MSDCN-LSTF: A Multiscale and Multidimensional Collaboration Network for Land Surface Temperature Spatiotemporal Fusion Based on Transformer and Convolution

High spatiotemporal resolution land surface temperature (LST) data is crucial for dynamic environmental monitoring. Spatiotemporal fusion is an effective means to obtain high spatiotemporal resolution remote sensing data. However, existing spatiotemporal fusion methods are still insufficient in dealing with the uncertain temporal fluctuations and complex spatial distribution of LST data. In this article, we propose a multiscale and multidimensional collaboration network (MSDCN-LSTF) combining convolutional neural network (CNN) and Transformer to fuse Landsat and MODIS LST products, which explicitly models spatial scale discrepancies, multidimensional feature interactions, and unstable spatiotemporal variation patterns. Specifically, a multiscale and multidimensional attention module (MSMDAM) is designed to enable collaborative interaction across channel-spatial dimensions and adjacent scale levels, enhancing the representation of complex spatial structures in LST. To mitigate redundant information in self-attention computation, an adaptive sparse transformer (AST) is introduced to jointly model sensor bias and temporal variations. Additionally, a novel weighting strategy (STS weighting) that considers the stability of spatiotemporal variation trends is designed to further address the rapid temporal and spatial changes of LST. Experiments conducted over three study areas demonstrate that MSDCN-LSTF generally outperforms existing advanced spatiotemporal fusion methods, achieving an average RMSE below 1.53 K and an average SSIM exceeding 0.97, with RMSE reductions ranging from 0.39 K to 1.40 K. Ablation studies further verify the effectiveness of MSMDAM and STS-weighting, which reduce RMSE by 0.28 K and 0.39 K, respectively. These results indicate that MSDCN-LSTF provides a robust and effective approach for high-precision LST spatiotemporal fusion.