Assessing the performance of individual and ensembled models in identifying areas with infiltration potential

Estimating groundwater recharge using other affecting factors such as hydro-meteorological, and physical factors, is the main way to understanding, predicting the sustainability and availability potential of aquifers. The objective of this research was investigating the efficiency of some individual and ensembled models and the effect of the ensembling on promoting the efficiency of Bayesian and Random forest models.

The required environmental layers (DEM, aspect, slope, curvatures (profile and plan curvature), lithology, landuse, soil texture, NDVI, fault distance, river distance, SPI, drainage density, TWI), were prepared by ArcGIS 10.6 software for the study area. the Bayesian theory and Random Forest models and ensembling of these models were avaluated. To consider the ensemble effect of these models, input layers were used in two models and then the models were ensembled by several scenarios, which were based on the principles of basic mathematics. The double-ring infiltrometer method were used for understanding the spatial variability of groundwater recharge potential (GWRP). The performance of the models was evaluated using statistical measures. ROC, CCI, and TSS indices were used for evaluating the results of implemented models. Finally, GWRP mapping prepared and the percolation of the study area was classified into five classes: very high infiltration, high infiltration, medium infiltration, low infiltration and very low infiltration.

The sandy-clay-loom soil texture and the Quaternary sediments (Qft2), rangeland and agriculture areas (in landuse layer), showed maximum percolation. The results indicated that the random forest (RF) model was identified as the superior model compared to Bayesian and ensembled models, by ROC, TSS and CCI indices (ROC= 0.983, TSS= 0.86, and CCI= 93.9, respectively). Also, among the ensembled models, the RFBa5 model (based on the fifth scenario) was evaluated as the most efficient model through ROC, TSS, and CCI indices (ROC= 0.984, TSS= 0.76, and CCI= 87.94, respectively). Based on the first individual model (RF), 11% of the study area had moderate to very high infiltration potential. This despite the fact that the 89% of the study watershed was found with low and very low potential. While according to RFBa5, 30% of the study area were estimated with moderate to very high potential and 70% with low and very low potential. The Bayesian model was observed as the weakest model. But based on the ensemble of this model with the random forest model, under different scenarios, the strengthening of this model was observed. These results show the positive effect of ensembling the models.

The GWR potential maps are useful in planning with more accuracy for implementation of artificial GWR, soil protection, aquifers and watershed management projects in order to protect water and soil resources by directing runoffs to preventing the soil erosion and aquifers recharge. It is recommended to study different suitable models and their ensembling in this field or other watersheds and select the best model to get the best performance and obtain an accurate map.