Prediction of underground water potential in Khorramabad plain based on convolutional neural networks

Today, due to population increase, industrial development, excessive exploitation, droughts, exploitation of underground water has multiplied. Therefore, identifying areas with underground water as one of the important sources for providing drinking water, agriculture, and various industries is considered to be one of the important and necessary issues in water resources management. The purpose of this research is to investigate and zonate the areas with underground water in Khorram Abad plain located in Lorestan province using convolutional neural network method. For this purpose, maps of nine factors affecting underground water were first prepared in the ArcGist environment. In the convolution method, the number of samples was determined as the ratio between the training set and the test set was 70:30, and the convolution neural network framework was used as 2 convolution layers and 2 integration layers, 2 complete connections. layers and finally the sigmoid layer was used for classification from the 3-3 convolution kernel, the Relu function as the activation function and the cross entropy function as the loss function. The obtained maps were classified into 5 classes: very good, good, average, low and very low. Confusion matrix was also used to validate the results of the model. 30% of the real data was used for evaluation, which resulted in an overall accuracy of 92%, that is, the model was able to correctly identify 92% of the data as underground water and 93% as the absence of underground water. The analysis of the groundwater potential map of the convolutional neural network model shows that about 57% of the area is in low groundwater conditions and 43% of the area is in good groundwater conditions.