Evaluating feature importance in AI-driven crop water stress models

Effective irrigation control is important to minimizing crop water stress, when there is climate variability and scarce water. The current paper suggests a predictable machine learning architecture to forecast crop water stress based on predicates of the weather condition, including temperature, humidity, soil moisture, rainfall, velocity, and NDVI. A semi-synthetic dataset was constructed to respond to a wide range of agro-climatic situations and SMOTE was used to deal with imbalance in the classes. Six classifiers such as Random Forest, Gradient Boosting, SVM, KNN, Decision Tree, and Logistic Regression were tested in a uniform pipeline with the highest accuracy of 83.39% being seen in Random Forest. SHapley Additive exPlans (SHAP) and permutation importance were applied to determine influential features to enhance transparency. The paper shows how an explainable AI can help to use data-driven, transparent, and scalable solutions to smart irrigation applications in precision agriculture.