Development of an Offline Intelligent Tool for Early Detection of Maize Lethal Necrosis for Smallholder Farmers

Maize lethal necrosis (MLN) is a destructive viral disease caused by the co-infection of Maize chlorotic mottle virus and Sugarcane mosaic virus. It threatens food security across East Africa, including Tanzania, where smallholder farmers grow maize as a primary staple crop. Early diagnosis in the field is difficult because the initial leaf symptoms closely resemble those of several other common maize problems, and access to trained plant pathologists in rural areas is very limited. This paper presents M-Skana, an offline-capable progressive web application (PWA) that detects MLN from smartphone leaf photographs using a two-stage deep learning pipeline based on MobileNetV2 and TensorFlow.js. A needs-assessment questionnaire administered to 41 smallholder maize farmers confirmed strong demand for an offline, Kiswahili-supported diagnostic tool. Stage 1 of the pipeline distinguishes maize leaves from non-maize images, reaching an accuracy of 98.3 percent and a weighted F1-score of 0.983. Stage 2 classifies confirmed maize leaves as healthy, MLN infected, or other disease, reaching an accuracy of 95.7 percent and a weighted F1-score of 0.957. Both models run entirely in the phone browser, and the complete two-stage prediction was measured at between 7 and 14 seconds across four tested Android devices, with no data leaving the phone. These results suggest that a lightweight, on-device, two-stage approach can bring reliable MLN screening within reach of rural farmers who have unreliable or no internet access.A live demonstration of the M-Skana application is available at: mskana.vercel.app