Repo with documentation for "3D-PV-Locator: Large-scale detection of rooftop-mounted photovoltaic systems in 3D" published in Applied Energy.

In case you would like to explore the code with which we created the image datasets and pre-processed the CityGML files, please have a look at the following GitHub repo.

3D-PV-Locator is a joint research initiative between Stanford University, University of Freiburg, and LMU Munich that aims at democratizing and accelerating the access to photovoltaic (PV) systems data in Germany and beyond.

To do so, we have developed a computer vision-based pipeline leveraging aerial imagery with a spatial resolution of 10 cm/pixel and 3D building data to automatically create address-level and rooftop-level PV registries for all counties within Germany's most populous state North Rhine-Westphalia.

For every address equipped with a PV system in North Rhine-Westphalia, the automatically produced address-level registry in GeoJSON-format specifies the respective PV system's:

• geometry: Real-world coordinate-referenced polygon describing the shape of the rooftop-mounted PV system
• area_inter: The total area covered by the PV system in square meters
• area_tilted: The total area covered by the PV system in square meters, corrected by the respective rooftop tilt
• capacity_not_tilted_area: The total PV capacity in kWp of area_inter
• capacity_titled_area: The total PV capacity in kWp of area_tilted
• location of street address in latitude and longitude
• street address
• city and
• ZIP code

For every rooftop equipped with a PV system in North Rhine-Westphalia, the automatically produced rooftop-level registry in GeoJSON-format specifies the respective PV system's:

• Azimuth: Orientation of the rooftop-mounted PV system, with 0° pointing to the North
• Tilt: Tilt of the rooftop-mounted PV system, with 0° being flat
• RoofTopID: Identifier of the respective rooftop
• geometry: Real-world coordinate-referenced polygon describing the shape of the rooftop-mounted PV system
• area_inter: The total area covered by the PV system in square meters
• area_tilted: The total area covered by the PV system in square meters, corrected by the respective rooftop tilt
• capacity_not_tilted_area: The total PV capacity in kWp of area_inter
• capacity_titled_area: The total PV capacity in kWp of area_tilted
• street address
• city and
• ZIP code

For a detailed description of the underlying pipeline and a case study for the city of Bottrop, please have a look at our spotlight talk at NeurIPS 2020:

• Paper
• Slides
• Recorded Talk

You might also want to take a look at other projects within Stanford's EnergyAtlas initiative:

• EnergyAtlas
• DeepSolar for Germany: Publication and Code

Please note that apart from the pipeline code and documentation, we also provide you with

• A pre-trained model checkpoint for PV classification on aerial imagery with a spatial resolution of 10cm/pixel.
• A pre-trained model checkpoint for PV segmentation on aerial imagery with a spatial resolution of 10cm/pixel.
• A 100,000+ image dataset for PV system classification.
• A 4,000+ image dataset for PV system segmentation.
• Pre-processed 3D building data in .GeoJSON format for the entire state of North Rhine-Westphalia.

In case you would like to explore the code with which we created the image datasets and pre-processed the CityGML files, please have a look at the following GitHub repo.

When using these resources, please cite our work as specified at the bottom of this page.

NOTE: All images and 3D building data is obtained from openNRW. Labeling of the images for PV system classification and segmentation has been conducted by us.

git clone https://github.com/kdmayer/3D-PV-Locator.git
cd 3D-PV-Locator

Download pre-trained classification and segmentation models for PV systems from our public AWS S3 bucket. This bucket is in "requester pays" mode, which means that you need to configure your AWS CLI before being able to download the files. Instructions on how to do it can be found here.

Once you have configured your AWS CLI with

aws configure

you can list and browse our public bucket with

aws s3 ls --request-payer requester s3://pv4ger/

Please download our pre-trained networks for PV system classification and segmentation by executing

aws s3 cp --request-payer requester s3://pv4ger/NRW_models/inceptionv3_weights.tar models/classification/
aws s3 cp --request-payer requester s3://pv4ger/NRW_models/deeplabv3_weights.tar models/segmentation/

To create PV registries for any county within North Rhine-Westphalia, you need to

1. Download the 3D building data for your desired county from our S3 bucket by executing and replacing <YOUR_DESIRED_COUNTY.geojson> with a county name from the list below:

    aws s3 cp --request-payer requester s3://pv4ger/NRW_rooftop_data/<YOUR_DESIRED_COUNTY.geojson> data/nrw_rooftop_data/
   

