Machine Learning 2022

Canberra Solar Footprint Mapping

Identified building footprints in Canberra equipped with solar panels by training a VGG16 binary classifier on aerial imagery, enhancing renewable energy mapping efforts for urban planning.

Project Type

Research Project

Canberra Solar Footprint Mapping - Solar panels aerial view

Key Results

High

Classification Accuracy

City-wide

Coverage

Supports

Urban Planning

Scalable

To Other Cities

The Challenge

Manual identification of solar panel installations across a city is time-consuming and quickly becomes outdated as new installations occur.

Key challenges included:

  • No automated method to detect solar installations
  • Manual surveys expensive and quickly outdated
  • Varying roof colors and materials complicate detection
  • Need for city-wide coverage at building level
  • Integration with existing urban planning systems

Urban Planning Context

Understanding solar adoption patterns helps urban planners assess grid capacity needs, identify areas for targeted incentive programs, and track progress toward renewable energy goals.

Renewable Energy

Track solar adoption citywide

Grid Planning

Inform infrastructure decisions

Our Solution

Trained a VGG16 deep learning model on labeled aerial imagery to automatically classify buildings with solar installations. Achieved high accuracy through careful data augmentation and transfer learning techniques.

Model Architecture

Input

Building footprint aerial tiles

VGG16 Base

Pre-trained ImageNet features

Fine-tuning

Custom classification layers

Output

Solar/No-solar classification

Transfer Learning

Leveraged VGG16 pre-trained on ImageNet as base model, fine-tuning final layers for solar detection task.

Data Augmentation

Applied rotation, flipping, and color augmentation to increase training data diversity and model robustness.

GDAL Processing

Used GDAL for efficient aerial imagery processing, tiling, and georeferencing.

Building Footprint Masks

Applied cadastral building footprints as masks to focus classification on individual buildings.

Binary Classification

Simple but effective binary classifier determining presence or absence of solar panels.

Confidence Scoring

Output includes confidence scores for quality filtering and manual review prioritization.

Technology Stack

PyTorch
VGG16
Python
GDAL
NumPy
Aerial Imagery
Rasterio
scikit-learn

Project Impact

Technical Achievement

  • High classification accuracy on Canberra buildings
  • Robust to varying roof materials and conditions
  • Efficient processing of city-wide aerial imagery
  • Methodology documented for reproducibility

Planning Applications

  • Supports renewable energy policy decisions
  • Identifies areas for targeted solar programs
  • Scalable approach for other Australian cities
  • Foundation for ongoing monitoring capability