AI-powered maize leaf disease identification using field-image deep learning classifiers
University of Pretoria researchers are developing an AI-based image classifier that detects grey leaf spot and other maize leaf diseases from real-world field photos. Current proof-of-concept exceeds 70% accuracy on field-collected images, with a target above 90%. The long-term goal is an in-field diagnostic app for farmers and agricultural extension staff across sub-Saharan Africa.
Overview
Maize production across South Africa and sub-Saharan Africa is threatened by foliar diseases such as grey leaf spot (GLS), Northern leaf blight, common rust, and Phaeosphaeria leaf spot. Early, accurate identification in the field is critical for timely crop management decisions. This project develops an artificial intelligence classifier that can automatically identify maize leaf diseases from photographs captured under real field conditions, with a long-term goal of deploying the technology as an in-field diagnostic mobile app for farmers and extension officers.
Technical specifications
- Convolutional neural network (CNN) classifiers trained on a curated image database of approximately 2,500 manually annotated photographs of the four main maize leaf diseases, collected during field trips to KwaZulu-Natal and the Eastern Cape using cellphones to replicate real-world capture conditions
- Proof-of-concept baseline using the PlantVillage lab-image dataset achieved above 95% accuracy for individual diseases
- Binary CNN classifier for GLS presence or absence in field images has reached above 70% accuracy and is being refined toward the 90% target
- Planned improvements include expanding the training set by more than five times, annotating lesions with spatial coordinates, and applying advanced AI methods such as image segmentation and automated parameter tuning
- Designed to extend to multi-disease identification across GLS, Northern leaf blight, common rust, and Phaeosphaeria leaf spot
Technology readiness level
The technology is at an early-to-mid research and development stage. Proof of concept has been demonstrated on controlled lab images, and a field-image classifier has been built and partially validated on a South African dataset. Remaining work focuses on expanding and enriching the annotated training set, improving model accuracy, and developing a field-capture mobile application. The research team is seeking paid annotation support, app development collaboration, and field-image collection partnerships for the next growing season.
About University of Pretoria
UP is a comprehensive, multi‑campus public research university in Pretoria with a broad disciplinary base and significant postgraduate activity. Industry engages through co‑located, large‑scale labs at the Engineering 4.0 campus—home to national reference and training facilities developed with SANRAL and partners—and through clinical pathways via Steve Biko Academic Hospital. For streamlined collaboration, Enterprises University of Pretoria serves as a single contracting gateway for contract research and consulting, while a dedicated Technology Transfer Office manages IP and licensing. Research is supported by South Africa’s National Research Foundation and the South African Medical Research Council, alongside international sponsors. Startup creation is enabled by the on‑campus TuksNovation incubator.
delivered to your inbox.