Evaluation of temporal and spatial changes of irrigation water quality classes in Qazvin Plain, Iran, using machine learning models

The present study aimed to determine the quality and quantity of groundwater resources for agricultural purposes in the Qazvin Plain (northwest of Iran) using the spatial and temporal distribution maps of agricultural water quality classes prepared by machine learning models during three study periods of spring 2012, 2016, and 2020. Modeling was performed based on geological maps, annual precipitation maps and 12 hydrogeochemical parameters measured for 63 piezometric wells. Appropriate hydrochemical parameters were selected for each statistical period to model agricultural water quality using the machine learning models of Random Forest (RF), Boosted Regression Tree (BRT) and Multinomial Logistic Regression (MnLR). The results introduced the best models to be RF in 2012 (kappa coefficient (κ)=0.54, overall accuracy (OA)= 69%) and MnLR in 2016 and 2020 (κ= 0.83 and 0.75; OA=88 and 84%), respectively. The percentage of area for C4-S3 class (very high salinity with high sodium) increased from 5% in 2011 to 23.9% in 2019. Giving the increased precipitation in 2019, the agricultural water quality class in the southern region changed from C4-S3 in 2015 to C4-S2 (very high salinity with medium sodium) in 2019. Additionally, the simulated maps showed an elevation in the percentage of C4-S3 class area from 2012 to 2020 in the central part of the region where agricultural lands are concentrated. Our findings revealed the trend of adverse changes in water quality at different regions of Qazvin Plain during the years of study, highlighting the need to make purposeful management decisions. And The study utilized both advanced machine learning algorithms and traditional classification methods, including the Wilcox diagram, to assess agricultural water quality based on twelve physicochemical parameters. And The 12 parameters used in the study were selected based on data availability and relevance to agricultural water quality standards. Due to inconsistent data across years, variables such as nitrate and organic matter were excluded.