Apppublic

Prabhakar-Yadav05/Urban_Feature_Extraction

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

App README

GeoVision — Urban Feature Extraction Platform

A unified web platform for satellite image classification and segmentation, combining multiple deep learning models trained on panchromatic and multispectral imagery of Indian urban regions.

Models

Model	Type	Input	Classes	Dataset
JNPA V2 Patch Classifier	Random Forest	1-band PAN	5	JNPA 2.5m
CARTOSAT V3 4-Class Patch Classifier	Random Forest	1-band PAN	4	Cartosat
JNPA Dense U-Net V128	U-Net (dense)	1-band PAN	5	JNPA 2.5m
CARTOSAT Dense U-Net V128	U-Net (dense)	1-band PAN	5	Cartosat
JNPA U-Net V1	U-Net	1-band PAN	5	JNPA 2.5m
Vashi U-Net (4-Band)	U-Net	4-band Sentinel-2	7	Vashi
GBNM U-Net (3-Band RGB)	U-Net	3-band RGB	7	Greater Bhiwandi-Nijampur-Mira
Cartosat ResNet-18 Classifier	ResNet-18	3-band RGB	6	Cartosat
Vashi Mask R-CNN	Mask R-CNN	3-band RGB	2	Vashi (building detection)

Project Structure

Feature-Ext(Prabhakar)/
├── web_platform/                    # Unified Flask web platform
│   ├── app.py                       # Entry point
│   ├── config/
│   │   ├── model_registry.json      # All model configurations
│   │   └── prabhakar_models/        # Summary JSONs for multi-band models
│   └── platform_app/
│       ├── inference.py             # Core inference service
│       ├── prabhakar_inference.py   # Multi-band model inference
│       ├── prabhakar_models.py      # U-Net, ResNet-18, Mask R-CNN architectures
│       ├── model_registry.py        # Model loading and metadata
│       ├── image_utils.py           # Image loading and tiling
│       ├── unet_model.py            # Panchromatic U-Net architecture
│       ├── routes.py                # Flask API endpoints
│       ├── config.py                # Platform settings
│       ├── static/                  # CSS and JS
│       └── templates/               # HTML templates
│
├── BTP_Urban_Feature_Extraction(Prabhakar_)/
│   └── Models_Training/
│       ├── Aerial_RGB/              # GBNM RGB U-Net training
│       ├── Cartosat/                # Cartosat ResNet-18 training
│       └── Vashi_sentinel2/         # Vashi U-Net + Mask R-CNN training
│
├── BTP_Urban_Feature_Extraction(Vansh)/
│   ├── trained_models/              # Panchromatic model checkpoints
│   ├── training_code/               # Training scripts
│   ├── notebooks/                   # Training notebooks
│   └── final_results/               # Evaluation results
│
└── Test_Model/                      # Test satellite imagery

Setup

bash

git clone https://github.com/Prabhakar-Yadav/Urban-Feature-Ext.git
cd Urban-Feature-Ext

# Pull model checkpoints (requires Git LFS)
git lfs pull

# Create virtual environment
python -m venv venv310
venv310\Scripts\activate        # Windows
# source venv310/bin/activate   # Linux/Mac

# Install dependencies
pip install flask torch torchvision rasterio geopandas shapely joblib scikit-learn opencv-python-headless pillow pandas

# Run the platform
cd web_platform
python app.py

Open http://localhost:5050 in your browser.

Usage

1.Upload a satellite image (PNG, JPEG, GeoTIFF up to 500 MB)
2.Select Auto mode for automatic model selection, or Manual to pick a specific model
3.View segmentation results, class statistics, and download outputs (GeoTIFF, Shapefiles, CSV)

Features

—Automatic model selection based on domain similarity scoring
—Supports panchromatic (1-band), RGB (3-band), and multispectral (4-band) imagery
—Streaming mode for large GeoTIFFs (40M+ pixels)
—Sliding-window inference with Hann blending for smooth U-Net predictions
—Instance-level building detection via Mask R-CNN
—GeoTIFF and Shapefile export with CRS preservation
—Per-class area statistics and confidence metrics

Authors

—Prabhakar Yadav — Multi-band models (U-Net, ResNet-18, Mask R-CNN), platform integration
—Vansh — Panchromatic models (Patch Classifier, Dense U-Net), web platform UI, scoring system