A novel surface energy balance-based approach to land surface temperature downscaling
Mohammad Karimi Firozjaei,Naeim Mijani,Majid Kiavarz,Si-Bo Duan,Peter M. Atkinson,Seyed Kazem Alavipanah
DOI: https://doi.org/10.1016/j.rse.2024.114087
IF: 13.5
2024-03-02
Remote Sensing of Environment
Abstract:Spatial downscaling satellite sensor-derived land surface temperature (LST) is of great importance for various environmental applications. However, the energy balance at the land surface is complex, especially in urban environments. As a result, the complexity of land surface thermal processes and the resulting LST cannot be accurately modeled using common downscaling methods. Here, we propose a new physical approach for LST downscaling based on surface energy balance (SEB) components which can capture this complexity and, thereby, increase the accuracy of downscaling. Experiments to demonstrate the effectiveness and accuracy of the proposed SEB approach were undertaken across 23 test sites in Europe and the United States. The research data included the Landsat 8 and Terra satellite sensor imagery products, land cover data, a digital elevation model, surface imperviousness maps and ground station data. Downscaled LST (DLST) maps at different spatial resolutions were produced using the developed SEB-based approach and compared to common benchmark methods including random forest regression (RFR), geographically weighted regression (GWR), area-to-point regression kriging (ATPRK), disaggregation of radiometric surface temperature (DisTrad), and three layers composition (TLC). Using the observed and In-situ LST, and soil temperature data in different land covers and seasonal conditions as the reference, the accuracy of the SEB and common benchmark approaches for LST downscaling were evaluated and compared. The average RMSE (bias) between the observed LST and DLST derived by SEB, RFR, GWR, ATPRK, DisTrad and TLC approaches for the different test areas were 1.15 (−0.42), 1.45 (0.91), 2.11 (1.02), 2.36 (−0.8), 2.79 (−2.31), and 2.45 (−1.48) K, respectively. Accuracy evaluation based on In-situ LST data showed that the RMSE decreased by 0.47–1.96 K (∼1.28 K on average) and the bias decreased by 0.15–0.98 K (∼0.57 K on average) when the SEB method was applied for LST downscaling in place of the other approaches, respectively. After the SEB approach, RFR approach had higher accuracy than other approaches. The mean RMSE between the observed LST product and the SEB (RFR) derived DLST at spatial resolutions of 480, 320, 240, 192, 120 and 30 m in the warm season were 0.89 (0.95), 0.93 (1.10), 0.99 (1.42), 1.07 (1.52), 1.21 (1.66), and 1.23 (1.94) K, respectively. In the cold season, these values were 0.52 (0.60), 0.58 (0.75), 0.67 (0.93), 0.78 (1.04), 0.87 (1.18), and 0.92 (1.36) K, respectively. Importantly, the effect of target spatial resolution (scale) on DLST accuracy was less acute for the SEB approach than the RFR approach. The mean RMSE between the DLST obtained from SEB (RFR) and the the observed LST for built-up land, cropland, barren land, grassland, shrubland and forest were 11.51 (2.32), 1.21 (1.49), 1.48 (2.04), 1.43 (1.62), 1.51 (1.53), and 1.12 (1.25) K, in the warm season and 0.92 (1.43), 0.76 (0.91), 0.86 (1.20), 0.76 (1.08), 0.77 (1.20), and 0.61 (0.72) K, in the cold season, respectively. The results of this research, thus, demonstrate the high efficiency of the proposed, novel SEB-based approach for downscaling satellite LST at different scales and under different climatic, geographical and environmental conditions.
environmental sciences,imaging science & photographic technology,remote sensing