Estimation of the deep drainage for irrigated cropland based on satellite observations and deep neural networks

Abstract

The deep drainage of irrigated cropland, which is one of the main causes of water losses in irrigation, plays an important role in water resource management. However, its magnitudes and variations remain unclear at regional scales. In this study, we established a remote sensing deep drainage model for irrigated cropland (ICDD_RS model) based on satellite observations and a deep neural network (DNN). First, deep drainage for natural vegetation was calculated by the water balance equation. Then, the ICDD_RS model was established by linking the calculated deep drainage and remote sensing data for natural vegetation based on a DNN. Finally, the model was used to predict the deep drainage of irrigated cropland through remote sensing data. The model was tested in the Heihe River Basin of China and the results showed the following: (1) The ICDD_RS model can well describe the relationship between deep drainage and selected influencing factors (0.83 < R2 < 0.86). At the site scale, it is highly consistent with in-situ observations with R2 values of 0.86, 0.74, and 0.71 for the Daman, Yingke, and Pingchuan stations, respectively. At the catchment scale, a low bias of 12% is obtained. (2) Deep drainage is large, accounting for 6–17% of the irrigation water. This model should be valuable in providing more detailed and effective information on agricultural water resource management.