A machine learning approach for predicting soil heat flux

Soil heat flux (SHF) is a crucial component of the energy balance Bowen ratio method, which can be used to estimate the surface layer optical turbulence structure parameter (Cn 2) in real-time. Using three years (2016-2018) of INCOMPASS data from three locations in India (Berambadi, Dharwad, and Kanpur), we developed a predictive model for SHF. After quality checking and scaling the data, we applied Lasso Regression to select the most important variables from a set of candidates. We identified soil temperature rate of change, soil temperature gradient, soil temperature at 0.05m, relative humidity, and net radiation as the key predictors. We then trained 3 different regression models: linear regression, random forests, and gradient boosted regression at each location. Ultimately, Gradient Boosted Regression proved to be the most effective model in predicting SHF. Our results suggest that the soil temperature gradient along with net radiation and soil temperature can be a reliable predictor of SHF across varied climates and spatial locations and demonstrates the potential for using machine learning models to estimate SHF in real-time.