STSTFN: Spatiotemporal–Spectral Land Surface Temperature Fusion Network

Abstract

Land Surface Temperature (LST) plays a crucial role in climate change research, ecological monitoring, and agricultural management. However, acquiring LST data with both fine temporal and fine spatial resolution remains a significant challenge. Existing downscaling methods are primarily constrained by insufficient capture of fine spatial details, and error accumulation inherent in multistep processing. To overcome these limitations and achieve robust, end-to-end LST downscaling, this article proposes a deep learning-based Spatiotemporal–spectral fusion network (STSTFN), which uses MODIS and Landsat LST products as base data and incorporates Sentinel-2’s near infrared (NIR) band as auxiliary input to enhance spatial detail. A multiscale convolutional fusion network is employed to learn the nonlinear mappings between the input and output data, enabling finer reconstruction of LST. Furthermore, STSTFN introduces a bi-temporal fusion strategy using a blur-weight difference-based weighting mechanism, which effectively integrates predictions from both temporal directions. Compared with a traditional method [spatiotemporal integrated temperature fusion model (STITFM)], and two deep learning-based methods [two-stage hierarchical spatiotemporal fusion network (THSTNet), and spatiotemporal temperature fusion network (STTFN)], STSTFN demonstrated superior performance in both quantitative evaluation (achieving an average root mean square error (RMSE) of 1.68°, a structural similarity index measure (SSIM) of 0.936, and a correlation coefficient (CC) of 0.931, which represent improvements of 32.9%, 6.2%, and 3.6% over the compared methods, respectively) and visual assessments, confirming its effectiveness and superiority in generating high-quality LST products.