   Example for the county of Essen:

    aws s3 cp --request-payer requester s3://pv4ger/NRW_rooftop_data/Essen.geojson data/nrw_rooftop_data/
   

2. Specify the name of your desired county for analysis in the config.yml next to the "county4analysis" element by choosing one of the counties from the list below:

   Example:

    county4analysis: Essen
   

3. OPTIONAL STEP: Obtain your Bing API key for geocoding from here and paste it in the config.yml file next to the "bing_key" element

   Example:

    bing_key: <YOUR_BING_KEY>
   

   NOTE: If you leave <YOUR_BING_KEY> empty, geocoding will be done by the free OSM geocoding service.

Once the data and models are in place, we build and run the docker container with all required dependencies in interactive mode and mount the /data and /log directory in the container to our local machine. Mounting the /data and /log directories allows us to share the code outputs between the container and our local machine.

docker build -t 3d_pv_docker .
docker run -it -v <YOUR_ABSOLUTE_PATH_TO_THE_PROJECT_REPO>/data/:/app/data/ <YOUR_ABSOLUTE_PATH_TO_THE_PROJECT_REPO>/logs/:/app/logs/ 3d_pv_docker

Please ensure that <YOUR_ABSOLUTE_PATH_TO_THE_PROJECT_REPO> corresponds to your absolute path to the 3D-PV-Locator repo on your local machine, e.g., /Users/kevin/Projects/Active/3D-PV-Locator/ in my case.

Note: Depending on how many tiles you want to download, you will need to adjust the memory of your Docker container with the following flag for the docker run command:

--memory=<memory>

Having the docker container in interactive mode, we can now decide which pipeline steps we want to run by putting a "1" next them.

Example:

    run_tile_creator: 1

    run_tile_downloader: 1

    run_tile_processor: 1

    run_tile_coords_updater: 0

    run_registry_creator: 1

In the interactive Docker container, we then execute the pipeline with:

  python run_pipeline.py

After successful completion, the resulting PV registry for your area of interest will be written to /data/pv_registry.

Please choose the county you would like to run the pipeline for from the following list:

• Düren
• Essen
• Unna
• Mönchengladbach
• Solingen
• Dortmund
• Gütersloh
• Olpe
• Steinfurt
• Bottrop
• Coesfeld
• Leverkusen
• Köln
• Soest
• Mülheim-a.d.-Ruhr
• Münster
• Heinsberg
• Oberhausen
• Euskirchen
• Krefeld
• Warendorf
• Recklinghausen
• Bochum
• Rhein-Kreis-Neuss
• Rheinisch-Bergischer-Kreis
• Herne
• Kleve
• Bonn
• Minden-Lübbecke
• Herford
• Rhein-Sieg-Kreis
• Düsseldorf
• Hagen
• Paderborn
• Wuppertal
• Oberbergischer-Kreis
• Viersen
• Rhein-Erft-Kreis
• Märkischer-Kreis
• Städteregion-Aachen
• Remscheid
• Mettmann
• Lippe
• Ennepe-Ruhr-Kreis
• Hochsauerlandkreis
• Gelsenkirchen
• Höxter
• Borken
• Hamm
• Bielefeld
• Duisburg
• Siegen-Wittgenstein
• Wesel

For the German state of North Rhine-Westphalia (NRW), OpenNRW provides:

• Aerial imagery at a spatial resolution of 10cm/pixel
• Extensive 3D building data in CityGML format

MIT

Please cite our work as

@article{MAYER2022,
title = {3D-PV-Locator: Large-scale detection of rooftop-mounted photovoltaic systems in 3D},
journal = {Applied Energy},
volume = {310},
pages = {118469},
year = {2022},
issn = {0306-2619},
doi = {https://doi.org/10.1016/j.apenergy.2021.118469},
url = {https://www.sciencedirect.com/science/article/pii/S0306261921016937},
author = {Kevin Mayer and Benjamin Rausch and Marie-Louise Arlt and Gunther Gust and Zhecheng Wang and Dirk Neumann and Ram Rajagopal},
keywords = {Solar panels, Renewable energy, Image recognition, Deep learning, Computer vision, 3D building data, Remote sensing, Aerial imagery},
}