← All workApplied AI researchongoing2026

ZimCropGuard: multi-crop disease detection under domain shift

On this page
  1. In short
  2. The idea
  3. The central hypothesis
  4. Architecture
  5. Training strategy
  6. Recognition
  7. Deployment vision
  8. Where it stands

Status: research in progress. The results below are preliminary. This page describes an active programme, not a finished product, and any numbers should be read as work in progress rather than final claims.

In short

  • WhatCrop disease from one phone photo, across five staple crops.
  • HowEfficientNetB0 and ViT-B/16, fused, trained in stages for field conditions.
  • ResultPreliminary and ongoing; Best Poster, Deep Learning IndabaX Zimbabwe 2026.

The idea

ZimCropGuard is a hybrid deep-learning framework for detecting crop disease across multiple Zimbabwean staple crops, from a single phone photo, under the messy conditions farmers actually shoot in. It builds on my completed BSc Honours dissertation, a hybrid CNN-Transformer for tomato leaf disease that is already deployed, and scales the same robustness-first philosophy to a much larger, multi-crop problem.

Scale. 5 Zimbabwean staple crops, 21+ disease classes, and 54,000+ training images.

The central hypothesis

Models trained on clean laboratory imagery, such as PlantVillage, systematically underperform on real African field conditions. The distribution shifts across lighting, camera quality, background clutter, cultivar, and leaf position, and a model that has only ever seen lab data does not survive the move outdoors. The hypothesis under test is that domain-adaptive training measurably closes that gap. Quantifying the shift, and the recovery, is the point of the research.

Architecture

Two complementary backbones read each image in parallel:

  • EfficientNetB0 captures fine local texture, the lesion-level detail that distinguishes one disease from another.
  • ViT-B/16 captures global structure, the arrangement and context across the whole leaf.

Their features are concatenated and fused through learned dense layers, a learned-fusion ensemble rather than a simple average of two predictions, so the model can weigh local and global evidence per case.

ZimCropGuard learned-fusion architectureEach image is read in parallel by EfficientNetB0, for local texture, and ViT-B/16, for global structure. Their features are concatenated and fused through learned dense layers, a learned-fusion ensemble rather than an average of two predictions, before the disease class is predicted across 5 crops and 21 or more classes.Leaf photofield or laboratoryimageEfficientNetB0local texture:lesion-level detailViT-B/16global structure acrossthe leafFeatureconcatenationLearned dense fusionlayersweighs local and globalevidence per caseDisease class5 crops, 21+ classesZimCropGuard learned-fusion architectureEach image is read in parallel by EfficientNetB0, for local texture, and ViT-B/16, for global structure. Their features are concatenated and fused through learned dense layers, a learned-fusion ensemble rather than an average of two predictions, before the disease class is predicted across 5 crops and 21 or more classes.Leaf photofield or laboratoryimageEfficientNetB0local texture:lesion-level detailViT-B/16global structureacross the leafFeatureconcatenationLearned densefusion layersweighs local andglobal evidence percaseDisease class5 crops, 21+ classes

Training strategy

Four stages take the model from laboratory data to field calibration.

ZimCropGuard staged trainingFour training stages: freeze and pretrain the backbones on the combined dataset; train the fused head with the backbones frozen; fine-tune end to end at a low learning rate; then calibrate to the field domain on mixed laboratory and field images with focal loss.1. Freeze and pretrainbackbones on the combineddataset2. Head trainingfused classifier, backbonesfrozen3. Full fine-tuneend to end at a low learningrate4. Domain calibrationmixed laboratory and fieldimages, focal lossZimCropGuard staged trainingFour training stages: freeze and pretrain the backbones on the combined dataset; train the fused head with the backbones frozen; fine-tune end to end at a low learning rate; then calibrate to the field domain on mixed laboratory and field images with focal loss.1. Freeze and pretrainbackbones on the combined dataset2. Head trainingfused classifier, backbones frozen3. Full fine-tuneend to end at a low learning rate4. Domain calibrationmixed laboratory and field images, focal loss
Technical detail

A staged pipeline, designed for a large and imbalanced multi-crop dataset:

  1. Freeze and pretrain the backbones on the combined dataset.
  2. Head training on the fused classifier while the backbones stay frozen.
  3. Full fine-tune end to end at a low learning rate.
  4. Domain calibration on a mixed set of laboratory and real field images, with focal loss to handle heavy class imbalance across 21+ diseases.

Recognition

  • Best Poster, Deep Learning IndabaX Zimbabwe 2026.
  • Selected to represent Zimbabwe at the Deep Learning Indaba 2027.

Deployment vision

The intended delivery is a Streamlit web app plus an offline-capable progressive web app (PWA), so the tool works with intermittent connectivity in the field. Multi-language support is planned for English, Shona, and Ndebele, alongside a self-improving loop that learns from farmer-submitted uploads over time.

Where it stands

The dissertation-stage tomato model proved the robustness approach end to end and is deployed as a live demo. ZimCropGuard extends it to five crops with a larger dual-backbone architecture and an explicit domain-shift study. That study is ongoing, and the preliminary results are promising enough to have earned the poster award, with the full evaluation still in progress.

Want to talk about this work? Email me or download my CV.