iany-crop-v1

An on-device crop-health classifier — point a camera at a leaf and get its crop + condition (healthy / disease). A MobileNetV2 trained from open leaf datasets + community photos, exported to ONNX to run in the browser via onnxruntime-web. Built for iAny, the offline, on-device Khmer AI platform.

⚠️ Status: Experiment. Try it at iany.app/crop-scan — fully on-device, nothing uploaded. Help improve it by contributing photos at iany.app/crop.

What it does

Classifies a single crop leaf into a <crop>_<condition> class, for early, offline triage — "healthy vs a problem" — with advice that can be read aloud in Khmer. It is a suggestion, not an agronomist or a lab.

Labels (output order — this order matters)

10 classes (alphabetical — the softmax output order). labels.txt in this repo is authoritative and grows as more crops/conditions are trained.

0 cashew_disease
1 cashew_healthy
2 cassava_disease
3 cassava_healthy
4 maize_disease
5 maize_healthy
6 mango_disease
7 mango_healthy
8 vegetable_disease
9 vegetable_healthy

vegetable bundles tomato/potato/bell-pepper (not distinct iAny crops).

Input / preprocessing (important)

• Input: float32, shape [1, 224, 224, 3] (NHWC).
• Normalization: MobileNetV2 — scale pixels [0,255] → [-1,1] (i.e. x/127.5 - 1).
• Center-crop the frame to a square before resizing to 224×224. Identical input contract to iany-waste-v1, so the same runtime serves both.

Usage

Python (onnxruntime)

import onnxruntime as ort, numpy as np
from PIL import Image

labels = open("labels.txt").read().split()
img = Image.open("leaf.jpg").convert("RGB").resize((224, 224))
x = (np.asarray(img, np.float32) / 127.5 - 1.0)[None]      # [1,224,224,3], [-1,1]
sess = ort.InferenceSession("model.onnx")
probs = sess.run(None, {sess.get_inputs()[0].name: x})[0][0]
print(labels[int(probs.argmax())], float(probs.max()))

Browser (onnxruntime-web) — see src/lib/wasteOnnx.ts (generic [-1,1] MobileNetV2 classifier) and src/views/CropScanView.tsx for the live-camera implementation.

Training data

v1 is bootstrapped from open datasets (not yet fine-tuned on Cambodian photos):

• CCMT (Cashew, Cassava, Maize, Tomato; Ghana, field-captured, expert-validated) — cashew, cassava, maize, and vegetable (tomato)
• MangoLeafBD (CC BY 4.0) — mango
• PlantVillage — bell-pepper / potato / tomato → vegetable
• (optional) Cassava Leaf Disease 2020 (Kaggle) — harder real-field cassava

Base: MobileNetV2 (ImageNet weights), transfer learning. Full recipe: docs/CROP-MODEL.md.

Limitations

• v1 / Experiment — do not trust the headline accuracy. Reported validation accuracy is high (~0.99) but in-distribution: each crop comes from one dataset, so the model separates classes partly by dataset style (background, camera), not pathology. On real Cambodian phone photos accuracy is materially lower. The /crop-scan UI labels this a guess for exactly this reason.
• The fix is data, not training. Folding in real /crop field photos (and retraining) is what makes it trustworthy — expect the number to drop then, which is the point.
• Coarse conditions (healthy / disease) by design; CCMT's maize "diseases" include some pests. Naming the exact disease needs a later, finer model.
• Best on one leaf filling the frame, decent light.

Intended use

Offline crop-health triage + education; the /crop-scan experiment; pre-filling labels in the /crop collector. Not a certified diagnostic or a replacement for an agronomist — present results as guidance and encourage a second opinion for serious decisions.

License & attribution

Released under CC-BY-4.0 — please credit the source datasets (PlantVillage, MangoLeafBD — CC BY 4.0, Cassava, rice sets, iBean per their terms) and iAny. Verify each source dataset's terms before commercial redistribution.

Credit & recipe

Trained and released by iAny (E-KHMER Technology). Full training + deploy recipe: github.com/sengtha/iAny · docs/CROP-MODEL.md.