Plant Disease Prediction

Built for — Growers and students who want a quick first check from a photo of a leaf.

• Organised a three-class leaf image dataset (Healthy, Powdery, Rust) into train, validation, and test splits.
• Built the training pipeline with on-the-fly augmentation (rescale, shear, zoom, horizontal flip) using Keras ImageDataGenerator.
• Designed and trained a compact CNN, plotted training vs. validation accuracy, and saved the model for reuse.
• Wrapped the model in a Flask app: upload an image, it is resized and normalised exactly as in training, and the predicted class is returned.

1. D1

   A small CNN as a deliberate baseline

   Two convolution + max-pooling blocks (32 and 64 filters), a 64-unit dense layer, and a 3-way softmax. Small enough to train on a laptop CPU in minutes, which makes it a clear baseline to improve on.

2. D2

   Augmentation instead of more data

   With roughly 440 training images per class, random shear, zoom, and flips help the model generalise rather than memorise the training set.

3. D3

   Identical preprocessing in training and serving

   The Flask app resizes to 225×225 and scales pixels to [0, 1] exactly as the training generator does — a common source of silent accuracy loss when it drifts.

Training images

1,322

Validation images

60

Test images

150

Classes

3

Figures taken directly from the repository.

• After 5 epochs the model reached 83.3% accuracy on the 60-image validation set (training accuracy 83.3%), with validation loss still falling.
• The 150-image test split was not formally evaluated in the notebook, and a sample Rust test image was predicted Healthy with near-equal probabilities (0.51 vs 0.49) — a clear sign the model needs more training and a proper test-set report.

• →Evaluate on the held-out test set with a per-class confusion matrix.
• →Try transfer learning from a pretrained backbone such as MobileNet or EfficientNet.
• →Add early stopping and train for more epochs while validation loss is still improving.