Satellite Imagery-Based Property Valuation

A multimodal deep learning system that predicts property prices by combining tabular data with satellite imagery.

Overview

This project implements a Multimodal Regression Pipeline that predicts property market value using:

• Tabular Data: Property features (bedrooms, bathrooms, sqft, grade, etc.)
• Satellite Imagery: Visual environmental context captured via Google Maps API

The system uses a stacking ensemble approach:

1. XGBoost: Processes tabular features
2. ResNet18: Extracts visual features from satellite images
3. Meta-Learner: Combines both predictions for final output

Project Structure

satval/
├── data_fetcher.py          # Satellite image download script
├── preprocessing.ipynb      # Data cleaning and EDA
├── model_training.ipynb     # Multimodal model training
├── code_with_output.ipynb   # Complete notebook containing code with outputs
├── README.md                # This file
│
|
|(Below directories will be added once you run the code_with_output.ipynb file)
├── data/
│   ├── raw/                 # Original Excel files
│   ├── processed/           # Processed CSV files
│   └── images/              # Downloaded satellite images
│       ├── train/
│       └── test/
│
└── models/                  # Saved models and visualizations
    ├── cnn_model.pth
    ├── xgb_model.pkl
    ├── meta_learner.pkl
    └── gradcam_samples.png

Setup Guide (Kaggle) - Option- A

Directly go to below link and run the notebook. I have already downloaded the images which exists in /kaggle/working directory, so it will take less time to run. Make sure the accelerator is selected to GPU T4 X2.

Notebook - https://www.kaggle.com/code/mohitagarwal24/cdc-satellite-imagery-based-property-valuation

If you wish not to download images then I have made data uploaded to kaggle as well which can be accessed below for direct use, but these must be copied to /kaggle/working directory and should look like as follows:

Dataset - https://www.kaggle.com/datasets/mohitagarwal24/cdc-satellite-image-based-property-valuation-data

Setup Guide (Kaggle) - Option- B

1. Create Kaggle Notebook

• Go to kaggle.com/code
• Create new notebook with GPU T4 x2

2. Upload Data

• Download train.xlsx and test.xlsx
• Upload to Kaggle or add as dataset

3. Initial Setup

import os
from pathlib import Path

# Create directories
for d in ['data/raw', 'data/processed', 'data/images/train', 'data/images/test', 'models']:
    Path(d).mkdir(parents=True, exist_ok=True)

# Copy data (after uploading to Kaggle)
!cp /kaggle/input/*/*.xlsx data/raw/

# Set your API key
os.environ['GOOGLE_MAPS_API_KEY'] = 'YOUR_GOOGLE_MAPS_API_KEY'
os.environ['GOOGLE_MAPS_SIGNING_SECRET'] = 'YOUR_SIGNING_SECRET'

4. Run Notebooks

1. Run preprocessing.ipynb - Data cleaning & EDA
2. Run model_training.ipynb - Train models & generate predictions

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      Input Data                              │
├─────────────────────────────┬───────────────────────────────┤
│     Satellite Image         │       Tabular Features         │
│       (224×224)             │       (31 features)            │
└──────────────┬──────────────┴───────────────┬───────────────┘
               │                               │
               ▼                               ▼
        ┌──────────────┐              ┌──────────────┐
        │   ResNet18   │              │   XGBoost    │
        │    (CNN)     │              │  (GBM 200)   │
        └──────┬───────┘              └──────┬───────┘
               │                               │
               │ Predictions                   │ Predictions
               │                               │
               └───────────┬───────────────────┘
                           │
                           ▼
                  ┌────────────────┐
                  │  Meta-Learner  │
                  │    (Ridge)     │
                  └────────┬───────┘
                           │
                           ▼
                    Price Prediction

Results

Model	Val RMSE	Val R²
XGBoost (Tabular)	~0.166	~0.897
ResNet18 (Image)	~0.351	~0.552
Stacking (Fusion)	~0.159	~0.910

Model Explainability

Grad-CAM is used to visualize which image regions influence predictions:

• Green areas (vegetation) → Higher value
• Water proximity → Higher value
• Urban density patterns → Variable impact

Dataset Features

Original Features (18)

• Property: bedrooms, bathrooms, sqft_living, floors
• Quality: grade (1-13), condition (1-5), view (0-4)
• Location: lat, long, waterfront
• Temporal: yr_built, yr_renovated

Engineered Features (13)

• house_age, years_since_renovation
• total_rooms, bath_bed_ratio
• living_lot_ratio, quality_score
• is_luxury, has_basement, and more

API Setup

Google Maps Static API

1. Go to Google Cloud Console
2. Create project → Enable "Maps Static API"
3. Create API key → Enable billing

Dependencies

torch>=2.0.0
torchvision>=0.15.0
pandas>=2.0.0
numpy>=1.24.0
scikit-learn>=1.3.0
matplotlib>=3.7.0
seaborn>=0.12.0
pillow>=10.0.0
opencv-python>=4.8.0
tqdm>=4.65.0
openpyxl>=3.1.0