Unsupervised remote sensing image segmentation based on a dual autoencoder

Abstract. Aimed at solving the limitation of the traditional unsupervised segmentation method of remote sensing images that rely more on features and prior knowledge, an unsupervised remote sensing image segmentation method combined with the atrous convolution autoencoder and the bidirectional long- and short-time memory autoencoder autoencoder (BiLSTM) network is proposed. The multispectral remote senses images using the dimensional reduction capacity of the autoencoder to obtain the transverse low-dimensional spatial features of the image through the atrous convolution autoencoder, and they use the BiLSTM autoencoder to learn the longitudinal low-dimensional spectral features of the multiband sequence of the remote sensing image. The batch normalization layer is added to optimize the whole network, and the visualization method is used to dynamically train and adjust the model. The data clustering of two kinds of low-dimensional features is obtained by the softmax algorithm, and unsupervised image segmentation of the multispectral data is finally realized by fusing the clustering results. In the experiment, the original data of Landsat8 remote sensing images were clustered and segmented directly. The results show that the accuracy is up to 83.25% through the clustering of dual autoencoder network in the water information extraction, which is obviously better than the common clustering algorithm and can get better image segmentation results.