Poultry disease classifier (MobileNetV2)
Classifies a photo of chicken droppings as Coccidiosis, Healthy, Newcastle, Salmonella. Built to explore low-cost, phone-friendly early warning for smallholder poultry farmers in East Africa.
Results (held-out test set, 1022 images)
- Accuracy: 0.979
- Macro F1: 0.972
| Class | Precision | Recall | F1 | Test images |
|---|---|---|---|---|
| Coccidiosis | 0.994 | 1.000 | 0.997 | 316 |
| Healthy | 0.974 | 0.981 | 0.977 | 308 |
| Newcastle | 0.903 | 0.982 | 0.941 | 57 |
| Salmonella | 0.985 | 0.959 | 0.972 | 341 |
For reference, the dataset authors report 98.0% test accuracy with a fine-tuned MobileNetV2 (Machuve et al., 2022).
Data
Machine Learning Dataset for Poultry Diseases Diagnostics, farm-labelled images from Arusha and Kilimanjaro, Tanzania (Zenodo record 4628934). Split 70/15/15, stratified by class, seed 42.
Intended use and limits
- A screening aid to prompt a laboratory test or vet visit. It is not a diagnosis.
- Farm labels can be wrong; the authors also publish a smaller PCR-confirmed set (Zenodo 5801834) that should be used to check real-world accuracy.
- Trained on images from two Tanzanian regions and three breed types; accuracy elsewhere is unknown until tested.
- Newcastle disease has the fewest examples, so its errors deserve the closest attention.
Citation
Machuve D, Nwankwo E, Mduma N, Mbelwa J (2022). Poultry diseases diagnostics models using deep learning. Frontiers in Artificial Intelligence 5:733345.
Model by Mosi Africa.