Spatial prediction of groundwater potential and driving factor analysis based on deep learning and geographical detector in an arid endorheic basin. (September 2022)

• Record Type: Journal Article
• Authors: Wang, Zitao; Wang, Jianping; Han, Jinjun
• Is Part Of: Ecological indicators. Volume 142(2022)
• Publication Date: 2022-09
• Subjects: Groundwater potential mapping (GPM) -- Driving factors -- Geographical detector -- Deep learning -- Qaidam Basin; Environmental monitoring -- Periodicals; Environmental management -- Periodicals; Environmental impact analysis -- Periodicals; Environmental risk assessment -- Periodicals; Sustainable development -- Periodicals; 333.71405
• DOI: 10.1016/j.ecolind.2022.109256
• Languages: English
• ISSNs: 1470-160X

Abstract

Highlights: We used a Random Forest, a deep neural network, and a convolutional neural network (CNN) to predicte the groundwater potential based on 17 factors in the Qaidam Basin. The CNN model had the best performance in three algorithms, so we used geographic detector method to quantify driving factors of the groundwater potential predicted by CNN. Geomorphy had the strongest impact on desertification, followed by evaporation. Comprehensive consideration of multiple factors enhanced study of groundwater potential. Abstract: Substantial mineral resources are enriched in the arid endorheic basins; however, due to environmental constraints, these areas face water shortages as well as its extremely uneven spatiotemporal distribution, which restricts the development of local industry and agriculture. Identifying these areas of the high groundwater potential are useful for groundwater supply and its sustainable planning. In this study, the Qaidam Basin in Northwest China was taken as an example. We collected 17 conditioning factors (i.e., precipitation, evaporation, geology, soil, Topographic Wetness Index, Fractional Vegetation Cover, distance to rivers, river density, distance to roads, road density, distance to faults, fault density, slope, curvature, residential density, landcover, and geomorphology) affecting groundwater resources in arid areas. We also collected 139 groundwater samples and used random forest (RF), deep neural network (DNN) and convolutional neural network (CNN) (associated with one-hot encoding) to predict the groundwater potential in this area. The Qaidam Basin was discretized into 420, 000 sample points calculated in turn by the above three models. Receiver operating characteristic (ROC) and area under the curve (AUC) were used to test the accuracy of the three methods. Results indicated that the prediction scores for the three methods were 0.742, 0.790, and 0.817, and the AUC was 0.783, 0.811, and 0.846, respectively. The result provided by CNN was more precise than the results provided by RF and DNN. Additionally, this study aims to investigate the effects of the aforementioned factors on groundwater potential. A total of 17 factors were combined with the Geodetector model to quantify their impacts and interactions on the groundwater potential of the Qaidam Basin. Results revealed that the critical factors affecting groundwater potential in the Qaidam Basin were geomorphology (0.183) and evaporation (0.144), and their combined contribution was 0.457. The influence of arbitrary two-factors on groundwater potential is larger than that of themselves, demonstrating linear or nonlinear enhancement between them and confirming that the factor selections were sensible. The method based on CNN-Geodetector provides a novel approach for calculating groundwater potential, selecting appropriate evaluation indicators and quantifying the driving factors in the arid endorheic basins.