Network Architecture Search-Based Remote Sensing Image Change Detection

The preceding chapters have demonstrated the effectiveness of deep learning in both homogeneous and heterogeneous remote sensing change detection. However, reliance on fixed, manual architectures limits adaptability across diverse sensors, scenes, and modalities. Neural architecture search (NAS) offers a promising solution by automatically discovering task-specific network architectures in a data-driven manner, leading to more discriminative representations for change detection. In this chapter, two self-adaptive neural architecture search frameworks are presented to enhance the flexibility and robustness of change detection systems. First, an evolutionary NAS framework is introduced, which employs adaptive gene encoding, specialized genetic operators, and a self-adaptive selection mechanism to automatically discover simple, efficient networks for SAR images change detection. Second, a semi-supervised adaptive ladder network is introduced that adjusts its dual-input architecture and generates pseudo-labels by fusing semi-supervised and unsupervised outputs, enabling effective adaptation to homogeneous and heterogeneous image pairs with minimal labeled data.