Authors - Alfred ADINSI, Pelagie HOUNGUE Abstract - This systematic review evaluates AI-based techniques for rice dis-ease detection with a focus that existing surveys have not adopted: their deploy-ability in West African smallholder conditions, using Benin as the reference case. Based on 220 studies selected from 390 Scopus publications (2019–2025) via PRISMA, it goes beyond performance benchmarking to assess what actually works under resource constraints. Rice blast (70.9% of studies), brown spot (60.9%), and bacterial blight (44.5%) dominate the literature. Deep Learning accounts for 64.5% of approaches, hybrid methods for 21.8%, and classical Machine Learning for 13.6%. Mean accuracy reaches 94.2% for pure Deep Learning and 95.8% for hybrid architectures. Res-Net+ViT (96.4% ± 2.1%) and CNN+SVM (94.1% ± 4.1%) are the strongest per-formers, but performance alone is not the right metric for Benin. While 85% of studies apply to tropical climates, only 30.5% propose solutions running on limited hardware. Three approaches clear both bars: MobileNet+SVM (89.4%), optimized YOLOv8 (89.2%), and ResNet-based Transfer Learning (91–94% after fine-tuning). That AI can detect rice diseases accurately is no longer in question. The harder problem is which systems beninese farmers and extension agents can actually use. This review provides an answer.
Open Zoom