Paddy Leaf Disease Detection using Transfer-Learning CNNs with Grad-CAM Explanations

NelScan is a research-oriented M.Sc Computer Science dissertation that classifies common paddy (rice) leaf diseases from photographs and explains each prediction with a Grad-CAM heatmap. Paddy is the principal crop of the Cauvery delta and much of Tamil Nadu, and diseases such as bacterial leaf blight, brown spot, leaf smut and blast can cut yields sharply if they are not recognised early. Extension staff are few, and a farmer who photographs a leaf needs an answer and a reason to trust it.

The project uses public rice-leaf disease image datasets — for example the UCI Rice Leaf Diseases dataset (three classes) — optionally supplemented with a small set of field photographs collected with the farmer's consent. Images are cleaned, resized and augmented, and transfer learning is applied with three ImageNet-pretrained backbones: MobileNetV2, EfficientNet-B0 and ResNet50, in TensorFlow-Keras. Class imbalance is handled with class weights and balanced augmentation, and every model is evaluated with stratified 5-fold cross-validation, confusion matrices and macro-F1.

Grad-CAM shows which region of the leaf drove each decision, so the student can check whether the network is looking at lesions or at the background. A lightweight Streamlit demo lets a user upload a photo and see the predicted class, confidence and heatmap. An optional MATLAB colour-feature baseline gives a classical comparison, linking the work to the Digital Image Processing and AI & ML courses.

Abstract

Early identification of paddy leaf diseases helps farmers apply the right treatment before losses spread. This dissertation investigates transfer learning with three convolutional neural network backbones — MobileNetV2, EfficientNet-B0 and ResNet50 — for classifying paddy leaf images into disease classes, and uses Gradient-weighted Class Activation Mapping (Grad-CAM) to explain predictions. The study follows a documented protocol: stratified 5-fold cross-validation, class-imbalance handling, a held-out test split and an optional field-photo test to measure the drop from laboratory images to real conditions. Results are reported in a pre-defined table; no accuracy is claimed before the experiments are run.

Introduction

In Tamil Nadu, paddy is grown across the kuruvai, samba and thaladi seasons, and leaf diseases are usually identified by eye by the farmer or an agricultural officer. Photographs taken on a phone are an easy input, and convolutional neural networks have become the standard tool for image classification. Training a network from scratch needs tens of thousands of labelled images; transfer learning reuses features learned on ImageNet and fine-tunes them on a few hundred or thousand leaf images.

Existing Approaches and Research Gap

• Classical pipelines segment lesions and extract colour and texture features (for example colour histograms or GLCM texture) followed by SVM or k-NN; they are interpretable but brittle under changing light and background.
• CNN studies on public leaf datasets often report very high accuracy from a single random split, on clean images with plain backgrounds, and rarely show why the model decided.
• Gap: a careful comparison of lightweight and heavy backbones under stratified k-fold, with explicit imbalance handling, Grad-CAM inspection of failure cases and an honest field-photo test.

Proposed Work

• Assemble and audit the dataset (duplicates, label noise, class counts).
• Build an augmentation and preprocessing pipeline.
• Fine-tune three backbones with identical protocol; compare accuracy, macro-F1, model size and inference time.
• Generate Grad-CAM heatmaps for correct and incorrect predictions and analyse them.
• Deploy the best lightweight model in a Streamlit demo.

Feasibility

• Technical: Python, TensorFlow and OpenCV are free; training runs on a laptop GPU or a free cloud notebook GPU.
• Economic: no licence cost; MATLAB is used only where the college lab already has it.
• Operational: the demo needs only a photo upload.

Paddy leaf diseases such as bacterial leaf blight, brown spot, leaf smut and blast reduce yield when they are not identified early, yet farmers depend on visual judgement and occasional visits from extension staff. Automated image classifiers exist in research papers, but many are evaluated on a single random split of clean laboratory images, give no explanation for their predictions and are too heavy to run on modest hardware, so their reported accuracy says little about real use.

This dissertation addresses the problem of building a reliable and explainable paddy leaf disease classifier: comparing transfer-learning CNN backbones of different sizes under a rigorous stratified cross-validation protocol with class-imbalance handling, measuring performance on independent field photographs, and using Grad-CAM to verify that predictions are driven by disease lesions rather than background artefacts.

In scope

• Leaf-level image classification into a fixed set of disease classes plus healthy (where the dataset has a healthy class).
• Three pretrained backbones, one training framework (TensorFlow-Keras), one explanation method (Grad-CAM).
• Stratified k-fold evaluation, a held-out test split and an optional small field-photo test set.
• A local Streamlit demo.

Out of scope

• Pest identification, nutrient deficiency, whole-plant or drone imagery.
• Severity estimation or pesticide recommendations; the demo advises consulting an agricultural officer.
• Mobile app deployment (suggested as future work with TensorFlow Lite).

The dissertation follows an experimental research design in six stages over about 24 weeks.

Evaluation protocol. Hold out 15% of images as a final test set, stratified by class. On the remaining 85%, run StratifiedKFold(n_splits=5, shuffle=True, random_state=42); augmentation is applied only to training folds. Report mean ± standard deviation of accuracy and macro-F1 across folds, the confusion matrix on the test set, model size (MB), parameters and average CPU inference time per image. If field photos are collected, report them separately — they are never mixed into training.

Results table to fill: rows = SVM baseline, MobileNetV2, EfficientNet-B0, ResNet50 (each with and without class weights); columns = CV accuracy, CV macro-F1, test accuracy, test macro-F1, field macro-F1, size, inference time.