Leaf image
Drop a leaf photo here
or
One leaf, filling the frame, plain background, even light.
Prediction
notebooks/01_disease_model.ipynb, which writes
web/models/disease_model.json. Until then the feature
extractor still runs, so you can verify segmentation and inspect the
97 features below.
Extracted features (97)
Field inputs
10 sensor & agronomic inputsWater balance & decision
FAO-56 · Allen et al. 1998How this is computed
Total available water TAW = 1000 (θFC − θWP) Zr,
readily available water RAW = p · TAW, and root-zone depletion
Dr = 1000 (θFC − θobs) Zr
— FAO-56 (Allen et al., 1998). The classifier sees these derived
quantities alongside the raw sensor readings; the regressor predicts the
applied depth in mm.
Scan history
Models in this page
Why classical machine learning
Published CNNs reach 99%+ on the PlantVillage benchmark, but weigh tens to hundreds of megabytes and need a server at inference time. This page carries hand-designed colour, lesion-chromaticity, GLCM texture, LBP and shape features into a conventional classifier, so the deployed model is small enough to ship inside the page itself. Nothing you load here leaves your device, and the page keeps working with the network switched off.
Honest limitations
- The disease model is trained on laboratory-condition imagery. Accuracy on real field photographs is measurably lower — the project reports that gap rather than hiding it.
- The irrigation model learns the FAO-56 reference policy from noisy sensor observations. Its residual error is dominated by soil-moisture sensor noise, not by model capacity.
- This is decision support, not a prescription. Confirm any treatment with a local extension officer before applying agro-chemicals.