Real-Time Irrigation Need Assessment Using Adaptive Machine Learning with Explainable AI for Sustainable Agriculture

Abstract

Efficient water management has become a critical challenge in modern agriculture due to increasing climate variability, water scarcity, and the growing demand for sustainable food production. In addition, oil spills and petroleum contamination in agricultural water sources can adversely affect irrigation practices by degrading water quality, altering soil properties, and potentially reducing the suitability of water for crop irrigation. Such contamination can influence irrigation requirements and complicate real-time irrigation decision-making, particularly in environments affected by environmental pollution. This study proposes an adaptive real-time irrigation need assessment framework that integrates Machine Learning (ML), Continual Learning (CL), and Explainable Artificial Intelligence (XAI) to support intelligent irrigation decision-making. The framework employs a comprehensive data preprocessing pipeline, including missing value imputation, categorical feature encoding, and feature normalization, followed by a Random Forest classifier as the baseline predictive model. To address the non-stationary nature of agricultural environments, a CL strategy is introduced to incrementally process streaming sensor data in sequential batches, enabling the model to continuously adapt to changing environmental conditions while minimizing catastrophic forgetting. Model interpretability is enhanced through multiple XAI techniques, including feature importance analysis, permutation importance, partial dependence plots, and local explanation methods, providing transparent insights into the environmental variables that influence irrigation recommendations. Experimental evaluation using five-fold cross-validation demonstrated robust predictive performance, achieving fold accuracies of 95.50%, 96.33%, 95.67%, 97.33%, and 95.50%, with a mean accuracy of 96.07±1.41%. Furthermore, CL experiments across ten sequential data batches preserved the model performance without degradation, maintaining a final accuracy of 98.40%, thereby exhibiting zero performance loss during incremental adaptation. These results confirm that the proposed framework effectively combines predictive accuracy, adaptability, and interpretability, making it a robust and scalable solution for precision irrigation and sustainable agricultural water resource management.