Deep Neural Networks-Based Heterogeneous Remote Sensing Image Change Detection

With the diversification of remote sensing observation platforms, change detection has increasingly shifted from homogeneous to heterogeneous multi-source scenarios. The core challenge lies in learning modality-invariant feature representations. In this chapter, two deep learning—based approaches are proposed to address this challenge. First, a self-guided autoencoder is introduced that iteratively refines pseudo-labels through fusion of multiple change maps, enabling robust change detection without requiring transformation or alignment. Second, a multi-layer composite autoencoder leverages minimal labeled data to supervise change detection across multiple feature layers, iteratively refining pseudo-labels through fused prediction confidence to improve performance.