Hyperspectral image recognition based on lightweight causal convolutional network

Abstract

Due to the phenomenon of mixed pixels in hyperspectral images, the spatial relationship between pixels in feature recognition is more important than in common image classification. Traditional convolutional neural networks can extract spatial structure features, but it cannot accurately express the spatial causality between image pixels. A lightweight deep learning network architecture is proposed based on causal convolution, which can effectively mine the intrinsic causality between pixels in hyperspectral images. Through comparative experiments in three public datasets, it is shown that compared with the traditional convolutional neural network, this network not only maintains the accuracy in ground objects recognition but also greatly reduces the model parameters. The structure proposed in this paper has high practical value in the recognition task of hyperspectral image.