PlantCare AI - efficientnet_b3

Fine-tuned tf_efficientnet_b3.ns_jft_in1k on the PlantCity field dataset (12 crops, 52 classes).

metric value
test accuracy 0.9956
test macro-F1 0.9946
ECE (calibrated) 0.0030
input size 300x300

Split: group-aware stratified 72/14/14, near-duplicates removed, original photos only, on-the-fly augmentation. Divide logits by temperature (see plantcare_meta.json) before softmax for calibrated confidences.

Limitations: trained on leaves from two districts of Khyber Pakhtunkhwa (Charsadda, Chitral); expect a drop on other regions, cameras and crops. Use as decision support, not a diagnosis.

Dataset: Khan et al. (2025) PlantCity: A Comprehensive Image Based on Multi Crop Leaves disease in Pakistan, Mendeley Data (check its licence before releasing weights).

Downloads last month
23
Safetensors
Model size
10.9M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support