Improved Causal Temporal Fusion Transformer for Cloud Shadow Removal under Multi-temporal Satellite Reflectance Conditions

Abstract

Cloud shadow contamination produces spectrally depressed and spatially irregular reflectance observations in optical satellite image sequences, reducing the reliability of land-cover monitoring and biophysical index retrieval. This paper develops an improved Causal Temporal Fusion Transformer, termed Causal-TFT-SR, for cloud shadow removal under multi-temporal satellite reflectance conditions. The model integrates shadow-aware band embedding, strictly causal temporal attention, gated static-environment conditioning, and a reflectance-consistency reconstruction head. A multi-region experiment was designed with five land-surface types, seven spectral bands, and a 16-date temporal window. Compared with temporal linear interpolation, U-Net restoration, bidirectional LSTM fusion, and a non-causal transformer baseline, the proposed model reduced mean absolute reflectance error to 0.0127, increased SSIM to 0.913, decreased spectral angle to 3.41 degrees, and reduced NDVI error to 0.026. Ablation results further show that causal masking, variable selection, and radiometric consistency are all necessary for stable sequential restoration. The results indicate that causal temporal fusion can recover shadow-affected reflectance without relying on future observations, making the method suitable for near-real-time optical satellite production chains.