啟發自因果關係的單域泛化方法於遙測場景分類

Abstract

深度學習常假設訓練與測試資料獨立同分佈，但遙測影像因地理與季節變化，往往違反此假設。領域泛化，尤其是單域泛化，儘管在其他領域已有所研究，卻在遙測應用中鮮少被探討。為填補這一空白，我們基於 SEN12MS 與 SEN12MS‐CR 資料集，構建了首個單域泛化基準，將影像劃分為春、夏、秋、冬等四個領域，由於各個季節對應一組不重疊的地區集，因此此基準同時反映了地理與季節的差異。此外，我們也提出了一種基於因果關係且參數高效、適用於各種遙測基礎模型的微調方式，稱為因果微調。該方法透過逐樣本因子分解，從因果關係的角度出發，強化影像特徵的跨域不變性，並透過因果因子選擇機制，聚焦最具預測力的因子。我們在 CROMA 與 DeCUR 兩個遙測基礎模型上測試因果微調，結果顯示無論在晴空或多雲條件下，其性能均優於全參數微調、線性探針，以及六種先進的單域泛化方法，與全參數微調相比，因果微調提升了約 3-4% 的 F1 分數，為穩健的遙測跨域分類提供了實際可行的解決途徑。

Deep learning models often assume that training and test data are independent and identically distributed (IID), yet in remote sensing, this assumption often breaks due to geographic and seasonal variations. While domain generalization (DG) and particularly single‐domain generalization (SDG) serve as promising solutions, they are underexplored in the field of remote sensing. To fill this gap, we introduce a novel SDG benchmark constructed from SEN12MS and SEN12MS-CR, where images are partitioned into four domains based on season and geography. Each domain represents a unique season and contains a distinct, non-overlapping set of regions. This design captures both geographic and seasonal variations, reflecting real-world spatiotemporal shifts in remote sensing data. Building on this benchmark, we propose Causal Tuning, a causality-inspired parameter‐efficient, model‐agnostic tuning approach for remote sensing foundation models (RSFMs). Causal Tuning combines instancewise factorization, which enhances domain invariance from a causality perspective, with causal factor selection, a class‐guided cross‐attention module that aggregates only the most predictive factors. Integrated with two recent RSFMs, CROMA and DeCUR, Causal Tuning outperforms full fine‐tuning, linear probing, and six leading SDG baselines, achieving 3-4% F1 gains under both clear and cloudy conditions compared to full fine-tuning, paving a practical path to robust out‐of‐distribution (OOD) remote sensing scene classification.