Partial unified learning for dynamic change detection in hyperspectral images

Hyperspectral image change detection (HSI-CD) aims to identify changes in bi-temporal hyperspectral images (HSIs) captured at different times in the same location. Existing algorithms often overlook the inherent class imbalance in HSI-CD, leading to poor generalization in detecting changes while introducing redundant computation in unchanged regions. This paper introduces a novel mechanism based on Partial Unified Learning for Dynamic Change Detection (PUL-DCD) to address these limitations. Particularly, a novel partial unified learning network is proposed, whose backbone is trained using multiple datasets, whilst the task-specific networks are trained independently with each individual dataset. In so doing, the network can maintain outstanding performance on specific datasets while having strong generalization ability. Furthermore, an innovative dynamic architecture is introduced that distinguishes between easy and hard regions for change detection, thereby optimizing parameter configuration and enhancing detection performance in challenging areas, while mitigating redundancy regarding unchanged information. Experimental results on three datasets show that PUL-DCD is competitive in both accuracy and efficiency.