Journal of Epidemiology and Public Health
ISSN: 3065-9078 | DOI: 10.64030/3065-9078
Mirza Samiulla Beg
The scarcity of water is one of the primary risks to the sustainability of agriculture in India owing to the heterogeneous nature of climatic and soil characteristics. In our study, we built a prediction model using geospatial data, climatic factors such as precipitation and temperature, and agricultural factors like cropping intensity and irrigation, to predict whether the region would fall under Low, Medium, or High categories of the Water Stress Index using the Agro_India_States dataset. Categorical variables such as soil type and climatic zones were encoded first before deploying seven machine learning models: Gradient Boosting, Random Forest, Decision Tree, Neural Network, and three versions of Logistic Regression. Selection of the best model was based on KS statistic (also known as Youden index), accuracy, and ROC area. The Gradient Boosting model performed the best in terms of KS score at 0.50 and outperformed the rest of the models significantly; traditional Logistic Regression and the Neural Network did not converge well and scored a KS statistic of zero. The results of the feature importance analysis showed that precipitation is one of the most crucial determinants for water stress, with other features such as soil, climate, cropping intensity, and geographical coordinates (latitude and longitude) having substantial significance as well. The winning model generated a total lift of 4.44 in the top 10 percent regions, indicating its efficiency in recognizing highly stressed regions. This study thus shows that gradient boosting, a member of ensemble machine learning techniques, significantly performs better than traditional approaches when used to forecast water stress for agriculture in India due to its ability to deal with complex relationships between categorical and numerical predictors.