Evolutionary Ablation: Quantifying Architectural Progress in Rice Disease Diagnosis across Machine Learning, Deep Learning, Computer Vision and Multimodal Explainable AI

Abstract

This paper presents a comprehensive longitudinal ablation study that systematically evaluates the architectural evolution of rice (Oryza sativa) disease diagnosis systems across eight consecutive research publications from 2023 to 2025. Our research trajectory has progressed through four distinct technological epochs: (1) Traditional Machine Learning employing handcrafted features with KNN and Decision Trees, (2) Deep Convolutional Neural Networks with comparative architecture analysis and edge optimization, (3) Advanced Paradigms including Vision Transformers, hyperspectral-temporal fusion, and attention mechanisms, and (4) Explainable AI Systems with integrated interpretability modules. We conduct a unified evaluation across a consolidated multimodal dataset of 15,230 images encompassing RGB field images, laboratory samples, and hyperspectral sequences across eight disease classes. The ablation reveals that while the transition from ML to deep learning yields the largest accuracy gain (+22.7%), the integration of attention mechanisms provides the optimal accuracy-efficiency trade-off (+11.3% accuracy, +28ms overhead). Vision Transformers demonstrate superior performance on globally distributed disease patterns (+4.8% over CNNs), while hyperspectral CNN-LSTM fusion enables unprecedented pre-symptomatic detection capability (88.5% accuracy at 48 hours before visual symptoms). Surprisingly, explainability modules incur only 2.4-8.1% computational overhead while increasing diagnostic confidence by 68.3% among agricultural experts. This study establishes the first quantified efficiency-performance frontier for agricultural vision systems and provides an architectural roadmap for future research in precision agriculture diagnostics.