A new nonlinear method for downscaling land surface temperature by integrating guided and Gaussian filtering

Abstract

Land surface temperature (LST), retrieved from thermal infrared (TIR) bands of remote sensing satellites, is an important parameter for various climate and environmental models. TIR bands detect a range of low-energy wavelengths, resulting in a coarser spatial resolution than other multispectral bands, and limiting applicability in heterogeneous urban regions. In this study, a new nonlinear method for LST downscaling, called Three Layers Composition (TLC), was proposed. The TLC integrates large-scale temperature variations, re-constructed detailed characteristics of LSTs, and strong boundary information. The performance of TLC is compared with disaggregation of radiometric surface temperature (DisTrad), thermal imagery sharpening (TsHARP), and random forest (RF) for a complex landscape in Beijing city, which has agriculture, forest, and massive impervious surfaces. The scale effects on the downscaled LSTs (DLST) were analyzed from the aspects of spatial resolution and spatial contexts. The experimental results indicate that the nonlinear algorithms (TLC and RF) perform better than linear methods (DisTrad and TsHARP). Indicated by coefficient of determination (R2), centered root-mean-square error (CRMSE), and correlation coefficient (CC), TLC (R2 = 0.901, CRMSE = 0.319, CC = 0.951) was the most effective and workable technique for predicting LSTs, followed by RF (0.768, 0.502, 0.874), TsHARP (0.544, 0.652, 0.734), and DisTrad (0.518, 0.751, 0.719). Larger experimental regions and larger ratios between initial and target resolution weaken the accuracy of DLST. TLC indicated a stronger ability to resist the influence of such scale effects. Traditional downscaling methods (DisTrad, TsHARP, and RF) are trained with global LST-predictor relationships and predict the DLST point by point, which can result in significantly biased estimates for very high or very low temperatures. Addressing this issue, TLC advantageously preserves the texture similarity between LST and its predictors, and yields more precise DLST, which showed higher consistency with the reference LST. Considering high accuracy and low computation time, TLC may be a safe technique for LST downscaling in other regions and different remote sensing sensors.