💧 Water Segmentation using Deep Learning

This project focuses on image segmentation to detect water bodies in satellite images using deep learning. It includes a model built from scratch and another leveraging EfficientNetB0 as a pretrained encoder. The final model is deployed using Flask through a simple web interface.

Key Highlights

• Built a U-Net deep learning model from scratch.
• Applied transfer learning by using EfficientNetB0 as the encoder for U-Net.
• Engineered a new Water Index feature to improve segmentation quality.
• Used MinMaxScaler for input normalization.
• Developed a Flask web app for user-friendly deployment and image uploading.

Model Details

• Model 1: U-Net from scratch with custom convolutional layers.
• Model 2: U-Net with pretrained EfficientNetB0 encoder (ImageNet weights).
• Loss Function: Jaccard Loss
• Optimizer: Adam
• Metric: IoU Score
• Input Shape: 128×128×12
• Normalization: MinMaxScaler
• Additional Feature: Water Index added to input channels

Test Results

• Test Loss: 0.31085
• Test IoU Score: 0.82624

Deployment

• The model is deployed using a Flask app.
• Users can upload satellite images and receive predicted water masks.

🔗 Links

• Model Weights (Hugging Face):
  Segmentation_Feature_Extraction.h5

• Live Web App:
  Water Segmentation App