Tree Species Classification

Submitted By : Divija Arora

Learning Objectives

GOAL

• Learned to clean and preprocess a real-world image dataset.
• Learned to detect and handle duplicates and corrupted images.
• Building and training CNN models using TensorFlow and Keras.
• Gaining practical experience with image preprocessing and data augmentation techniques.
• Understanding and comparing different CNN architectures
• Learned to fine-tune hyperparameters to improve model accuracy.
• Using Google Colab for GPU-accelerated model training.
• Gained hands-on experience in model evaluation and performance visualization.

Tools and Technology used

• Python
• Google Colab
• TensorFlow / Keras
• Matplotlib
• Scikit-learn
• Pandas
• PIL (Pillow)
• EfficientNet
• CNN (Convolutional Neural Networks)
• GitHub (for version control)

Methodology

We implemented and compared the following three models:

​

​

• Basic CNN Model –

• A sequential CNN with 3 convolutional blocks, each followed by MaxPooling.

​

• Ends with a Flatten layer, Dense hidden layer, Dropout, and a Softmax output layer.

​

• Trained from scratch.

​

Methodology

• EfficientNetB0 –

• Helped stabilize training and improve convergence.

​

• Achieved the highest accuracy among the three.

• CNN with Batch Normalization –

• Loaded the EfficientNetB0 model pre-trained on ImageNet without the top layer.

​

• Kept the base model frozen initially to use learned features efficiently.

​

• This significantly improved accuracy.

All models were trained on a labeled dataset of tree images across 31 species.

Problem Statement:

Identifying tree species from images is a challenging task due to variations in lighting conditions, camera angles, leaf shapes, and complex backgrounds.

Manual identification can be time-consuming and prone to errors, especially when dealing with large datasets or visually similar species.

�Goal is to develop a robust and efficient deep learning model that can automatically and accurately classify tree species based on input images — regardless of lighting, background, or orientation.

Solution:

We built a deep learning pipeline that:

• Ensured consistent image dimensions and quality for model readiness.
• Improved data integrity by eliminating low-quality inputs (removes duplicates, outliers, and corrupted images).
• Applied transformations like rotation, zoom, and flipping to improve generalization.
• Trains three CNN models to classify tree species.
• Compares performance to choose the best model.

Github Link : https://github.com/Divija-Arora/edunet_internship

Screenshot of Output:

EfficientNetB0:

Achieved a low validation accuracy of approximately 8.1%, indicating underperformance.

Screenshot of Output:

Basic CNN:

Reached around 25.7% validation accuracy, showing gradual improvement over epochs.

Screenshot of Output:

CNN with Batch Normalization:

Performed best out of the three, with a validation accuracy of approximately 30.6%, demonstrating better generalization.

Screenshot of Output:

Final Tuned CNN Model:

Achieved ~98% training accuracy using a CNN enhanced with Batch Normalization and Dropout.

Conclusion:

• Built an end-to-end deep learning pipeline for tree species classification.
• Learned to handle real-world challenges like duplicates, varying image sizes, and data imbalance.
• Compared multiple CNN models and improved accuracy through experimentation.’
• Achieved ~98% training accuracy after fine-tuning EfficientNetV2S with regularization and early stopping.
• Used industry-standard tools like TensorFlow, Keras, and EfficientNet effectively.
• Future Scope : Expand to more tree species and global datasets.