ChangeFlow uses latent rectified flow to boost change detection in remote sensing by 1.3 F1 points
A new generative framework outperforms discriminative baselines on four benchmarks for change detection.
Remote sensing change detection (RSCD) traditionally relies on per-pixel discriminative classification that outputs a single mask per input pair. This approach fails to capture the context-dependent, often ambiguous nature of region-level annotations. ChangeFlow, developed by Blaž Rolih and colleagues, addresses these limitations with a generative framework that reframes change detection as the synthesis of a change mask in latent space using rectified flow. The model is conditioned on structured, lightweight signals and supports stochastic sampling, allowing multiple candidate masks to be aggregated for improved robustness and confidence estimation.
ChangeFlow achieves strong empirical results: an average F1 of 80.4% across four benchmark datasets, surpassing the previous state-of-the-art by 1.3 points on average, while maintaining inference speeds comparable to recent strong discriminative baselines. By generating coherent masks that respect global structure and providing per-pixel confidence through sample agreement, ChangeFlow offers a practical tool for real-world remote sensing applications where annotation ambiguity is common.
- ChangeFlow uses latent rectified flow to generate change masks, modeling a distribution of plausible outputs instead of a single deterministic prediction.
- It achieves an average F1 of 80.4% across four remote sensing benchmarks, improving by 1.3 points over the previous best method.
- Inference speed is comparable to strong discriminative baselines, while sampling-based ensembling provides confidence estimates for ambiguous regions.
Why It Matters
Better change detection in satellite imagery with built-in ambiguity handling and confidence scores for practical geospatial analysis.