Locating susceptible areas of karst springs in Kabgian watershed using Shannon maximum entropy algorithm

Introduction
Over the past decade, Maxent (maximum entropy) algorithm has been used extensively in natural resource studies, especially on topics related to animal habitat suitability, plant species distribution, and so in prediction of species presence potential areas. Since the presence of springs as a point feature at the watershed indicates the ability of the occurrence of a spring at that point, the Maxent model can be used to assess water resources potential of the watershed. Therefore, this study tries to evaluate the capability of Maxent algorithm to determine the potential areas for predicting and mapping presence of springs on the Kabgian karstic watershed.
Methodology
The data used in this study are divided to two categories of variables including dependent variable as points of presence of the spring, and independent variable as environmental information layers, including: Geographical Aspect, Lineament Density, Vertical Distance to the Channel Network (VDCN), Topographic Wetness ( TWI), Topographic Roughness (TRI), Topographic Position (TPI), Stream Density, Stream Power, Slope, Real Surface Area(RSA), Digital Elevation Model (DEM), Slope Curvature, Fault Density, Vegetation (NDVI), Vector Ruggedness Measure (VRM), precipitation and lithology. All modifications, layer preparation, classification, analysis, and map extraction was done using ArcGIS® 10.5, PCI Geomatica® 2018, SAGA GIS, MaxEnt® 3.3.3, Google Earth Pro 9, and Excel 2016 software.

Result and Discussion
The results showed that the accuracy of the model was very good with AUC 82.7 % on the ROC curve. According to Jackknife test, among 17 environmental variables, DEM, VDCN, rainfall, TPI and NDVI indices had the most impact on modeling, respectively. The least impact on modeling is VRM, geographical aspect, curvature, fault density, and lineament, respectively. The impact of the lithology index for predicting has been moderate; this can be attributed to the fact that more than 73% of the watershed is covered with limestone and gypsum which has a relatively uniform effect in modeling; This condition can also be seen in the lithology response curve which has given a significant response for Asmari and Gachsaran Formations. In general, the response curves of the most influential variables in the modeling were interpreted as follows: DEM response, first, increases in the range of 2200-2000 meters, then decreases to 2700 meters and again increases to 3700 meters. By moving away from the valley, the effect of the VDCN index is reversed; In other words, most springs have appeared near the valley. Areas with 1000 mm (and more) rainfall have been the most prone part of the basin. TPI index shows that with increasing basin surface bulge and slope, the areas prone to the presence of the spring decreases. The response of the NDVI index is positive and relatively uniform. The stream density response diagram has two significant breaks; it first decreases sharply at a density of 0.5 (km / km2) and increases sharply at a density of 1 to 1.3 (km / km2), then decreases again. The impact of watershed lithology in the area of Asmari and Gachsaran formations has been positive due to karstification and significant outcrop, and has had a negative response for other formations, especially Gurpi and Pabdeh formation due to being fine-grained. In general, the breaks in response curves are likely to be due to the interaction of other variables. The turning points of these curves and the range between these points, however, are the most effective part of the index that indicates whether the area is prone or not. According to the water resources suitability map, the areas with high, medium and low potential of the spring presence are 1894, 21795 and 63637 hectares, respectively. Areas with high potential are mainly located on the heights of Asmari and Gachsaran formations.
Conclusion
Between the 17 environmental indices used, none had acceptable predictive ability alone; this can be attributed to the complexity of karst systems and the difficulty of modeling it. The accuracy of the model was 82.7%, indicating the very good ability of the model in achieving the research goal. The final map of the areas prone to the presence of the spring is divided into three classes: high, medium and low, which include 2.2, 25 and 72.8 percent of the basin, respectively. Maxnet software, based on Shannon's entropy maximum, only needs the presence information of dependent variable and the information layers of the independent environmental variables to predict the most suitable regions for the presence of a dependent variable. Hence, the main advantage of this model is the accurate classification of potentials using layers of the presence of dependent variables; therefore, the possible error resulting from the absence or absence of points does not occur using this model. In overall, Shannon maximum entropy algorithm has a significant ability to determine the potential of groundwater resources in different areas and can be used for predicting and mapping karst water resources potential in watersheds.