SAR Image Classification

This repository contains the implementation of a Synthetic Aperture Radar (SAR) Image Classification pipeline. The project explores image filtering techniques, feature enhancement, and machine learning-based classification to segment SAR images into meaningful categories: urban, vegetation, and water.

Project Overview

Synthetic Aperture Radar (SAR) images provide valuable data for remote sensing applications but suffer from noise, such as speckle, which can obscure meaningful information. This project addresses these challenges by:

• Applying advanced filtering techniques to enhance image quality.
• Using machine learning models for SAR image classification.
• Evaluating model performance using Mean Squared Error (MSE) and classification accuracy.

The goal is to improve SAR image interpretability and enable robust segmentation for geospatial and environmental analysis.

Key Features

• SAR Image Preprocessing: Reads HH and HV polarization bands, enhances intensity, and performs multilooking.
• Speckle Noise Reduction: Implements five filtering techniques to improve image clarity.
• Automated Image Segmentation: Uses Unsupervised (K-Means) and Supervised (SVM, Decision Tree, Random Forest) classifiers.
• Performance Evaluation: Assesses filter effectiveness (MSE) and classifier accuracy based on segmented outputs.

Methodology

1. Preprocessing & Feature Enhancement

• Loaded SAR image bands (HH and HV polarization) using GDAL.
• Enhanced image intensity and performed multilooking for improved visualization.

2. Noise Reduction (Filtering)

Applied the following speckle noise filters:

• Boxcar Filter: Averages pixel values within a moving window.
• Refined Lee Filter: Balances noise reduction with edge preservation.
• Median Filter: Removes noise while maintaining edges.
• Mean Filter: Applies a kernel-based averaging approach.
• Gaussian Filter: Smooths image using a Gaussian kernel.

3. Image Classification

• Unsupervised Learning: Used K-Means clustering to segment SAR images.
• Supervised Learning: Implemented Support Vector Machine (SVM), Decision Tree, and Random Forest Classifiers to classify urban, vegetation, and water areas.

4. Model Evaluation

• Filtering Performance: Measured using Mean Squared Error (MSE).
• Classification Accuracy: Evaluated based on segmentation quality and classifier performance.

Results

Filter Performance (MSE)

Filter	Mean Squared Error (MSE)
Mean Filter	30.535
Median Filter	36.116
Gaussian Filter	35.177
Refined Lee	14.0005 (Best)
Boxcar Filter	29.559

Key Insights:

• Refined Lee Filter provided the lowest MSE, indicating better noise suppression and feature preservation.
• Median and Gaussian Filters retained more image details but resulted in higher MSE due to residual speckle noise.

Classification Results

• Unsupervised Learning (K-Means) provided a preliminary segmentation but lacked class labels.
• Supervised Learning:
  • SVM, Decision Tree, and Random Forest successfully segmented images into urban, vegetation, and water.
  • Random Forest performed best, achieving clearer segmentation and robustness to noise.

Requirements

• Python 3.8+
• Required Libraries:
  • GDAL
  • Pillow
  • numpy
  • scipy
  • scikit-learn
  • matplotlib
  • astropy

Challenges Faced

Unrealistically High Classification Accuracy: During model evaluation, the SVM, Decision Tree, and Random Forest classifiers yielded over 99% accuracy, which is highly unusual for SAR image classification. This suggests a potential issue with dataset partitioning, class distribution, or feature scaling that requires further investigation.