KAN-Enhanced Transformer with Multiscale Parallel Pooling for Cloud Removal

Synthetic aperture radar (SAR) imagery penetrates clouds and provides valuable structural information of land surfaces under severe cloudy conditions, complementing optical cloud removal tasks. However, SAR’s inherent speckle noise often degrades removal quality. In this work, we propose a multimodal cloud removal method called Kolmogorov-Arnold network (KAN)-enhanced transformer for cloud removal (KAN-CR). This method integrates SAR and optical data through a KAN-enhanced transformer fusion module (KTFM). The multiscale parallel pooling attention (MPPA) modules contained in enhanced feature processing units (EFPU) enhance contextual perception and channel filtering. We also present a global seasonal observation dataset for cloud removal (GSOD-CR), comprising 10,000 carefully curated SAR–optical triplets from diverse regions worldwide. Experimental results show that KAN-CR achieves a PSNR improvement of approximately 0.49 dB over current state-of-the-art models on the GSOD-CR dataset.