CPMF: An integrated technology for generating 30-m, all-weather land surface temperature by Coupling Physical model, Machine learning and spatiotemporal Fusion model

Although thermal remote sensing is the optimal method to measure large-scale Land Surface Temperature (LST), its application has been severely constrained due to the cloud contamination and the trade-off between temporal and spatial resolution. The integrated technology of LST gap-filling and downscaling is an effective method to break through these limitations. In this study, we proposed an integrated technology of gap-filling and downscaling to generate daily 30-m all-weather LST by Coupling Physical model, Machine learning, and spatiotemporal Fusion model, termed CPMF. CPMF comprises three modules: (I) estimating 1-km LST based on the surface energy balance theory (SEB-LST 1km ); (II) generating spatially complete 1-km LST coupling machine learning (CRLST 1km ); (III) all-weather 30-m LST from the CRLST 1km combining the spatiotemporal fusion downscaling and machine learning downscaling in an equal-weighted manner (CPMF-LST 30m ). Then, satellite data, reanalysis data, airborne data and in-situ LST data were used to evaluate the CPMF's performance. Results showed that: 1) SEB-LST 1km correlates well with clear-sky MODIS-LST (mean Pearson's R ≈ 0.70, mean RMSE ≈ 3.62 K). 2) CRLST 1km has a high correlation with MODIS-LST and reanalysis-LST, outperforming other four existing gap-filling products. 3) CPMF-LST30m achieves good accuracy, with Pearson's R of 0.86-0.96 (RMSE < 3.40 K) against Landsat-LST, R = 0.66 (P < 0.01) with airborne LST, and R = 0.97 (RMSE = 4.25) with in-situ LST, surpassing single-method downscaling. 4) Sensitivity analysis highlighted the importance of SEB-LST and CRLST in machine learning models, confirming the efficacy of the proposed physical model. CPMF provides a one-stop service for producing high-quality, long-term, all-weather LST data at a 30-m resolution.