27Kushal/PIXEL-Srishti-Segmentation

🔗 View the Full Project Source Code on GitHub

---

PIXEL-Srishti: Geospatial Auditing Models 🌍

This repository contains the core custom-trained vision models for PIXEL-Srishti (Jaldrishti).

These models are designed for Edge Inference (running locally on standard laptops) to audit MGNREGA projects, detect illegal construction, and track water body development without relying on high-bandwidth cloud APIs.

Models Included

1. Land Cover Semantic Segmentation (latest_seg_model.pth)

• Architecture: DeepLabV3+ (with ResNet34 backbone and ASPP)
• Dataset: Trained on LandCover.ai
• Purpose: Classifies every pixel in a satellite image into 5 strict classes to identify exactly what is on the ground.
• Classes:
  1. Background
  2. Buildings (Infrastructure)
  3. Woodlands (Vegetation)
  4. Water Bodies (Ponds/Lakes)
  5. Roads

2. Structural Change Detection (latest_cd_model.pth)

• Architecture: Siamese U-Net
• Purpose: Takes a "Before" and "After" satellite image (Optical or SAR) and extracts deep structural features to output a binary mask of exact terrain alterations.
• Use Case: Validating if a contractor actually dug a new pond or moved earth, while safely ignoring seasonal color changes or shadows.

How to Use (Inference)

These weights are designed to be dynamically hot-swapped into VRAM by our custom Python Orchestrator to prevent Out-Of-Memory (OOM) errors on 8GB GPUs.

To run them in the PIXEL-Srishti backend, place both files into the pixel_srishti_ml/checkpoints/ directory:

import torch

# Load Siamese U-Net (Change Detection)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
checkpoint_cd = torch.load('checkpoints/latest_cd_model.pth', map_location=device)

# Load DeepLabV3+ (Segmentation)
checkpoint_seg = torch.load('checkpoints/latest_seg_model.pth', map_location=device)