Self-Sufficient Deep Learning Super-Resolution of MODIS Land Surface Temperature Data: Integrating Kernel-Based Regression and Transfer Learning

Land Surface Temperature (LST) is a vital geophysical parameter influencing environmental and biophysical processes. However, acquiring high spatiotemporal resolution LST data (≤100 m, ∼daily) is challenging due to the trade-off between spatial and temporal resolutions of satellite thermal infrared data. Spatiotemporal fusion of cross-sensor LSTs suffers from overpass time mismatches, varying viewing geometries, and cloud contamination, limiting the availability of clear images for statistical models. To address these limitations, we propose a deep learning super-resolution (DL-SR) approach that requires no additional data once trained, to enhance the resolution of Moderate Resolution Imaging Spectroradiometer (MODIS) LST, referred hereafter as the self-sufficient DL-SR. We developed a novel transfer-learning-based super-resolution model, TNetSR, which leverages a pre-trained Very Deep Super-Resolution (VDSR) network and is refined via a Multi-Residual U-Net. Furthermore, we integrated a kernel-based Random Forest (RF) regression method with DL-SR, which generated high-resolution training data using surface reflectance, spectral indices, albedo, digital elevation model, and land cover. These outputs and low-resolution MODIS LST were used to train DL-SR models, enabling the self-sufficient DL-SR of MODIS inputs. By evaluating across 9 Australian OzFlux sites, TNetSR outperforms comparative models, achieving a mean squared error of 1.027 K (test sites) and 1.229 K (external test site), a peak signal-to-noise ratio of 28.196 dB (test sites) and 27.216 dB (external test site), and a structural similarity index of 0.842 (test sites) and 0.838 (external test site) against the high-resolution dataset. Comparisons with in-situ LST yield a Pearson correlation coefficient (R) ≥ 0.94, a bias of 2.97 K (test sites) and 3.0 K (external site), and an unbiased root mean squared error of 3.67 K (test sites) and 3.71 K (external site). The results demonstrated that TNetSR consistently improved LST accuracy across landscapes, climatic zones, and seasons.