CompletedPythonTensorFlowVS CodeGit

Rooftop Solar Energy System

A Deep Learning-powered tool that maps rooftop solar potential by extracting building footprints from satellite imagery using a ResNet50-U-Net model

Timeline4 months

RoleML Developer

TeamGroup

StatusCompleted

Source Code

Overview

The Rooftop Solar Energy Potential Map is a comprehensive geospatial analysis project that identifies and visualizes the potential for solar energy generation on rooftops within specific geographic areas. By leveraging advanced machine learning models and geospatial data analysis, the project automatically calculates solar energy consumption potential, helping governments and organizations make informed decisions about where to install solar panels.

The project addresses two key objectives:

Building Footprint Extraction: Detects and delineates building footprints from satellite images using a sophisticated ResNet50-U-Net deep learning model.
Solar Energy Estimation: Calculates potential solar energy generation based on solar radiance data and user-defined parameters.

Screenshots

App Screenshot

Project Architecture

ResNet50-U-Net Model

The project utilizes a hybrid deep learning architecture combining:

Encoder: ResNet50 extracts hierarchical features from input satellite images
Decoder: U-Net architecture reconstructs detailed segmentation masks to capture building footprints

Performance Metrics

Validation Accuracy: 95.21%
Mean IoU Before Post-Processing: 0.85
Mean IoU After Post-Processing: 0.92

Model Architecture

Project Components

1. Data Preparation

split_folders.ipynb

Organizes and splits the dataset into training and validation sets
Ensures proper data division for model training and evaluation

create_mask.ipynb

Generates binary masks from satellite images for model training
Creates masks highlighting building footprints (white on black background)
Essential for supervised learning of the segmentation model

2. Model Training

model_resnet.ipynb

Implements and trains the ResNet50-U-Net model
Handles the complete training pipeline including data loading, model compilation, training, and evaluation
Saves the trained model for inference

model_architecture_resnet.json

Contains detailed JSON specification of the ResNet50-U-Net architecture
Documents layer specifications, activations, and network connections

3. Prediction & Analysis

prediction.ipynb

Generates building footprint predictions using the trained model
Applies post-processing techniques to refine segmentation results
Handles inference on new satellite images

area_finder.ipynb

Calculates real-world area of buildings from model predictions
Converts pixel-based predictions to actual square meters
Essential for accurate solar potential calculations

Building Area Calculation

The project uses a sophisticated method to convert pixel-based predictions to real-world measurements:

Formula

Building Area (m²) = (Pixel Area (m²) / Conversion Factor) × Number of Pixels

Process

Extract Geographic Coordinates: Uses user-provided coordinates to define the geographic bounds of the image area
Calculate Conversion Factor: Determines the pixel-to-world conversion factor by relating real-world distances (latitude/longitude) to pixel dimensions
Compute Building Area: Multiplies the pixel area by the number of pixels representing each building footprint

Key Components

Pixel Area (m²): The area represented by a single pixel in square meters
Conversion Factor: Ratio of real-world dimensions to pixel dimensions, based on geographic coordinates and image resolution
Number of Pixels: Total count of pixels within each building's footprint as predicted by the model

This method ensures accurate area calculations by integrating geographic and pixel-based information, providing precise measurements of building footprints in real-world units.

Usage Workflow

Data Preparation Phase

Use split_folders.ipynb to organize your satellite image dataset
Generate training masks using create_mask.ipynb to create labeled data

Model Training Phase

Train the ResNet50-U-Net model using model_resnet.ipynb
Monitor training progress and validation metrics
Save the trained model for inference

Prediction & Analysis Phase

Use prediction.ipynb to generate building footprint predictions on new satellite images
Apply post-processing to refine segmentation results
Calculate building areas using area_finder.ipynb with geographic coordinates
Integrate solar radiance data to estimate energy potential

Technical Notes

Ensure file paths and directories are correctly configured in all notebooks
Review model_architecture_resnet.json for detailed model architecture information
Geographic coordinates must be accurate for precise area calculations
Post-processing significantly improves segmentation quality (IoU: 0.85 → 0.92)