Addressing the challenge of missing remote sensing data in cold and high-altitude regions: A Dual-branch Spatiotemporal Fusion GAN based on incomplete images

Continuous monitoring of surface dynamics in cold and high-altitude regions is of vital strategic importance. However, sensor limitations hinder the acquisition of complete high-spatiotemporal resolution images. Traditional spatiotemporal fusion models rely on cloud-free reference pairs, struggling with incomplete observations. To address this limitation, this study proposes a Dual-branch Spatiotemporal Fusion Generative Adversarial Network (DSTF-GAN) for incomplete images. The dual-branch architecture extracts features from temporal evolution and spatial detail, enabling the model to learn mapping relationships from cloud-affected multi-temporal pairs. Furthermore, a composite loss function and an artifact removal module are introduced to mitigate topographical distortions. This study presents a comprehensive comparative analysis against five state-of-the-art (SOTA) models, alongside a rigorous ablation study of the DSTF-GAN within a typical study area on the Tibetan Plateau. Experimental results demonstrate that the proposed model outperforms SOTA models in terms of structural similarity and spectral fidelity, especially maintaining robust reconstruction performance even under high missing rates in reference image. Furthermore, generalization experiments on heterogeneous sensor combinations confirm that the model possesses generalizability across diverse spatial resolution scales.