UC Berkeley

Irrigation is the most significant consumer of freshwater worldwide. Deciding on the right amount of irrigation is crucial for sustainable water management and food production. The Penman-Monteith (P-M) reference crop evapotranspiration (ETO) is the gold standard in irrigation management and scheduling; however, its calculation requires measurements from multiple sensors over an extended reference grass surface. The cost of land, sensors, maintenance, and water to keep the grass surface green impedes having a dense network of ETO stations. To solve this challenge, this research aims to develop an input-limited ETO estimation approach based on historical weather data and machine learning (ML) algorithms to relax the need for a reference grass surface. This approach, called “SolarET,” takes solar radiation (RS) data as its sole input. RS is the only meteorological driving factor of ETO that does not rely on the measuring surface. To test the generalizability of SolarET, we test its performance over unseen arbitrary locations across California. California is chosen as the case study since it is one of the world's most hydrologically altered and agriculturally productive regions. In total, 19,088,736 hourly data samples from 131 automated weather stations have been used in this study. The ML models have been trained over 114 stations and tested over 17 unseen stations, each representing a California climatic zone. Our findings point to the superiority of decision tree-based algorithms versus neural networks. SolarET works best in irrigation-oriented regions of California (e.g., Central Valley) and is less accurate in coastal and desert zones. Our results demonstrate the higher accuracy of SolarET using hourly (RMSE = 0.93 mm/day) and daily (RMSE = 0.97 mm/day) RS data in comparison to well-known empirical alternatives like Priestley-Taylor (PT) (RMSE = 1.35 mm/day) and Hargreaves-Samani (HS) (RMSE = 2.69 mm/day).