Satellite Imagery-Based Real Estate Valuation

Project Overview

This project implements a Multimodal Neural Network to value real estate properties by combining traditional tabular data (bedrooms, square footage, year built) with satellite imagery. This approach quantifies the "location value" of a property by visually analyzing its surroundings (greenery, density, proximity to highways) using Computer Vision.

Key Features

• Multimodal Architecture: Fuses a ResNet-18 CNN (for images) with a Multi-Layer Perceptron (for tabular specs).
• Visual Forensics: Uses Grad-CAM to visualize what features (trees vs. pavement) drive property value.
• Geospatial Analysis: Incorporates Latitude/Longitude to map price hotspots.
• Log-Price Modeling: handles the skewed distribution of high-value properties.

Repository Structure

.
├── 23117140_final.csv     # Final Predictions Submission for Test Set
├── 23117140_report.pdf    # Comprehensive Project Report
├── data_fetcher.py        # Script to download satellite images from Google/Mapbox
├── model_training.ipynb   # Main notebook: Training, Validation, and Explanation (Grad-CAM)
├── preprocessing.ipynb    # Notebook for EDA and generating visualizations
├── requirements.txt       # Python dependencies
└── README.md              # This documentation

Setup Instructions

1. Environment Setup

Create a virtual environment and install the required dependencies:

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Data Acquisition

This project requires a dataset of house metadata (train.csv) and corresponding satellite images. To download images (if you have an API key):

python data_fetcher.py

Note: Ensure you have a valid Google Maps Static API key stored in a .env file as GOOGLE_MAPS_API_KEY.

Usage

Training the Model

Open model_training.ipynb in Jupyter Notebook/Lab:

1. Run All Cells to:
   • Load and preprocess the dataset.
   • Train the Multimodal Network (ResNet + MLP).
   • Compare performance against an XGBoost baseline.
   • Generate a submission.csv (renamed to 23117140_final.csv).

Visual Analysis (EDA & Forensics)

Run preprocessing.ipynb to generate the project's visual artifacts:

• Price Distribution Histograms
• Geospatial Price Maps
• Sample Image Grids
• Grad-CAM Explanation Heatmaps

Results

The multimodal model demonstrates that visual signals from satellite imagery provide a measurable improvement in valuation accuracy over tabular-only baselines, specifically by capturing neighborhood context (density, greenery) that simple metadata misses.

License

MIT License