Spatial-temporal modeling of soil moisture using optical and thermal remote sensing data and machine learning algorithms

Spatiotemporal estimation and monitoring of soil moisture based on remote sensing observations (optical and thermal) is challenging due to its physical nature in high vegetation conditions, necessitating improving and increasing the accuracy of soil moisture estimation in these areas. Therefore, this research aimed to develop a new approach to estimating surface soil moisture in agricultural fields with dense vegetation using machine learning algorithms by incorporating optical and thermal remote sensing data and soil physical properties. For this objective, 16 Landsat-8 satellite images and more than 430 control locations were used during the sugarcane crop’s growth period in 2018-2019 at the Hakim Farabi Sugarcane Agro-Industrial company in the Khuzestan province of Iran. A set of 10 scenarios of various unique combinations of the available input variables were developed and then evaluated by five machine learning algorithms, including multiple linear regression (MLR), decision tree-based algorithms (CART and M5P), and ensemble learning-based algorithms (gradient-boosted regression trees (GBRT) and random forest regression (RFR)). According to the results, the highest correlation between input variables and surface soil moisture was observed in Soil Wetness Index (SWI) and Normalized Soil Moisture Index (NSMI) with R values of 0.79 and 0.69, respectively. Also, the highest accuracy of machine learning algorithms based on R2, RMSE, and MAE results was obtained in GBRT (0.99, 0.011, and 0.006) and RFR (0.99, 0.014, and 0.007), respectively. In general, the findings of this research show the importance of using variables based on Landsat-8 remote sensing data in combination with ensemble learning algorithms that can be independent of any ground measurements.