Estimation of groundwater level variation is important for establishing a sustainable development plan of groundwater resources. Thus, it is necessary to develop a method for accurately estimating groundwater level variation. In this study, we developed the machine learning model to estimate variations in the groundwater level, and applied it to Chungcheong province, South Korea using geological and hydrological factors.
First, we clustered 58 groundwater observation wells using eight geological factors (soil type, soil texture, drainage level, effective soil depth, land cover, digital elevation model, slope, and timber density). Based on the clustering results, we developed machine learning models (Random Forest (RF) and Support Vector Machine (SVM)). Four hydrological factors (precipitation, temperature, normalized difference vegetation index, and soil moisture) in addition to eight geological factors were used for classifying the groundwater level variations. The accuracy of the machine learning models was evaluated based on the Area Under the Curve (AUC) value. RF showed better performance with an AUC average value of 0.75 for each cluster, compared to the SVM of 0.73. Maps of class in the groundwater level variation were created using the result of the machine learning models (RF and SVM) and in situ data. The maps of RF and SVM represented more detailed and accurate variations than the interpolation map based on in situ groundwater observation.
The groundwater level variation map can be a useful tool for efficient groundwater management. The machine learning models developed in this study can be particularly unique in the areas where groundwater observation wells are scarce or do not exist. Furthermore, it will be possible to predict the groundwater level variations in the future through the improved machine learning models when geo-hydrological data become available.