AI-Driven Precision Agriculture for Sustainable Crop Production

Abstract

As a key enabling technology for precision agriculture, Artificial Intelligence (AI) promises the insight to turn the growing trove of available satellite, sensor, and imagery data into field-specific decisions – from crop yield forecasts to plant disease detection to input-resource optimisation. This paper reviews new (2021–2025) peer-reviewed literature on AI-based precision agriculture and presents a novel ensemble machine-learning model – an integrated Random-Forest/XGBoost stacked yield-prediction model with a fine-tuned convolutional-neuralnetwork disease-detection model, tested for both prediction performance and downstream resource-use and productivity impacts. Vegetative-stage NDVI, cumulative growing-season rainfall, and soil nitrogen status were the most important variables for predicting yield, and the proposed yield-prediction ensemble outperformed the baselines (individual RF, XGBoost, support-vector regression, gradient boosting, decision tree, and linear regression), with R² = 0.95 on hold-out data. The fine-tuned convolutional disease-classification model achieved 99.1% accuracy, between the 99.05% and 99.75% reported by state-of-the-art architectures in standard plant-disease benchmarks. Results from the field and deployment studies reviewed indicated that AI precision-agriculture adoption had the following range of inputuse reductions and productivity gains: 22-31% on variable-rate fertilizer (VRF) application, targeted pesticide application, and precision irrigation, while the biggest gains in decreased usage of farm inputs occurred in relation to autonomous machinery fuel and labor. The results show that ensemble machine-learning techniques can provide statistically significant and repeatable advances in prediction accuracy and the viability of on-farm practices in combination with multi-source remote-sensing and soil data. The key challenges to more widespread adoption are discussed, including data quality and availability, model interpretability, computational or connectivity needs, and equitable access to the smallholder farmers.