资源环境科技发展态势分析平台

Studies focusing on national/global water scarcity require water availability and water use to quantify the imbalance. In this regard, annual irrigation withdrawal data reported by the USGS every 5 years provide a valuable information on the water use patterns over the United States. This study develops an empirical model to estimate annual irrigation water withdrawal using irrigated area, climate information, and population as predictors. Given the hierarchy in the data sources, we propose a predictive linear hierarchical regression model to develop annual irrigation water withdrawal models using varying intercepts (VI) and varying intercepts and slopes (VIS) approaches. Estimates from hierarchical models are compared with pooled and unpooled classical regression models. Overall, both hierarchical models outperform the classical models with the adjusted R-2 between USGS-reported and modeled withdrawal estimates being above 0.6 in most states using county and climate division level data. However, due to the spatial difference between the supply (rural areas) and demand (urban areas) for agriculture products, climate division level estimates exhibit a higher adjusted R-2 than county level estimates. The VIS model is able to capture local effects better, particularly for states whose irrigation withdrawal patterns significantly differ from the national pattern. The performance of the models is also validated by leaving out the entire nation's water-use data out (i.e., leave-one-out cross-validation) to ensure the reported skill is not due to overfitting. Split-sample validation in predicting 2010 irrigation withdrawal also shows the potential of the developed hierarchical model in estimating the annual irrigation withdrawals for the years with no data within the once in 5 year USGS database.