Autonomous Crop Disease Surveillance

Problem

Crop disease is often detected late, after yield is already compromised. Visual inspection does not scale across smallholder landscapes, and expert agronomy is unevenly distributed.

Context

An applied computer-vision programme for detecting and monitoring plant disease from field imagery. The system uses image processing and machine learning techniques and was deployed as an Android application for use in the field.

Dataset

Field imagery and disease-labelled plant photographs collected for supervised visual recognition under real farming conditions rather than laboratory stills.

Methodology

Supervised visual recognition with convolutional models, evaluated around operational false-positive cost for farmers and extension officers who must act on alerts in the field.

Model / architecture

Deep convolutional neural networks for crop disease detection and classification from RGB field imagery, packaged for on-device inference in an Android application.

Results

A deployable mobile surveillance tool that brings computer vision from research into farmer-facing decision support, enabling earlier identification of crop disease without waiting for specialist visits.

Evaluation metrics

Modality

RGB field imagery

Task

Detection / classification

Deployment

Android application

Challenges

Domain shift between training images and farms, class imbalance across diseases, on-device compute constraints, and explanations farmers and extension officers can trust.

Impact

Earlier, cheaper surveillance of plant health: a concrete instance of AI for agricultural resilience, delivered where the farmer already is.

Tools / technology

• Python
• TensorFlow
• Computer vision
• CNN
• Android