Remote Sensing Semantic Segregation for Water Information Extraction: Optimization of Samples via Training Error Performance

Abstract

The corrupted information in training samples is an important factor affecting the accuracy and generalizability of the machine learning models. Due to the extremely high memory capacity of deep learning models, the interference of excessive corrupted information makes the model prone to bad generalization behavior. This paper proposes a method of estimating training sample quality using the value calculated by the loss function in the process of gradient descent optimization. The method includes a model accuracy variation degree algorithm and a sample quality analysis algorithm. The model accuracy variation degree algorithm provides a basis for determining the intervention time of the sample quality analysis algorithm by calculating the intensity of the model accuracy variation change. The data error evaluation algorithm analyzes the distribution characteristics of the training error and estimates the error degree of the training samples to control the quality of the input samples. This study includes a water segmentation experiment performed on GF1 remote sensing images, which demonstrates that the optimization method can significantly improve the model accuracy and training stability.