Application of nonlinear models and groundwater index to predict desertification case study: Sharifabad watershed

The level of groundwater can also be used in monitoring desertification and land degradation. In this study, three models, namely: partial least square regression, artificial neural networks (ANN), and adaptive neuro-fuzzy inference system, were used to monitor and predict the level of groundwater and the land degradation index via the Iranian Model of Desertification Potential Assessment method. The groundwater data of 24 Piezometric wells from 2002 to 2016 were also collated to predict the groundwater level. In all models, 70% of the data were applied for training, while 30% of data were employed for testing and validation. Monthly rainfall, topographic wetness index, distance of the river (m), latitude and longitude of Piezometers in the Universal Transverse Mercator coordinate system were the inputs, and the level of groundwater was the output of each method. The prediction performance of both training and testing sets is evaluated by R2 and MSE. Looking at statistical inferences, we found that ANN has the highest efficiency (R2 = 0.96, MSE = 0.71 m) which agree with other findings. We combined the results of ANN with ordinary kriging (OK) and produced a groundwater condition map. According to the potential desertification map and groundwater level index, the potential of desertification had become severe since 2002 and was at a rate of 60% of land area, which, due to incorrect land management in 2016, increased to almost 98% of the land surface in the study area. Again between 2002 and 2016, the land area with low degradation risk decreased from 38,030 ha (39% of the study area) to zero ha in 2016. In 2016, there was no moderate land degradation risk. Using ANN, we predicted that around 99% of the area (95,206 ha) was severely degraded in 2017 and according to groundwater level index, the land degradation increased by 100%. This implies that the area deserves urgent care and reclamation. We also used latitude and longitude of Piezometers as input variables which improved the model. In addition to the target variable, latitude and longitude play important roles in OK and decreased the total error of two combined models.