Application of convolutional neural network in predicting groundwater potential using remote sensing: a case study in southeastern Liaoning, China

With the rise of machine learning and artificial intelligence, back propagation (BP) neural network, support vector machine (SVM), random forest model, and others can be used to predict the distribution of groundwater. By using the existing sample data, learning, training, and forecasting for some unknown areas (unmanned areas, areas where people are not easy to reach, etc.) can save costs and improve the efficiency of machine learning. This paper took an area of 2000 km 2 in the southeast of Liaoning Province as the study area. This study used convolutional neural network (CNN) for data training and testing based on the results of groundwater assessment by remote sensing with lithology index, relief index, slope index, water density index, vegetation fraction index, soil humidity index, and land temperature index and field survey data and data of wells. With the coupling relationship between the results of groundwater potential assessment–based AHP and groundwater spatial distribution, the prediction model of groundwater distribution is established. Finally, after 1000 times of training, a good prediction model with the training set of 100% accurate and the test set of about 80% accurate was obtained. Subsequently, a ROC curve was done by using the survey data of the study and the results of prediction (one or zero) of the CNN model. The ROC curve showed that the AUC was 0.854, and the standard error was 0.08. Thus, the groundwater situation in the unsurveyed areas can be predicted by this model to guide the development and utilization of groundwater in the future.