Research on water body extraction based on a joint probability model of convolution neural network and spectral information

Abstract

This paper proposes a method of water body extraction from remote sensing images based on deep learning. Using Landsat images to establish a joint probability model of convolution neural network in which the full connection layer is replaced by the multi-layer perceptron and combined with spectral information. This paper trains the CNN model firstly and then puts the extracted features into the MLP classifier. Since the full connection layer of the CNN can obtain full information, using MLP to replace the full connection layer to train the neural network has a higher identification ratio. In addition, due to the insufficient utilization of spectral information from remote sensing images by CNN, and NDWI can highlight the water body characteristics in remote sensing images, using the joint probability model can effectively improve the water body extraction accuracy. In this paper, two water bodies in Jilin Province of China, Jinyue Lake, and Chagan Lake, are selected as the study areas. The water body extraction method based on a joint probability model of convolutional neural network and spectral information is compared with the common water body extraction methods, and the new algorithm has a kappa coefficient of 0.86 and an overall accuracy of 93.17% for the water body extraction of Jinyue Lake, and it has a kappa coefficient of 0.86 and an overall accuracy of 93% for the water body extraction of Chagan Lake. The new algorithm has improved the kappa coefficient and overall accuracy compared with other water body extraction methods, which proves the high identification ratio.