A Review of Gemma and PaliGemma Vision-Language Models for Plant Disease Detection in Africa

Crop disease costs an estimated 220 billion USD a year, and smallholder farmers carry most of that loss with no pathologist to call. Image classification has long been the standard answer, and scores on datasets like PlantVillage now sit above 99%. Literature has established that models can learn background cues instead of disease features, a flaw the field rarely accounts for. This review asks whether newer models, built on Google's Gemma and PaliGemma systems, do better. They classify an image and explain what they see in plain text. Five databases were searched under a PRISMA 2020 protocol, and results were screened against an eight-domain bias checklist covering dataset quality, field testing, and error reporting. Eighteen studies met the criteria, and their results were uneven. Across the diagnostic sub-tasks reviewed, disease naming accuracy typically fell between 82% and 88%, symptom description between 39% and 52%, and treatment recommendation as low as 28% to 43%, a spread of over 40 percentage points from what a model can name to what it can reliably advise. Only two of the eighteen studies checked for fabricated output. None reported a calibration score or a rule for admitting uncertainty. Field accuracy, tested in a minority of cases, ran 10 to 40 points below laboratory accuracy. A deployed cassava detector in East Africa still beat the alternative farmers have: 65% in the field against 40 to 58% for extension agents and 18 to 31% for farmers alone. Cassava, yam, millet, and sorghum remain barely studied, and no study paired accuracy with power use, latency, and farmer judgment together. Closing that gap, and not chasing higher laboratory scores, will therefore decide whether this technology reaches the farmers who need it.