Learning Semantic-Robust Change Detection via Semantic-Invariant Self-Distillation

Remote sensing change detection (CD) models often struggle with non-semantic variations—such as shadows, illumination shifts, and atmospheric changes—that do not represent actual land-cover changes. These models, typically pre-trained on single-view classification tasks, often misinterpret these appearance-level fluctuations as semantic changes, leading to high false alarm rates. The authors ask: can we explicitly train a model to distinguish between semantic changes and non-semantic environmental noise?

The authors propose SCDistill, a two-pronged framework designed to enhance robustness. First, they introduce a diffusion-based perturbation simulation pipeline that generates realistic, non-semantic variations (e.g., relighting, weather effects) on existing synthetic datasets. This exposes the model to diverse environmental conditions during training. Second, they employ a semantic-invariant self-distillation strategy. In this setup, a frozen Vision Foundation Model Encoder (VFME) acts as a teacher, providing stable semantic representations of clean images. A student encoder is then trained to produce identical semantic features even when the input image is subjected to the simulated non-semantic perturbations. This forces the student to learn representations that are invariant to appearance changes while retaining semantic content.

SCDistill achieves state-of-the-art performance across multiple semantic change detection benchmarks. By explicitly modeling non-semantic disturbances and distilling semantic invariance, the model significantly reduces false alarms caused by environmental fluctuations. Furthermore, the authors demonstrate that the learned representations generalize well to related tasks, including binary change detection and change captioning, confirming that the robustness gains are not limited to a single specific task.