High resolution annual irrigation water use maps in China based-on input variables selection and convolutional neural networks

Abstract

Water use and available water resources are the prerequisites for optimal water resources allocation, which is an efficient way to alleviate water scarcity all over the world. However, the spatial scales of water use and available water resources are often not matched due to their different ways for data surveying or observing. The spatial distributed (grid-scale) water use map is one of the ways to convert the spatial scales of the water use and available water resources. As irrigation is essential to guarantee sustainable development of agriculture, food security and economy, and accounts for most water use in China. A framework to generate high resolution (1 km) annual irrigation water use maps has been proposed. An iterative input variables selection (IIS) algorithm is adopted to select the optimal input variables among candidate input variables. After determining the optimal input variables, the convolutional neural network (CNN) is built to establish the relationship between the selected input variables and the irrigation water use. The spatial distributed irrigation water use maps can be produced by the trained CNN with the selected input variables at grid scale. The spatial distributed annual irrigation water use maps in China is produced at 1 km resolution from 1998 to 2013, and the performance of the framework has been proven to be credible with three metrics (i.e., root mean squared error, Nash Sutcliffe efficiency coefficient and relative error). Based on the spatial distributed annual irrigation water use maps, the spatial autocorrelation of annual irrigation water use in China is weak and the clustered pattern is random with 99% confidence. The temporal changing pattern mainly depends on precipitation and soil moisture in most prefectures, and the change of the cultivated area can also lead to dramatic change of the annual irrigation water use in China. The proposed framework can generate high resolution spatial distributed annual irrigation water use maps at grid scale, and the maps can be applied to ensure the optimal water resources allocation to realize efficient utilization of water resources and the agricultural policy-making.