Roofpedia: A Python repository from NUS Urban Analytics Lab

Roofpedia - Mapping Roofscapes with AI

Explore Sustainable Roofscapes Around the World »

This is the official repo of Roofpedia, an open registry of green roofs and solar roofs across the globe identified by Roofpedia through deep learning.

In this repo you will find:

A ready to use dataset of 1,812 manually labelled polygons of rooftop greenery and solar panels covering 8 cities.
A ready to use dataset of building footprints identified with Green Roofs and/or Solar roofs by Roofpedia. Download data and labels here
A straight forward pipeline to run prediction on your own satellite image dataset
A guide on how you can tweak the pipeline to detect and tag roof features to OSM building footprints (coming up)

Running Roofpedia

Steps:

Install prequisites
Download and extract weights and sample dataset
run predict_and_extract.py
get result!

1. Prerequisites

You could use environment.yml to create a conda environment for Roofpedia

conda env create -f environment.yml

For non-gpu users, use environment_cpu.yml instead.

conda env create -f environment_cpu.yml

2. Data Preparation

Download the pretrained weights and sample dataset here and extract them to the root folder of the repo.

For custom inference, datasets should be processed and placed in the results folder. See more details in later sections.

Prediction

Predictions can be carried out by running the following sample code. The name of the city depends on the name of each dataset.

  python predict_and_extract.py <city_name> <type>

A sample dataset is provided in the results folder with the name NY for prediction, just run

  python predict_and_extract.py NY Solar

for Greenroofs, run

  python predict_and_extract.py NY Green

See the result in NY_Solar.geojson or NY_Green.geojson in 04Result folder and visualise the results in QGIS or ArcGIS.

Custom Dataset

Custom Dataset pairs can be created with QGIS using tiling functions.

Create a WMTS satellite tile connection with any WMTS server. You can use Mapbox's WMTS server for good quality images.
With QuickOSM, query and download the building footprint of a desired area for prediction.
Save the building polygons to 01City folder.
Callup QGIS toolbar (Ctrl + Alt +T), in Raster Tools, choose Generate XYZ Tiles(Directory) to generate satellite tiles for the area by using Canvas Extent. Use Zoom 19 and save to 02Images/Cityname
You are now ready for prediction

A unified script in extracting building polygons and downloading satellite tiles from Mapbox is a work-in-progress.

Custom Dataset File Structure

The structure of the results folder is as follows:

📂results
┣ 📂01City
┃- ┗ 📊Cityname1.geojson
┃- ┗ 📊Cityname2.geojson
┣ 📂02Images
┃--- ┗ 📂Cityname1
┃--- ┗ 📂Cityname2
┣ 📂03Masks
┃--- ┗ 📂Green
┃--- ┗ 📂Cityname1
┃--- ┗ 📂Cityname2
┃--- ┗ 📂Solar
┃--- ┗ 📂Cityname1
┃--- ┗ 📂Cityname2
┣ 📂04Results
┃- ┗ 📊Cityname1_Green.geojson
┃- ┗ 📊Cityname1_Solar.geojson
┃- ┗ 📊Cityname2_Green.geojson
┃- ┗ 📊Cityname2_Solar.geojson

01City contains geojson files of building polygons
02Images contains a slippymap directory of satellite images. For the pre-trained models, a zoom level of 19 is required.
03Masks contains predicted masks of each tile according to object type 04Results contains final cleaned building footprints tagged with the specific object type

Training

By preparing your own labels, you can train your own model. Training options can be set under config/train-config.toml. The default folder to the dataset is the dataset folder. The dataset.py performs train-test-val split to the extracted XYZ file structure, named images for satellite images and labels for the polygon masks respectively. Once the data is prepared, run the following to train new models according to the labels. The labels are not limited to greenroof or solar panels, but can be any custom object pn the roof as long as sufficient labels are provided.

 python train.py

Paper

A paper about the work was published in Landscape and Urban Planning and it is available open access.

If you use this work in a scientific context, please cite this article.

Wu AN, Biljecki F (2021): Roofpedia: Automatic mapping of green and solar roofs for an open roofscape registry and evaluation of urban sustainability. Landscape and Urban Planning 214: 104167. doi:10.1016/j.landurbplan.2021.104167

@article{roofpedia,
  author = {Abraham Noah Wu and Filip Biljecki},
  doi = {10.1016/j.landurbplan.2021.104167},
  journal = {Landscape and Urban Planning},
  pages = {104167},
  title = {Roofpedia: Automatic mapping of green and solar roofs for an open roofscape registry and evaluation of urban sustainability},
  url = {https://doi.org/10.1016/j.landurbplan.2021.104167},
  volume = {214},
  year = 2021
}

Limitations, issues, and future work

Roofpedia is an experimental research prototype, which leaves much opportunity for improvement and future work.

As with all other machine learning workflows, the results are not always 100% accurate. Much of the performance of the predictions (e.g. classification of whether a building has a solar panel on its rooftop) depends on the quality of the input imagery. Therefore, some buildings are misclassified, especially in imagery in which it is difficult even for humans to discern rooftop greenery and photovoltaics, resulting in false positives and false negatives. However, when these results are aggregated at the city-scale, the results tend to be more accurate.

For future work, we hope to add more cities to our collection and add the temporal aspect to the project, tracking the evolution of greenery and solar panels through time.

License

Distributed under the MIT License. See LICENSE for more information.

Contact

Abraham Noah Wu, Urban Analytics Lab, National University of Singapore, Singapore

Acknowledgements

Roofpedia is made possible by using the following packages

PyTorch
GeoPandas
Robosat - loading of slippy map tiles for training and mask to feature function is borrowed from robosat

This research is part of the project Large-scale 3D Geospatial Data for Urban Analytics, which is supported by the National University of Singapore under the Start-Up Grant R-295-000-171-133.

We gratefully acknowledge the sources of the used input data.

Some of the aspects of the project and its name - Roofpedia - are inspired by Treepedia, an excellent project by the MIT Senseable City Lab to measure and map the amount of street greenery in cities from the pedestrian perspective, and compare cities around the world.

ualsg/Roofpedia