Plant disease classification from leaf images: a fine-tuned CNN, a corrected classical baseline to compare it against, and Grad-CAM to check that the model is looking at the lesion rather than the background.
Main window. The cursor-tracked spotlight drifts on its own when the mouse is idle.
Diagnosis result. Runner-up classes are shown alongside the top prediction, and a low-confidence warning appears when the margin between the top two is narrow.
v1 was a Random Forest over ten hand-crafted features, reporting 79% accuracy. Auditing that pipeline turned up two bugs that had quietly disabled most of it.
area_ratio was computed as countNonZero() on the k-means reconstruction.
After reconstruction the image contains exactly two colours, and both are
non-zero, so the ratio was always 1.0. Measured across synthetic leaves with
lesion coverage from 0% to 100%:
| Lesion coverage | v1 area_ratio |
v2 lesion_ratio |
|---|---|---|
| 0% | 1.0000 | 0.0000 |
| 25% | 1.0000 | 0.0796 |
| 50% | 1.0000 | 0.1572 |
| 75% | 1.0000 | 0.2214 |
| 100% | 1.0000 | 0.2826 |
One of ten features carried zero information. The classifier was training on nine.
GLCM ran after segmentation with levels=256. But the segmented image has
only 2 distinct grey levels, so the 256×256 co-occurrence matrix was 99.99%
empty and contrast/correlation/energy were close to meaningless. The original
image had 37 distinct levels; a real photograph has 200+.
Colour features had the same root cause — mean and standard deviation taken over a two-colour reconstruction discard the actual colour distribution, which is the strongest signal available for this task.
PlantVillage images sit on a uniform grey backdrop, so the dominant k=2 split is leaf vs. background, not lesion vs. healthy tissue. And on a healthy leaf, k=2 still forces a split, partitioning sensor noise into hundreds of speckle "lesions".
v2 masks the background first (HSV saturation + largest connected component), clusters within the leaf, picks the lesion cluster deterministically by greenness, and applies morphological opening plus an area filter so only spatially coherent regions survive.
These fixes are in baseline/features.py. The classical pipeline is kept
rather than deleted, because "the CNN beats the classical approach" only means
something if the classical approach was implemented correctly.
| Component | Implementation |
|---|---|
| Main model | EfficientNet-B0, ImageNet-pretrained, two-stage fine-tune |
| Baseline | Random Forest over 60 corrected hand-crafted features |
| Explainability | Grad-CAM via forward/backward hooks (no extra dependency) |
| Interface | CustomTkinter desktop app |
Training freezes the backbone and trains the new head first, then unfreezes and fine-tunes with a lower trunk learning rate. Model selection uses macro-F1, not accuracy — on an imbalanced dataset, accuracy is dominated by the largest classes and will happily pick a model that has given up on the rare ones.
pip install -r requirements.txt
export PYTHONPATH=$PWD/src # Windows: $env:PYTHONPATH = "$PWD\src"Download PlantVillage
and point --data-root at the color/ subdirectory (the archive also ships
grayscale/ and segmented/; passing the parent reads as three classes).
# Train
python -m agrivision.train --data-root "dataset/plantvillage dataset/color"
# Evaluate on the held-out test split
python -m agrivision.evaluate --data-root "dataset/plantvillage dataset/color"
# Classify an image
python -m agrivision.predict path/to/leaf.jpg
# See where the model looked
python -m agrivision.explain path/to/leaf.jpg
# Classical baseline, on the same splits
python baseline/train_baseline.py --data-root "dataset/plantvillage dataset/color"
# Desktop app
python app_ui.pyThe app runs without a trained model — it reports that the checkpoint is missing and tells you how to produce one, rather than failing when you click Run.
PlantVillage is laboratory data: every leaf sits on the same flat studio backdrop. A network can score near-99% by reading that backdrop and the leaf silhouette without ever attending to a lesion — and such a model collapses on field photographs. High accuracy here is routine and does not, on its own, indicate a useful model.
So evaluate.py re-scores the same test set with the background replaced by
random colour or noise, and reports the gap:
=== Randomised-background probe ===
accuracy 0.9xxx (-0.0xxx vs clean)
Accuracy drop from background swap: 0.0xxx
A large drop means the score was resting on the backdrop. Training with
--randomize-background 0.5 trades a little clean accuracy for a model that
generalises. Grad-CAM is the visual version of the same check: heatmap on the
lesion is good, heatmap on the background means the model is cheating no matter
what the metrics say.
| Model | Test accuracy | Macro-F1 | Accuracy w/ randomised background |
|---|---|---|---|
| v1 Random Forest — 10 features, 2 inert | 0.79 | — | — |
| v2 baseline — 60 corrected features | TBD | TBD | TBD |
| v2 EfficientNet-B0 fine-tuned | TBD | TBD | TBD |
Splits are computed once, written to artifacts/splits.json, and reused by every
stage including the baseline — so the comparison is like-for-like rather than
across two different random partitions.
Fixes beyond the feature pipeline:
- Three-way split. v1 used a single 80/20 split and reported test accuracy on data that had also driven hyper-parameter choices. There is now a test set touched exactly once.
- Checkpoints carry their schema. v1 saved a bare
.pkl; if the feature code changed, the old model kept predicting, just wrongly. Checkpoints now embed the class list, image size, normalisation and a schema version, and loading refuses on mismatch. - Stratified splits, balanced class weights, cross-validated baseline scores.
- Cached, parallel feature extraction in the baseline, so re-tuning doesn't re-extract 54k images serially.
- GUI:
state("zoomed")was Windows-only and raisedTclErrorelsewhere; inference ran on the UI thread behind a blocking sleep; a missing model file failed silently until the user clicked Run; disabled buttons kept their accent fill and still looked clickable; the placeholder advertised drag-and-drop that was never implemented. All fixed, and the app now shows runner-up classes and an explicit low-confidence warning.
app_ui.py Desktop GUI
src/agrivision/
data.py Splits, transforms, background-randomisation probe
model.py Backbone factory, staged freezing, checkpoint bundling
train.py Two-stage fine-tuning
evaluate.py Test metrics, confusion matrix, calibration, robustness
predict.py Reusable Predictor + CLI
explain.py Grad-CAM
baseline/
features.py Corrected hand-crafted features
train_baseline.py Random Forest on shared splits
Trained on laboratory images with controlled lighting and uniform backgrounds; field performance will be lower. Expected calibration error is reported because the confidence percentage shown to the user is misleading if the model is overconfident. This is a research prototype, not agronomic advice.
MIT — see LICENSE.