hodoly163/krishibondhu-jktk

KrishiDrishti-11X — Crop Disease Detection (YOLO11x, 116 classes)

KrishiDrishti-11X (render name; the weights file keeps the historical name jktk_x.pt) is the crop-disease detection model for KrishiBondhu AI — a Bangladeshi farming assistant by Team PentaFarmers. It detects plant diseases and healthy conditions across 116 classes / 30+ crops (rice, corn, potato, tomato, cassava, banana, citrus, grape, apple, and many more).

• Checkpoint: jktk_x.pt (~436 MB)
• Framework: Ultralytics (YOLO) + PyTorch
• Used in: KrishiBondhuVision and the KrishiBondhu demo Space

Provenance — verified from the checkpoint itself

The Ultralytics .pt checkpoint embeds its own training record. Loading jktk_x.pt with torch.load(..., weights_only=False) (or ultralytics) shows date: 2026-05-01T22:21:52, train_args, and train_metrics, giving an auditable evidence trail:

Claim	Embedded evidence
Base architecture	Ultralytics YOLO11x (name: yolo11x_optimized, task detect)
Training environment	Google Colab, GPU (device: 0, /content/... paths)
Run identity	project plant_disease_89class, run yolo11x_optimized
Continued from	resume: /content/drive/MyDrive/PlantDisease_TrainingBackup_20260501_171146/training_run/weights/last.pt — the team's May-2026 backup run
Epochs / batch / imgsz	100 epochs, 64 batch, 640 imgsz
Optimizer	AdamW, lr0 0.001, lrf 0.01, momentum 0.937
Augmentations	mosaic, mixup 0.2, copy-paste 0.2 (flip mode), RandAugment, HSV, random scale 0.5, erasing 0.4
AMP	enabled (amp: True)
Record date	2026-05-01T22:21:52 (UTC)

Datasets (merged, public)

1. 89-class public plant-disease detection base — classes 0–88 (the embedding run is literally named plant_disease_89class).
2. 27 field-disease classes from Roboflow Universe terratrendtest/planttest — cassava / corn / tomato field diseases form classes 89–115, matching that dataset's vocabulary exactly (e.g. Cassava Brown Leaf Spot, Corn Smut, Tomato bacterial wilt).

Recorded validation results (in-checkpoint train_metrics)

Metric	Value
mAP50	0.699
mAP50-95	0.686
Precision	0.635
Recall	0.682

Classes (116)

0: corn rust            1: cherry leaf           2: soybean leaf          3: strawberry leaf
4: corn northern leaf blight   5: coffee leaf    6: cauliflower alternaria leaf spot   7: celery leaf
8: strawberry anthracnose      9: basil leaf     10: citrus greening disease
11: grape downy mildew         12: blueberry rust        13: celery anthracnose
14: ginger leaf                15: cucumber angular leaf spot   16: broccoli downy mildew
17: blueberry leaf             18: cucumber bacterial wilt   19: grape leaf
20: garlic leaf                21: plum leaf             22: ginger sheath blight
23: rice leaf                  24: cabbage leaf          25: corn gray leaf spot
26: eggplant leaf              27: basil downy mildew    28: tomato bacterial leaf spot
29: cherry leaf spot           30: cucumber leaf         31: apple black rot
32: eggplant cercospora leaf spot   33: maple leaf       34: cucumber powdery mildew
35: bean rust                  36: rice sheath blight    37: zucchini yellow mosaic virus
38: bean leaf                  39: citrus canker         40: peach leaf curl
41: maple tar spot             42: tomato leaf mold      43: plum pocket disease
44: tomato early blight        45: cherry powdery mildew 46: grape leaf spot
47: bell pepper leaf spot      48: apple rust            49: cabbage alternaria leaf spot
50: tobacco mosaic virus       51: apple mosaic virus    52: potato late blight
53: tomato septoria leaf spot  54: potato leaf           55: lettuce mosaic virus
56: garlic leaf blight         57: grapevine leafroll disease   58: carrot cavity spot
59: cauliflower leaf           60: rice blast            61: tomato late blight
62: banana panama disease      63: ginger leaf spot      64: garlic rust
65: banana leaf                66: lettuce downy mildew  67: grape black rot
68: raspberry leaf             69: tomato mosaic virus   70: tomato leaf
71: squash powdery mildew      72: squash leaf           73: corn leaf
74: lettuce leaf               75: tobacco leaf          76: bean halo blight
77: apple leaf                 78: bell pepper leaf      79: broccoli leaf
80: apple scab                 81: strawberry leaf scorch  82: corn smut
83: coffee leaf rust           84: peach leaf            85: bean mosaic virus
86: celery early blight        87: potato early blight   88: tomato yellow leaf curl virus
89-115: the 27 Roboflow `terratrendtest/planttest` classes —
89: Cassava Bacterial Blight      90: Cassava Brown Leaf Spot   91: Cassava Healthy
92: Cassava Mosaic                93: Cassava Root Rot          94: Corn Brown Spots
95: Corn Charcoal                 96: Corn Chlorotic Leaf Spot  97: Corn Gray leaf spot
98: Corn Healthy                  99: Corn Insects Damages      100: Corn Mildew
101: Corn Purple Discoloration    102: Corn Smut                103: Corn Streak
104: Corn Stripe                  105: Corn Violet Decoloration 106: Corn Yellow Spots
107: Corn Yellowing               108: Corn leaf blight         109: Corn rust leaf
110: Tomato Brown Spots           111: Tomato bacterial wilt    112: Tomato blight leaf
113: Tomato healthy               114: Tomato leaf mosaic virus 115: Tomato leaf yellow virus

Usage

hf download hodoly163/krishibondhu-jktk jktk_x.pt --local-dir models

from ultralytics import YOLO

model = YOLO("models/jktk_x.pt")
res = model.predict("image.jpg", conf=0.25, imgsz=640, verbose=False)[0]
for box, cls, conf in zip(res.boxes.xyxy.tolist(), res.boxes.cls.tolist(), res.boxes.conf.tolist()):
    print(model.names[int(cls)], round(conf, 4), box)

Inference API (KrishiBondhu)

POST /predict multipart file → { detections: [{class, conf, box}], top }, then POST /advise returns RAG-grounded bilingual (EN/BN) treatment advice.

See the project repo.

Limitations

• No banana sigatoka class (documented coverage gap in our eval scripts).
• Roboflow planttest contribution is cassava / corn / tomato heavy — other Bangladeshi crops are represented through the 89-class base only.
• Verify field diagnoses with local agriculture extension officers.

License

AGPL-3.0 — this model was trained with Ultralytics tooling, whose AGPL-3.0 stewardship applies to derived work (the checkpoint also embeds Ultralytics' license notice). Our application code is separate. Contact the team for usage questions.