Domain-Incremental Remote Sensing Change Detection via Difference-Guided Adaptation and Frequency-Decoupled Distillation

Remote sensing change detection (RSCD) models often suffer from catastrophic forgetting when adapted to new domains (e.g., different sensors, seasons, or geographic regions). The authors investigate how to maintain high performance on new domains while preserving historical knowledge without needing to store or replay past training data.

The authors propose the DG-FDD framework, which consists of two primary components:

• Difference-Guided Dynamic Adapter (DGDA): Unlike standard adapters that process temporal branches independently, DGDA explicitly models bitemporal feature discrepancies. This allows the model to prioritize change-relevant features and suppress domain-specific noise (such as seasonal variations) during adaptation.
• Frequency-Decoupled Knowledge Distillation (FDKD-CS): This strategy separates structural information from domain-specific style in the frequency domain. By synthesizing cross-domain features—combining historical amplitude priors with current phase information—the model enables stable knowledge transfer from the frozen teacher to the student, even when the source and target domains have significant distribution shifts.

DG-FDD demonstrates strong performance in mitigating catastrophic forgetting across various incremental protocols. In experiments involving six two-domain sequences, the model achieved mean relative changes in F1 and IoU of only -0.23% and -0.45%, respectively. In more challenging three-domain sequences, the performance degradation remained minimal (-0.69% F1 and -1.31% IoU), indicating that the framework successfully balances stability (retaining old knowledge) and plasticity (learning new domains).

In real-world remote sensing, data distributions are rarely static. This research provides a memory-efficient way to deploy change detection systems that can evolve with new data without requiring massive retraining or the storage of sensitive historical datasets. By focusing on bitemporal discrepancies and frequency-domain structural consistency, the approach offers a robust solution for long-term environmental and urban monitoring.