GANet: A Network With Just 0.001 Million Parameters That Promises to Detect 38 Plant Diseases
A preprint proposes a lightweight AI model that, according to its authors, identifies 38 plant diseases with 99.99% accuracy and runs on a Raspberry Pi 4b and a Jetson Nano. The results have not yet been peer-reviewed.

Catching a crop disease in time usually requires an expert eye or a powerful computer, and neither is easy to come by in the field. A group of researchers wants to close that gap with an artificial intelligence model so small that, they claim, it runs on low-power devices such as a Raspberry Pi 4b or a Jetson Nano.
The model is called GANet (Group Equivariance Attention Network). Anubhab Tripathi, Bibek Das, Asfak Ali and Sheli Shinha Chaudhuri describe it in a preprint posted on arXiv and submitted on October 21, 2026. A preprint is a paper shared before peer review, meaning other specialists have not yet formally evaluated it. For now, all the figures below are those reported by the authors themselves.
The paper starts from the climate. The authors note that global warming and climate change are driving plant diseases and affecting food production worldwide, which calls for continuous monitoring and early detection. Deep learning models (neural networks that learn to recognize patterns in images) could help. But most of them, the authors warn, require a lot of processing power and perform poorly on modest electronics. In addition, many previous studies focus on diseases of specific crops, which limits their scope.
Their proposal is a network designed to predict and monitor 38 types of diseases. It incorporates two components that the authors present as new: a Group Attention block and a Group-Wise Feature Fusion block. According to the abstract, this combination makes the network both sparse and efficient.
The numbers are the most striking part. The model has just 0.001 million parameters, or about a thousand, a tiny fraction of what is typical in computer vision. At that size, the authors report 99.99% accuracy on PlantVillage, a plant disease dataset. In tests on edge devices (hardware that processes data on-site instead of sending it to the cloud), the system reaches 5 frames per second on the Raspberry Pi 4b and 11 on the Jetson Nano.
The abstract does not specify how that speed was measured or what kind of images the model was tested on beyond the cited dataset. Nor does it say whether the code is available for others to reproduce the results. Accuracy this high from a model this small is the kind of figure that should be confirmed by third parties before it is taken for granted.
For Latin America, where agriculture depends on affordable tools, the model's usefulness will depend on whether it is validated with real crops and real-world conditions, something the available material does not address. The things to watch are whether the authors release the code, whether the paper passes peer review and whether other teams can replicate the performance on low-cost hardware.
Source: arXiv