利用高解析度遙測影像進行土石流偵測在災害管理中至關重要，然而，由於地形結構的非均向性複雜度以及嚴重的光譜領域差異——特別是在偽色彩或多光譜資料中——這項任務仍充滿挑戰 。現有的語意分割模型與基礎模型在處理這些問題時往往遭遇瓶頸，導致在破碎區域出現嚴重的偽陰性，或因邊界擴張而產生過度分割的情況 。 為了解決上述限制，本研究提出一種新穎的混合架構，將具可訓練偏差之靜態區塊白化模組與 Swin-UMamba 解碼器進行結合 。其中，區塊白化模組在整個訓練過程中持續進行動態適應，能有效對輸入通道進行去相關並減輕特定領域的輻射雜訊，同時避免過早鎖定偏差值 。隨後，Swin-UMamba 解碼器負責捕捉長距離相依性，以精準描繪細緻且高解析度的土石流邊界 。 實驗結果證明，本研究所提出的方法在三個具挑戰性的資料集上，皆取得了極具競爭力且SOTA 的表現 。本模型全面超越了多種廣泛使用的卷積神經網路與近年基於 Transformer 的架構 。此外，與基於基礎模型的 TransLandSeg 相比，本方法在其中兩個基準測試中展現了更優異的成效，並在剩餘的一個資料集上維持了極為相近的結果 。 最重要的是，本架構在達成這些穩健結果的同時，展現了卓越的計算效率 。模型僅需 47.24M 的參數與 58.01G 的浮點運算次數 (FLOPs)，大幅降低了運算負擔，成功克服了大型視覺模型常見的推論瓶頸 。本研究證實，此方法能在不依賴大規模通用預訓練的情況下，成功在精準度與召回率之間取得平衡，並在高度破碎的土石流場景中維持嚴格的形狀保真度。;Landslide detection using high-resolution remote sensing imagery is critical for disaster management， yet it remains challenging due to complex anisotropic topographical structures and severe spectral domain gaps， particularly in false-color or multi-spectral data. Existing semantic segmentation models and foundation models often struggle with these issues， leading to severe false negatives in fragmented areas or over-segmentation resulting from boundary dilation. To address these limitations， we propose a novel hybrid architecture that integrates a Frozen Patch-Whitening with Trainable Biases module with a Swin-UMamba decoder. The Patch-Whitening module continuously adapts throughout the training process to effectively decorrelate input channels and mitigate domain-specific radiometric noise without prematurely locking the bias. Subsequently， the Swin-UMamba decoder captures long-range dependencies to accurately delineate precise， fine-grained landslide boundaries. Experiments demonstrate that our proposed method achieves highly competitive and state-of-the-art performance across three challenging datasets. Our model comprehensively outperforms a wide range of established CNN-based and recent Transformer-based architectures. Furthermore， when compared to the foundation-model-based TransLandSeg， our approach demonstrates superior efficacy in two benchmarks while maintaining closely comparable results on the remaining dataset. Crucially， our architecture achieves these robust results with remarkable computational efficiency. Requiring only 47.24M parameters and 58.01G FLOPs， our model significantly reduces the computational burden， overcoming the inference bottlenecks typical of massive vision models. The results validate that our approach successfully balances precision and recall， maintaining strict shape fidelity in highly fragmented landslide scenarios without relying on large-scale generic pre-training.
