due to some confidential reasons the model codes cant be shown before publishing of the research paper.

Road Extraction from CartoSat-3 Satellite Imagery

Overview

This project focuses on automated road extraction from high-resolution satellite imagery using deep learning-based semantic segmentation. The objective is to generate accurate pixel-wise road masks from multispectral satellite data, enabling applications in urban planning, mapping, disaster response, and geospatial intelligence.

---

Dataset

Source

• CartoSat-3 satellite imagery

  • PAN (0.45m resolution)
  • Multispectral MX (1.10m resolution)

Cities Used

• Training: Ahmedabad, Hyderabad, Ludhiana
• Validation (Unseen): Delhi

This cross-city setup ensures true generalisation evaluation.

Dataset Creation Pipeline

The dataset was built from scratch using a multi-step geospatial pipeline:

1. Band stacking (B, G, R, NIR)
2. CRS conversion to UTM
3. OSM road extraction (Overpass API)
4. Alignment correction (+3m East, +12m North)
5. Road buffering (width generation)
6. Rasterization to pixel masks
7. Tile generation (512×512)
8. Filtering empty tiles

---

Feature Engineering

Instead of using only RGB, a 5-channel input representation was designed:

Input = [Blue, Green, Red, NDVI, NDWI]

Derived Features

• NDVI (Vegetation Index) → separates vegetation from roads
• NDWI (Water Index) → separates water bodies from roads

This improves model discrimination in challenging regions.

---

Data Preprocessing

Normalization

• Dataset-specific mean & standard deviation
• Avoided ImageNet normalization (not suitable for satellite data)

Data Augmentation

Applied using Albumentations:

• Horizontal & vertical flips
• Rotation (0°, 90°, 180°, 270°)
• Shift, scale, rotate
• Brightness/contrast adjustments
• Gaussian noise

---

Models Implemented

Five segmentation architectures were implemented and compared:

1. DeepLabV3+ (ResNet-101 + ASPP + scSE)
2. DLinkNet34 (ResNet-34 + D-Block)
3. UNet (DenseNet169 + scSE)
4. LinkNet (DenseNet169)
5. FPN (DenseNet169 + scSE)

---

Training Setup

• Loss Function:

  Total Loss = 0.7 × Dice Loss + 0.3 × Binary Cross-Entropy
  

• Input size: 512 × 512

• Channels: 5

• Activation: Sigmoid

• Threshold: 0.5

All models were trained using the same pipeline to ensure fair comparison.

---

Evaluation Metrics

• IoU (Intersection over Union)
• F1 Score
• Precision
• Recall

---