Deep Neural Networks-Based Remote Sensing Image Change Detection

With its powerful representation learning capability, deep learning effectively alleviates the poor robustness of manual feature extraction in remote sensing change detection. In this chapter, two deep learning—based approaches for remote sensing image change detection are presented. First, a generative representation learning network combined with cyclic clustering is proposed for unsupervised multiple change detection in multispectral images, which extracts spatial-temporal-spectral features via recurrent learning and adaptively infers the number of change categories through cyclic training. Second, a superpixel-level deep change feature analysis network is introduced to suppress noise and outliers in high-resolution remote sensing change detection. It selects training samples unsupervisedly and fine-tunes superpixel representations through supervised learning for robust classification.