Deep Learning Leaf Species Classifier

Fine-grained image classification across 176 leaf species using transfer learning with ResNet-18. Trained on ~14.7K labeled leaf images, achieving ~87% validation accuracy.

Overview

Identifying plant species from leaf photographs is a challenging fine-grained recognition task — many species share similar shapes, venation patterns, and colors. This project builds an end-to-end pipeline from data exploration to a deployable inference script.

Key results:

• 87% validation accuracy on 176 classes using ResNet-18 pretrained on ImageNet
• Clean CLI inference script for batch predictions on new images
• Full reproducible training pipeline with fixed seeds

Sample Predictions

Below are sample training images from the dataset spanning various species:

Project Structure

├── report.ipynb        # Full training notebook — data exploration, training, evaluation
├── predict.py          # Standalone CLI inference script
├── label_map.json      # Class index → species name mapping
├── requirements.txt    # Python dependencies
└── README.md

Dataset

• Training set: 14,682 labeled images across 176 species
• Test set: 3,671 unlabeled images
• Format: JPG images of varying sizes, resized to 224×224 for training

The dataset is not included in this repository. It consists of leaf photographs with species labels such as quercus_alba, acer_rubrum, pinus_echinata, etc.

Approach

Model: ResNet-18 with ImageNet pretrained weights, final fully connected layer replaced for 176-class output.

Training details:

• 80/20 stratified train/validation split (seed=42)
• Adam optimizer, learning rate 1e-3 with StepLR decay
• Data augmentation: random resized crop, horizontal flip
• Normalization: ImageNet mean/std
• Trained for 10 epochs on Google Colab (T4 GPU)

Preprocessing (train): RandomResizedCrop(224) → RandomHorizontalFlip → ImageNet normalize

Preprocessing (inference): Resize(256) → CenterCrop(224) → ImageNet normalize

Results

Metric	Value
Final Validation Accuracy	~87%
Number of Classes	176
Model Parameters	~11.2M
Training Time	~10 min (T4 GPU)

Common failure modes observed from misclassified samples:

• Confusion between species within the same genus (e.g., pinus_densiflora vs pinus_echinata)
• Visually similar leaves across different genera
• Unusual poses, partial views, or inconsistent lighting

Quick Start

Install dependencies

pip install -r requirements.txt

Run inference

python predict.py \
    --data_dir /path/to/dataset \
    --input_csv /path/to/test.csv \
    --model_path model.pth \
    --label_map label_map.json \
    --output_csv predictions.csv \
    --device cpu

The output CSV contains two columns — image (path copied from input) and label (predicted species name).

Train from scratch

• Open report.ipynb, set DATA_DIR to your dataset root, and run all cells. The notebook saves model.pth and label_map.json on completion.

• Model weights are not included due to file size. Run report.ipynb to train and generate model.pth.

Applications

This pipeline generalizes readily to other image classification problems with minimal modification:

• Agriculture & forestry: Crop disease detection, weed identification, tree species inventory for forest management — swap the dataset and retrain.
• Biodiversity monitoring: Field biologists can use this to identify species from phone photos, feeding into ecological surveys and conservation tracking.
• Invasive species detection: Train on native vs. invasive species to flag threats in new regions — the fine-grained classification approach handles subtle visual differences well.
• Food & produce quality control: Classify fruit/vegetable varieties or grade produce quality on a sorting line — same ResNet backbone, different labels.
• Medical image classification: Skin lesion classification, cell type identification, or pathology slide screening all follow the same transfer learning pattern with domain-specific data.
• Retail & e-commerce: Product categorization from photos — clothing types, furniture styles, or any catalog with visual categories.

The core architecture (pretrained ResNet + replaced classifier head + CLI inference script) is a general-purpose template. Change the dataset, adjust the number of output classes, and retrain.

What I'd Do Next

• Stronger backbone: Swap ResNet-18 for ResNet-50 or EfficientNet-B3 for better feature extraction
• Aggressive augmentation: Add color jitter, random rotation, CutMix/MixUp to improve generalization
• Learning rate scheduling: Use cosine annealing with warmup instead of StepLR
• Test-time augmentation (TTA): Average predictions over multiple crops/flips at inference
• Grad-CAM visualization: Show which leaf regions the model focuses on for each prediction
• Lightweight deployment: Export to ONNX or TorchScript for faster CPU inference

Tech Stack

Python · PyTorch · torchvision · scikit-learn · Pandas · Matplotlib · PIL

License

MIT