Artificial intelligence and remote sensing for spatial prediction of daily air temperature: case study of Souss watershed of Morocco

Abstract

Air temperature (Tair) is a fundamental variable in climate research and climate impact management. Conventional field observations do not accurately capture its spatial distribution due to the sparse and uneven distribution of weather stations, especially in remote areas where the local variability is high. To circumvent this problem, in this study, remote sensing and weather station data were used to estimate Tair in the Souss watershed in Morocco. Two statistical methods, including linear regression and partial least squares (PLS), and four machine learning algorithms, namely k-nearest neighbors, random forest (RF), extreme gradient boost, and Cubist, were used for modeling and predicting Tair and its performance were evaluated using random subsets and cross-validation. Moderate resolution imaging spectroradiometer predictors, including Terra band 32 emissivity, Terra nighttime land surface temperature, Terra local time of night observation, Aqua band 31 emissivity, Aqua daytime land surface temperature, and Aqua nighttime land surface temperature (ALSTN), and auxiliary inputs, including sky-view, elevation, slope, and hillshade, were used as inputs for modeling. The results showed that the Cubist and RF were the most accurate models (RMSE = 2.09°C and 2.13°C, R 2 = 0.91 and 0.90, respectively), while PLS had the lowest predictive power (RMSE = 2.71°C; R 2 = 0.83). The overall performance of the models for estimating Tair in the study area was generally satisfactory, with RMSE limited to less than 3°C for all models. Nevertheless, the station data reliability was still an issue, with only four of the seven stations marked by complete meteorological data.