学部4年 石橋君が研究内容を，2026年8月9日（日）～14日（金）にWashington D.C.で開催された国際会議 IGARSS 2026 – 2026 IEEE International Geoscience and Remote Sensing Symposiumで口頭発表しました．
本研究は，季節による植生の変化や照明，影などの見かけの変化に左右されにくいリモートセンシング画像の変化検出を目指し，拡散モデルによる画像補完で生成した反実仮想的な変化前画像と，ラベルに基づくピクセル単位の対照学習を組み合わせた学習手法を提案するものです．
学会サイト：https://2026.ieeeigarss.org
Counterfactual-Guided Contrastive Learning for Robust Change Detection
O. Ishibashi, K. Ozaki and J. NakazawaAbstract:
Remote sensing change detection (CD) models often overfit to nuisance appearance shifts—seasonal foliage, illumination, shadows, and sensor artifacts—that spuriously correlate with change labels in a dataset-dependent manner. We propose a training-only framework that improves robustness to such spurious background cues by (i) generating counterfactual pre-change views using an existing diffusion-based inpainting model, and (ii) applying pixel-wise, label-conditioned contrastive learning that explicitly enforces invariance on unchanged pixels while maintaining discriminability for changed regions. A hard curriculum activates the contrastive term only after the segmentation branch stabilizes. Importantly, inference remains identical to a standard early-fusion baseline and requires only the original bi-temporal input pair. Experiments on three public benchmarks (LEVIR-CD, EGY-BCD, and CDD) show consistent gains on datasets with clear nuisance/label entanglement, particularly in low- to medium-data regimes, while also revealing boundary cases where aggressive invariance can be counterproductive.
