Hyperspectral Remote Sensing Image Segmentation Based on the Fuzzy Deep Convolutional Neural Network

Abstract

The "synonyms spectrum" and "foreign body with the spectrum" of remote sensing images have caused the traditional segmentation methods to be greatly limited. Existing segmentation methods represented by deep convolution neural network have made breakthrough progress. However, traditional deep learning is a completely deterministic model, which can not describe the data uncertainty well. To solve this problem, a new fuzzy deep neural network is proposed in this paper, called RSFCNN (Remote Sensing image segmentation with Fuzzy Convolutional Neural Network). The network integrates fuzzy unit and traditional convolution unit. Convolution unit is used to extract discriminant features with different proportions, thus providing comprehensive information for pixel-level remote sensing image segmentation. Fuzzy logic unit is used to deal with various uncertainties and provide more reliable segmentation results. In this paper, end-to-end training scheme is used to learn the parameters of fuzzy and convolution units. Experiments were carried out on the data set of ISPRS Vaihingen. According to the experimental results, the proposed method has higher segmentation accuracy and better performance than other algorithms.