From Rule-Based Systems to Retrieval-Augmented LLMs: A Comparative Review of AI Approaches for Maize Pest and Disease Management in Punjab

Abstract

Spring-sown maize in Punjab faces three recurring problems every season: shoot fly (Atherigona spp.) at the seedling stage, Maydis leaf blight (Bipolaris maydis) through the vegetative phase, and banded leaf and sheath blight (Rhizoctonia solani f. sp. sasakii) once the canopy closes. The Punjab Agricultural University (PAU) and the ICAR-Indian Institute of Maize Research (ICAR-IIMR), both based in Ludhiana, publish advisories for all three, and farmers in principle have access to this guidance every season. What has changed over the last four decades is not the advisories themselves but the computational tools built around them, moving from hand-coded rule-based expert systems in the 1980s and 1990s to classical machine learning on hand-crafted image features in the 2000s to convolutional neural networks and vision transformers from roughly 2015 onwards and now to retrieval-augmented large language models. Each of these four approaches has its own literature, and they rarely talk to each other: the deep-learning papers cite almost none of the expert-systems work, and the handful of agricultural RAG papers that exist barely reference either. The deep-learning literature also has a specific blind spot worth naming up front, which is that it leans heavily on the PlantVillage benchmark, a dataset of laboratory-photographed leaves that generalizes poorly once a model is pointed at an actual field and one that under-represents spring maize and the Punjab-specific disease complex in any case. This paper traces all four eras for one crop and one region rather than treating them as separate literatures, and it tries to hold itself to a stricter standard than is typical of a survey: every technical or agronomic claim is checked against at least two independent sources before it is stated. Where the practical, farmer-facing metrics that would actually distinguish these four approaches, hallucination rate, response latency on a cheap phone, and cost of running a village deployment simply have not been measured for Punjab maize by anyone, we say so and propose a protocol for measuring them rather than inventing numbers. The paper closes by naming, concretely, the gap in Punjab-specific field imagery and loss data that currently limits how far deep learning and RAG can be trusted here.