Image Segmentation and Classification of SAR Water Images

Introduction

This project focuses on leveraging deep learning models for the classification and segmentation of Synthetic Aperture Radar (SAR) water images. The primary objectives are:

1. Image Classification: Determine whether an image contains water using ConvNeXt, Vision Transformer (ViT), and ResNet models.
2. Image Segmentation: Segment regions containing water using U-Net, U-Net++, and Vision Transformer (ViT) architectures.

Data Used

• Dataset Source: EUROSAT and kaggle(https://www.kaggle.com/datasets/franciscoescobar/satellite-images-of-water-bodies)
• Data Description:
  • SAR images with varying resolutions.
  • Labels for classification: water and non-water.
  • Masks for segmentation: true binary masks indicating water regions.
• Preprocessing:
  • Normalization and augmentation techniques applied to improve model performance.
  • Data split: [80% training and 20% testing].

Classification Models and Performance

Models Used

• ConvNeXt
• Vision Transformer (ViT)
• ResNet

Performance analysis(Loss and Accuracy Curves):

ConvNeXt (Small):

ConvNeXt (Tiny):

ViT:

ResNet:

Segmentation Models and Performance

Models Used

• U-Net: A widely used architecture for biomedical image segmentation.
• U-Net++: An advanced version of U-Net with nested and dense skip connections.
• ViT: Adapted for image segmentation.

Visualization of Results

Below are examples of true masks and predicted masks:

UNET results:

Unet++ results:

ViT results:

Future work

• Exploring additional datasets for generalization.
• Implementing other advanced architectures.
• Fine-tuning hyperparameters for improved performance.