A New Approach Based on Deep Neural Networks and Multisource Geospatial Data for Spatial Prediction of Groundwater Spring Potential

Abstract

Groundwater spring plays a crucial role in human life, including water resource management and planning; therefore, developing accurate prediction models for groundwater spring potential mapping is essential. The objective of this research is to introduce and confirm a new modeling approach based on TensorFlow Deep Neural Networks (TF-DNN) and multisource geospatial data for spatial prediction of groundwater spring potential, with a case study in the tropical province in the central highland of Vietnam. For this task, the TF-DNN model structure with three hidden layers with 32 neurons each was established; therein, the Adaptive Moment Estimation (ADAM) algorithm was used as an optimizer, whereas the Rectified Linear Unit (ReLU) was used as the activation function, and the sigmoid was used as the transfer function. A geospatial database for the study area, consisting of 733 groundwater spring locations and 12 influencing factors, was prepared in ArcGIS Pro. Then, it was used to develop and verify the TF-DNN model. Decision Tree, Support Vector Machine, Logistic Regression, Random Forest, and Classification and Regression Trees were used as a benchmark for the model comparison. The results demonstrate that the proposed TF-DNN model (Accuracy = 80.5%, F-score = 0.797, and AUC = 0.864) achieves a high global prediction performance, outperforming the benchmark models. Thus, the TF-DNN represents a novel and effective tool for spatially predicting groundwater spring potential mapping. The groundwater spring potential map generated in this study has the potential to assist provincial authorities in formulating strategies concerning water management and socio-economic development.