Spatial prediction of loose aquifer water abundance mapping based on a hybrid statistical learning approach

In order to study and prevent water hazards in coal mines under thick loose strata, we need to make correct assessments of their water abundance levels based on the limited borehole data. According to the multi-source information composite principle, five main influencing factors of water abundance are chosen, and they are: the aquifer thickness, the thickness ratio between sandy and clayey layers, the consumption of the drilling fluid, the core recovery rate, and the hydraulic conductivity. Their spatial variations in the whole study area could be inferred from Kriging interpolation. Next, we have developed a novel off-the-shelf two-step assessment approach. In the first step, we apply a dimensionality reduction technique called Fisher discriminant analysis (FDA) to project the original normalized data into a low-dimensional space, which is convenient for data visualization. In this projection process, we want to keep the original information as much as possible. In the second step, we train three classification algorithms on the same transformed low-dimensional data to predict the water abundance level, and leave-one-out cross-validation is used to validate our proposed method due to data sparsity. Finally, the comprehensive zoning map of the study area’s water abundance level is established, which can provide scientific guidance for the mining operations and prevention of mine water hazards in this region. The whole process is further elaborated through a case study of the Baodian coal mine, from which we are able to know the location with the highest water abundance level.