Maize Crop Disease
Detection System

Project Overview

Crop disease costs smallholder farmers across Sub-Saharan Africa a large share of their yield, and most have no early warning system. Northern Leaf Blight can take an entire maize harvest, and without an affordable way to detect it, farmers often find out too late to stop the spread.

EWB Cornell's Digital Agriculture Subteam builds open-source unmanned aerial and ground vehicles (UAVs and UGVs) that run computer vision models in the field, so farmers learn a plant is infected while there is still time to act.

The Problem

• Northern Leaf Blight and other maize diseases cut Sub-Saharan harvests every year
• Smallholder farmers lack affordable early-detection technology
• Visual crop monitoring across large fields is impractical by hand
• Late detection leads to preventable losses affecting entire communities
• Commercial agricultural drones cost far more than a smallholder farm can spend

Our Technical Approach

The team builds custom UAV and UGV platforms designed to be affordable to construct using locally available components. The vehicles carry RGB and multispectral cameras that capture imagery as they traverse fields. Onboard deep learning models trained on annotated crop disease datasets analyze this imagery in real time to flag potentially diseased plants.

Machine Learning Pipeline

• Custom dataset of maize leaf imagery labeled for Northern Leaf Blight and healthy tissue
• PyTorch-based convolutional neural network for real-time disease classification
• Optimized for edge deployment on low-power embedded hardware
• Transfer learning from existing agricultural disease models accelerates training
• Field accuracy validation through controlled test plots

Hardware Platforms

Both UAV (drone) and UGV (ground vehicle) platforms are under development to address different field geometries and terrain types. The UAV provides rapid aerial survey capability for large fields, while the UGV can navigate between rows for close-up imagery and more detailed disease assessment. Both are designed with modularity and repairability as core principles.

• UAV platform with 20+ minute flight time and stabilized camera gimbal
• UGV with autonomous row-following navigation using LiDAR and camera fusion
• Modular sensor bay for interchangeable camera payloads
• Open-source hardware designs published for global replication

Intended Impact

By publishing all hardware designs, software, and training datasets as open source, the team aims to enable local engineers and agricultural organizations across Sub-Saharan Africa to build and deploy these systems independently. The goal is a self-sustaining ecosystem of affordable crop monitoring tools, not a product dependency.

Visit ewb-dig-ag.org for more details on the full project.