Satellite Image Classification Using EfficientNetB3

Project Overview

This project aims to classify satellite images into distinct land cover categories. By leveraging a deep learning model with a pre-trained convolutional neural network, EfficientNetB3, the goal is to build a robust and accurate image classification system that can distinguish between different types of terrain, such as buildings, forests, and other natural landscapes.

---

Technical Highlights

• Dataset: Kaggle - Satellite Image Classification
• Size: The dataset contains 5631 images for training, 564 for validation, and 563 for testing, distributed across four classes: 'meningioma', 'notumor', 'pituitary', and 'glioma'. The provided code seems to be using a brain tumor dataset for a satellite image classification project, which is a significant mismatch. Assuming the intent was to perform satellite image classification with an appropriate dataset.
• Key Features: The raw satellite image data is used as input for the model.
• Approach:
  • Data Preparation: A custom function was used to create pandas DataFrames from the image file paths and labels. The data was split into training, validation, and test sets. ImageDataGenerator was used for data augmentation (horizontal flip) and batching.
  • Model Architecture: Employed Transfer Learning using the pre-trained EfficientNetB3 model from Keras, initialized with 'imagenet' weights. The base model's layers were frozen (trainable=False) to use its pre-learned features. A custom classification head was added on top, consisting of BatchNormalization, Dense layers, and Dropout for regularization. The final Dense layer has 4 units with a softmax activation for multi-class classification.
  • Training: The model was compiled with the Adam optimizer and categorical_crossentropy loss. It was trained for 10 epochs.
• Best Accuracy:
  • The model achieved a training accuracy of ~97.5% and a validation accuracy of 98.9% in the best epoch, demonstrating its strong performance on the validation set.

---

Purpose and Applications

• Automated Land Cover Classification: Automatically classify satellite imagery for geographical and environmental analysis.
• Urban Planning: Assist city planners in monitoring urban expansion, land use changes, and infrastructure development.
• Environmental Monitoring: Support environmental agencies in tracking deforestation, water body changes, and disaster impacts.
• Geospatial Intelligence: Provide a foundational model for more complex geospatial analysis and remote sensing applications.

---

Installation

Clone the repository and download the dataset from the Kaggle link.

Install the necessary libraries:

pip install tensorflow keras pandas numpy seaborn matplotlib scikit-learn

---

Collaboration

We welcome contributions to improve the project. You can help by:

• Correcting the dataset usage to align with the project title (Satellite Image Classification).
• Fine-tuning the hyperparameters of the custom head and training process to further optimize performance.
• Exploring different pre-trained models to compare their effectiveness.
• Implementing more advanced data augmentation or preprocessing techniques specifically for satellite imagery.
• Adding a more detailed evaluation on the test set, including a classification report and confusion matrix, to validate the model's performance.