plantvillage-mobilenetv2

Plant leaf disease classifier for the capstone project AI-Powered Plant Disease Detection and Farmer Assistance (IIIT-H, Group 4). A farmer uploads a photo of a crop leaf; this model returns the most likely PlantVillage class and a confidence score. A LangGraph agent then explains the result and suggests treatment from a curated agricultural knowledge base.

• Base model: google/mobilenet_v2_1.0_224, fine-tuned end to end
• Classes: 38 (14 crops)
• Trained: 2026-10-11 · Hub id: neghani/plantvillage-mobilenetv2

Intended use

• Input: a single RGB photo of one leaf, roughly centred, taken in daylight.
• Output: PlantVillage class name (for example Tomato___Early_blight) plus softmax probability.
• Treat a top-1 probability below 0.5 as "not sure" and ask for a clearer photo or expert confirmation (see threshold sweep below).
• Not a substitute for a plant pathologist. Treatment decisions must be confirmed with a local agricultural extension officer.

Training data

• PlantVillage color images, stratified per-class split with seed 42: train 43445 · validation 5430 · test 5430.
• Stage 2 (tag v1.1 and later) also uses Cropped-PlantDoc (Singh et al. 2020, CC BY 4.0): 27 of 38 classes; the other 11 stay lab-only. See run_info.json for the PlantDoc commit and data manifest.
• Augmentation (phone-photo robustness): random resized crop (70-100% PlantVillage, 50-100% PlantDoc), horizontal and vertical flip, rotation up to 180°, brightness/contrast/saturation jitter, random Gaussian blur.
• Evaluation preprocessing: resize shortest edge, centre crop 224, normalise (same as the bundled preprocessor_config.json).

Results on the held-out PlantVillage test split

Confusion matrix: reports/confusion_matrix.png · Full per-class report: reports/classification_report.txt

Per-class precision / recall / F1

Confidence threshold sweep

Coverage is the share of test images with top-1 probability at or above the threshold; the remaining images would be answered with "not sure". Recommended CONFIDENCE_THRESHOLD for the API: 0.5 (lowest threshold whose accuracy on covered images is at least 97% with coverage of at least 50%).

Real-world images (field and web photos)

How to read the numbers

PlantVillage photos are taken under controlled conditions: one leaf, plain background. Accuracy on the held-out PlantVillage split above is therefore an upper bound, not what a farmer should expect from a phone photo. Images of the same leaf can appear in both the training and test splits, which inflates it further. The real-world figures below are the more honest ones.

Result

Test set: 25 images from Cropped-PlantDoc, spanning 25 classes that map to PlantVillage labels. Split evaluated: final. Evaluated on 2026-10-12 with the same chain as the live app (shared quality gate, crop strategy B, then this model on the 256×256 crop). The strategy and threshold were chosen on the tune split with model/eval_strategies.py.

Project target for the end-to-end pass rate: 50.0%. The point estimate is below the target. With few images the confidence interval is wide, so treat the figure as indicative. Stage 2 (v1.1) quotes the one-shot final test in model/final_report.md (284 images); that file is the only source of real-photo numbers. Do not mix it with a 25-image eval_real_world_set run.

Confidence threshold used: 0.85. It was fixed on the tuning split only; the final split is reported once.

Most common mistakes:

Per-image results, the per-class table and run.json (what was tested): reports/real_world_set/. Image list, labels and sources: real_world_set/ (images are not redistributed here). Set fingerprint 65f860f5c078 · model weights f1a98e77629c

The section is regenerated by the Colab notebook whenever the set, the weights or the threshold change.

Reproducing

python model/eval_real_world_set.py --set_dir model/data/real_world_set --split final \
  --predictor app_predictor:diagnose --strategy B --threshold 0.85

Limitations

• PlantVillage images are lab-style photos on plain backgrounds; accuracy on field photos is lower. The real-world section above reports the gap.
• Only the 38 PlantVillage classes are known. Other crops, other diseases, nutrient deficiencies and pest damage will be forced into the nearest known class; rely on the threshold and the "not sure" path.
• One leaf per photo. Mixed infections and whole-plant shots are out of scope.
• Confidence is a softmax probability. Stage 2 ships a fitted temperature in calibration.json (NLL on PlantDoc validation). Treat remaining overconfidence as a limitation.
• PlantDoc and the 50-image real-world set are web photos; they overstate phone-photo performance. The one-shot comparison is final_report_v1_1.md (not used for any choice).

Training details

• base_model: google/mobilenet_v2_1.0_224
• init_from: google/mobilenet_v2_1.0_224
• plantdoc_commit: ``
• code_commit: 01d9caf11d17ffbc918d6eb27ef4af90ef627da3
• code_dirty: False
• label_smoothing_factor: 0.0
• train_rows: 43445
• num_labels: 38
• epochs: 5.0
• learning_rate: 0.0005
• per_device_train_batch_size: 64
• weight_decay: 0.01
• warmup_steps: 0.05
• lr_scheduler_type: SchedulerType.COSINE
• fp16: True
• seed: 42
• wall_time_sec: 381.5
• device: cuda:0

Use in the API (Phase 5)

from transformers import pipeline

clf = pipeline("image-classification", model="neghani/plantvillage-mobilenetv2", revision="v1.0")
preds = clf("leaf.jpg", top_k=3)
# [{'label': 'Tomato___Early_blight', 'score': 0.93}, ...]
top = preds[0]
if top["score"] < 0.5:
    print("Not sure - please retake the photo or consult an expert")

Class names match the keys of api/data/knowledge/diseases.json in the project repository, so the prediction can be looked up directly for symptoms and treatment guidance.

Project repository: https://github.com/neghani/iiith-group4-capstone-plat-leaf-decease-detection

Disclaimer

Treatment suggestions produced from this model's output are general guidance only and must be validated with local agricultural experts before critical action.