The following people contributed equally to this repository (in alphabetical order):

• Daniel Bumke
• JJX
• Corstiaen Versteegh

A presentation of the project is available here.

This repository is forked on https://github.com/jjx3455/solar_berlin and https://github.com/cverst/solar_berlin to reflect the work of all contributors.

Estimating the potential photovoltaic production of Berlin rooftops

Prerequisites

• Install Python 3.7 (other versions may work) and the packages listed in requirements.txt
• To run the unet model effectively, gpu-accelertion in tensorflow is advisable - https://www.tensorflow.org/install/gpu
• To run the Mask R-CNN model additionally install the packages listed in maskr/requirements.txt and maskr/requirements-detectron.txt - gpu-acceleration for pytorch is advisable
• It may be necessary to install gdal into your operating system and pin gdal in the requirements.txt to the same version as your system (see https://pypi.org/project/GDAL/)
• In order to convert the original raster tiles from ecw to GeoTIFF GDAL must be built with ecw support (see https://trac.osgeo.org/gdal/wiki/ECW)
  • Alternatively, skip downloading the tiles and use the pre-processed GeoTIFFs available as converted.zip (72GB)
  • Alternatively 2, skip the extraction too and use the pre-extracted sample data available as extracted.zip (45GB)

Getting the original data

To download and convert the original data:

• Create a subfolder "data" in the repository
• If not using the pre-processed GeoTIFFs, download the original data:
  • Download map and vector files: cd conversion && sh download.sh
  • Convert from ecw to GeoTIFF: sh convert_rasters.sh
  • Clip the vector map to the same size as the raster tiles: python clip_shape.py

Extracting the data for the model

To extract the data into usable map/mask tiles and select a sample for the model:

• Adjust parameters in select_data.py for desired sample size and tile size
• Run python select_data.py
• (This will take a while as it extracts all tiles of that size from the original data. The tiles are saved in the data/extracted folder so sampling again from the same tiles will be much faster.)

Running the current models

Two models can be run on the available data.

The first is a U-Net model with a pre-trained MobileNetV2 as the backbone for semantic segementation of rooftops (pixel-by-pixel prediction of roof/no-roof).

To run the unet model:

• If not using the pre-extracted sample tiles:
  • modify the parameters in run_model.py to point at a different data selection folder
  • Run python run_model.py
  • Inspect results in logs The best model is unet/unet_resnet101v2.py

The second is a Detectron2 Mask R-CNN model with a pre-trained Resnet50 backbone for instance segmentation of roofs along with their PV category.

To run the model (Mask R-CNN):

• Run the model:
  • modify the parameters in maskr/run_mask_r_cnn.py to point at a different data selection folder
  • Run python maskr/run_mask_r_cnn.py
• Generate predictions:
  • modify the parameters in predict_from_mask_r_cnn.py
  • Run python predict_from_mask_r_cnn.py

Running Tests

Data for testing is managed with DVC (https://dvc.org/). To run the tests:

• Download the test data dvc pull
• Run python run_tests.py