County-Level Irrigation Water Demand Estimation Using Machine Learning: Case Study of California

Irrigated agriculture is the largest consumer of freshwater globally. Despite the clarity of influential factors and deriving forces, estimation of the volumetric irrigation demand using biophysical models is prohibitively difficult. Data-driven models have proven their ability to predict geophysical and hydrological phenomena with only a handful of influential input variables; however, the lack of reliable input data in most agricultural regions of the world hinders the effectiveness of these approaches. Attempting to estimate the irrigation water demand, we first analyze the correlation of potential influencing variables with irrigation water. We develop machine learning models to predict California’s annual, county-level irrigation water demand based on the statistical analysis findings over an 18-year time span. Input variables are different combinations of deriving meteorological forces, geographical characteristics, cropped area, and crop category. After testing various regression machine learning approaches, the result shows that Gaussian process regression produces the best results. Our findings suggest that irrigated cropped area, air temperature, and vapor pressure deficit are the most significant variables in predicting irrigation water demand. This research also shows that Gaussian process regression can predict irrigation water demand with high accuracy (R2 higher than 0.97 and RMSE as low as 0.06 km3) with different input variable combinations. An accurate estimation of irrigation water use of various crop categories and areas can assist decision-making processes and improve water management strategies. The proposed model can help water policy makers evaluate climatological and agricultural scenarios and hence be used as a decision support tool for agricultural water management at a regional scale.