An AI-IoT Convergence Framework for Autonomous Weed Detection and Pest Suppression in Sustainable Precision Farming

Abstract

Weeds and pests are estimated to cause 30-40% of global agricultural production losses, while conventional control relies heavily on manual labour, mechanical intervention and chemical pesticides and herbicides - each carrying scalability, cost or environmental drawbacks. This paper proposes an integrated AI-powered IoT architecture for precision agriculture that couples deep-learning-based detection (weed/crop classification, flying-insect and swarm detection, crop-infection severity assessment) with automated, chemical-free intervention (robotic laser weed elimination and Electronic Swarm Suppression, ESS). Rather than presenting the architecture only conceptually, this study implements and genuinely evaluates its core detection component as a transparent, reproducible pilot: a synthetic crop-versus-weed image dataset (300 images, procedurally generated) was built and used to train and compare a classical hand-crafted-feature baseline (Excess Green Index and shape descriptors with Logistic Regression) against a compact convolutional neural network implemented from scratch in NumPy - deliberately avoiding multi-gigabyte GPU-framework dependencies so the pilot remains fully reproducible on modest hardware. On an identical, stratified 70/15/15 train/validation/test split, the classical baseline achieved 95.6% test accuracy (F1 = 0.952, AUC = 0.998) and the from-scratch CNN achieved 97.8% test accuracy (F1 = 0.977, AUC = 0.994), both consistent with, and reported alongside, real literature benchmarks reviewed in this paper (including 99.5% locust-detection accuracy reported for a VGG19+classifier pipeline, and up to ~95% wheat-weed detection accuracy reported for YOLOv7/v8-based methods). The physical-intervention modules are evaluated as explicit, literature-grounded Monte Carlo simulations rather than physical hardware tests: a laser weed-elimination simulation shows this pilot's own classical detector supports only a 28.9% simulated weed-kill rate (21.5% crop damage) due to its centroid-localisation imprecision, while a projection using real trained-detector positional accuracy reported in the literature reaches 93.2% weed-kill with zero simulated crop damage, closely matching independently reported real-world field results; a synthetic harmful-versus-beneficial insect discrimination classifier (97.8% accuracy) combined with a real, field-measured laser-lethality rate on a genuine crop pest projects a 94.8% pest-suppression rate with 2.2% beneficial-insect collateral risk. These results are explicitly framed as a small-scale proof-of-concept and literature-grounded projection for the detection and intervention modules rather than a field-validated or hardware-validated deployment, since the robotic laser-elimination and ESS hardware components require physical implementation beyond this study's software-only environment. A structured literature review, comparative analysis, and research-gap synthesis are provided alongside the pilot, together with a concrete field-validation and hardware-implementation roadmap for future work..