning, construction, and maintenance of silt storage dams, as well as ecological protection and high-quality development of the Yellow River Basin.