Feature Engineering to Early Detection of Plant Disease Using Image Processing and Artificial Intelligence: A Comparative Analysis

Abstract

Abstract. Plant diseases are a critical barrier to agricultural sustainability, contributing to annual crop losses of 30–40% in some regions. Such diseases arise from diverse pathogens, including fungi, bacteria, and viruses, and can severely degrade crop yield and quality if undetected. Early diagnosis is crucial for timely intervention, reduced pesticide use, and long-term soil health. Traditional approaches, which rely on manual visual inspection of leaves, remain slow, subjective, and impractical for monitoring large or remote farms. In recent years, the convergence of artificial intelligence (AI), computer vision, and image-based analysis has enabled automated disease detection systems that are both scalable and real-time. Classical techniques first employed color, texture, and shape descriptors combined with machine learning models such as Support Vector Machines (SVM), k-Nearest Neighbor (KNN), and Random Forests. These methods achieved moderate success but struggled with variability in lighting, background, and leaf orientation. The introduction of deep learning, particularly Convolutional Neural Networks (CNNs), transformed plant disease detection by allowing end-to-end feature learning directly from raw images. Modern lightweight architectures like MobileNet and EfficientNet further enhance deployability on mobile and edge devices. In parallel, segmentation-assisted and hybrid models improve robustness under complex field conditions. This review consolidates these developments, evaluating methods in terms of accuracy, generalization, computational efficiency, and field-readiness. It also identifies persistent challenges—such as limited annotated datasets and the opaque decision-making of deep models—and highlights future directions in explainable AI, multi-modal sensing, and IoT-integrated precision agriculture.

Code and data availability

This is a review/survey article with no original phenotyping measurements, datasets, images, code, or models of its own. All datasets and methods mentioned (e.g., PlantVillage, Taiwan Tomato) belong to cited prior work, and no author code or data availability statement appears.

No evidence-backed public reproduction asset is currently recorded.