Aerial deep learning detection of fall armyworm and cocoa swollen shoot virus disease in maize and cocoa

Overview

karaAgro AI is extending its crop disease detection capabilities to aerial platforms by training deep learning object detection models that identify fall armyworm (FAW) in maize and cocoa swollen shoot virus disease (CSSVD) in cocoa from drone imagery. The goal is to provide farmers and agronomists with a rapid, scalable method for detecting crop stress and generating prescription maps that support variable-rate pesticide application. By shifting disease scouting from ground-level manual inspection to drone-based automated detection, the technology aims to reduce scouting time, enable earlier intervention, and minimize unnecessary pesticide use.

The work builds directly on karaAgro AI's existing portfolio of convolutional neural network models that detect diseases in maize (common rust, gray leaf spot, northern leaf blight) as well as tomato, sweet pepper, potato, grapes, peach, strawberry, apple, and cherry. These models are currently integrated into the karaAgro AI Android application, which is in beta testing. The proposed project adapts this proven detection framework to the unique challenges of aerial imagery and two priority West African crops.

Technical specifications

Detection targets:

• Fall armyworm (Spodoptera frugiperda) damage signs in maize
• Cocoa swollen shoot virus disease symptoms in cocoa

Imaging platform:

• Unmanned aerial vehicle (UAV) equipped with a 4K RGB camera, based on validation work by The Hughes Lab at Penn State using a DJI Phantom 4
• Operational flight altitude of 10 meters or lower above the crop canopy, as determined by prior field assessments
• Planned integration of a near-infrared (NIR) sensor to support vegetation stress analysis

Modeling approach:

• Deep learning object detection models trained on annotated aerial imagery
• Bounding-box annotation of visible disease and pest signs performed on pathologist-confirmed infected fields
• GPU-accelerated training infrastructure for model development and validation

Intended outputs:

• Automated detection results deliverable through the karaAgro AI platform
• Variable-rate application maps for targeted pesticide deployment

Technology readiness level

The underlying convolutional neural network framework has already been validated at 98.3% accuracy for three maize foliar diseases and is operational within the karaAgro AI Android app in beta. The aerial detection extension is at an early research and development stage. Initial validation by The Hughes Lab at Penn State has confirmed the feasibility of using a DJI Phantom 4 drone at 10 meters or lower for human-visible detection of fall armyworm. The next phase involves collecting pathologist-confirmed aerial image datasets, annotating disease and pest symptoms, and training and evaluating object detection models over an estimated eight-month project timeline. The team is seeking funding for drones, GPUs, NIR sensors, and key personnel including computer vision engineers and plant pathologists, as well as scientific guidance from industry partners on data collection, annotation, and modeling practices.