An AI-Driven Hybrid Image-Based Framework Using IoT Leaf Sensors for Sustainable Agriculture

Abstract

This study proposes an AI-driven hybrid image-based decision-support framework integrated with IoT leaf sensors for sustainable diagnosis and management of crop nutrient deficiencies and diseases. Unlike conventional machine-learning pipelines that primarily focus on classification accuracy, the main contribution of this work lies in the explicit integration of probabilistic classification outputs into a mathematical optimization model for fertilizer and pesticide management. Leaf images are processed using multi-feature extraction methods—Local Binary Pattern (LBP), Local Directional Pattern (LDP), Block-Wise Truncation (BWT), and Hough Transform—to capture complementary texture and geometric characteristics. Class imbalance is addressed using the Bayesian Gaussian Mixture Model–Synthetic Minority Oversampling Technique (BGMM-SMOTE), while feature redundancy is reduced through the Flower Pollination Algorithm (FPA). A Random Forest classifier is then employed to estimate class probabilities rather than merely discrete labels. The key novelty of the proposed framework is the transformation of these probabilistic outputs into actionable agronomic decisions through a mathematically formulated optimization model grounded in Diagnosis and Recommendation Integrated System (DRIS) principles and disease-treatment constraints. This model minimizes fertilizer and pesticide use while maintaining nutrient balance and effective disease control, thereby closing the gap between diagnosis and intervention. Experiments on two public datasets—oil palm nutrient deficiency and plant disease—were validated using real-world IoT sensor data, achieving accuracies of 89.6% and 93.8%, respectively. More importantly, the optimization-driven recommendations achieved substantial reductions in agricultural inputs, including 2,445.6 kg of potassium fertilizer, 330.25 kg of boron fertilizer, 14.2 liters of insecticide, and 5.0 liters of fungicide. These results confirm that the principal contribution of this work is not incremental feature or classifier combination, but a mathematically grounded decision framework that translates AI-based diagnosis into measurable sustainability gains in precision agriculture.