FEATURE ENGINEERING AND EXPLAINABLE AI FOR PLANT DISEASE DETECTION: A COMPREHENSIVE SURVEY

Abstract

Plant diseases represent a significant threat to global food security, causing yield losses estimated between 20% and 40% annually. Accurate, timely, and scalable disease detection is therefore critical for modern agriculture. This survey provides a comprehensive review of feature engineering approaches and deep learning techniques applied to plant disease detection, with special focus on Gradient-weighted Class Activation Mapping (GradCAM) and Explainable Artificial Intelligence (XAI) methods that bridge the gap between black-box model predictions and human interpretability. We examine traditional handcrafted feature engineering methods including color histograms, texture descriptors (GLCM, LBP, Gabor), and shape features, followed by deep feature extraction using Convolutional Neural Networks (CNNs) and transfer learning architectures such as ResNet, EfficientNet, VGG, DenseNet, and MobileNet. The survey covers key datasets including PlantVillage and PlantDoc, discusses data augmentation strategies, and critically evaluates explainability tools including GradCAM, GradCAM++, LIME, and SHAP. Emerging trends such as MLOps integration, lightweight models for edge deployment, and transformer-based architectures are also discussed. The paper highlights challenges including limited real-world data, class imbalance, and the need for domain-trustworthy interpretability, and outlines future research directions toward robust, explainable, and deployable plant disease detection systems.