ixrbhii/fig-disease-convnext

Fig leaf disease classifier

Tells which of 4 classes a fig leaf photo shows: Blight, Brown Spot, Healthy, Rust.

ConvNeXt-Small (convnext_small, from convnext_small.fb_in22k_ft_in1k), 224 × 224 RGB, 99 MB (half precision). 99.5 % test accuracy, macro-F1 0.993, on 372 held-out photos. Test photos come from the same datasets the model learned from, so real field photos score lower.

Class	Recall	F1	Test photos
Blight	100.0%	1.000	124
Brown Spot	97.1%	0.986	70
Healthy	100.0%	1.000	98
Rust	100.0%	0.988	80

Use

import timm, torch
from PIL import Image

model = timm.create_model("hf-hub:ixrbhii/fig-disease-convnext", pretrained=True).eval()
cfg = timm.data.resolve_data_config({}, model=model)
x = timm.data.create_transform(**cfg)(Image.open("leaf.jpg").convert("RGB")).unsqueeze(0)
p = model(x).softmax(-1)[0]
labels = model.pretrained_cfg["label_names"]
for i in p.argsort(descending=True)[:3]:
    print(labels[i], f"{100 * p[i]:.1f} %")

The training images were resized to 224 × 224 without cropping; timm's transform above (resize + 224 centre crop) gives the same result for square photos. label_descriptions in config.json maps each name to the dataset's original label. Best results: one leaf filling most of the photo, in daylight and in focus. Below about 45 % confidence, ask for a better photo.

In an app: LULC Fetch (Agri ▸ Diagnose crop disease) recognises the crop first and then runs the right disease model, with a photo-quality check and a disease map from geotagged photos.

Related

• All 44 models (2 crop detectors, 42 crops) as separate repositories: the collection
• All 44 models in one repository: ixrbhii/multicrop-disease-models
• Symptoms, treatment and pests for each disease: ixrbhii/crop-disease-qa

Training data and licence

Trained on LeafNet (CC-BY-4.0) and the authors' own field photos. Released under CC-BY-4.0: please credit "Multi-crop disease models, IXR (agritechixr)" and the training datasets. A diagnosis supports, but doesn't replace, a local agriculture expert.