Multi-Station Nitrate Prediction via Artificial Intelligence and Soft Computing Tools

The diffusion of nitrate pollution in watersheds is due to complex biochemical and hydrological procedures linked to the cycle of nitrogen and water. Nitrate load comes from different sources such as wastewater treatment plants, runoff of fertilized lawns and cropland, failing on-site septic systems, runoff of animal manure storage areas, and industrial discharges that contain corrosion inhibitors. Loss of nitrate to surface and groundwater can reduce farm productivity, harm the environment, and affect drinking water quality. Large uncertainties and limited physical understanding of the water quality such as nitrate barricade the process-based modeling and seek a black box relationship between driving and resultant variables. Therefore, in this paper Multi-Station (MS) modeling of nitrate of the Little River Watershed (LRW) has been done. Hence, MS nitrate modeling is considered whereby nitrate loads of the inside and outlet of the LRW could be predicted. As a more explanation, the nitrate of upper sub-basins are employed for predictions of the interior sub-basins nitrate loads, and then, central sub-basins are participated in outlet nitrate prediction of the LRW. So, MS model can prepare a reliable platform to get information about the amount of nitrate in crucial places of the LRW. For this purpose, two scenarios with distinct views are used for MS nitrate modeling to identify the suitable strategy for future hydro-environmental researches. In the first scenario, Markovian characteristics of the streamflow-nitrate process are proposed as the base of the MS model, where antecedent of streamflow and nitrate time series of sub-basins are shared in nitrate modeling. On the other hand, non-linear feature extraction criterion of MI that is more suitable measure regarding the linear measure of Correlation Coefficient (CC) is employed for the selection of appropriate inputs of the Least Square SVM (LSSVM) and Feed Forward Neural Network (FFNN) models to avoid from the time consuming trial-error process of input selection. In the second scenario, seasonality-based characteristics of the streamflow-nitrate process are focused. Where, streamflow and nitrate time series of the sub-basins are decomposed by the wavelet transforms at a suitable level for clarifying spectral and temporal information of the time series. Then, as a new feature extraction method, both SOM and MI are respectively employed for clustering homogeneous sub-series and selecting cluster's proper agents, to be fed into LSSVM and FFNN models for MS nitrate load modeling of the LRW.