1. Install (see instructions below)
2. Run the jupyter notebooks in order (optionally, you can read the blogposts at the same time)

Directories:

• train: train a model
• compute: compute CNN features by batches
• predict: use the model on CNN features

Common modules :

• cnn.py: use a custom VGG16 CNN
• download.py: download building images from mapbox
• geo.py: geographic and cartographic functions
• load.py: load data, download if not already cached

First, clone this repository.

Be sure you use python3.

Install GDAL version 2.x. If your favorite package manager does not provide version 2, you can download it from http://download.osgeo.org/gdal/, and then :

./configure
make
make install

Test the installation with gdal-config --version.

Install PROJ4 :

apt-get install libproj-dev

If pip install GDAL fails, retry after

export CPLUS_INCLUDE_PATH=/usr/include/gdal
export C_INCLUDE_PATH=/usr/include/gdal

If the package osgeo cannot be imported from python, retry after installing libgeos-dev libwebp-dev unixodbc-dev libxerces-c-dev libjasper-dev libnetcdf-dev libhdf4-alt-dev libgif-dev.

Install postgresql >= 9.6 and postgis.

If your favorite package manager does not provide a recent version, follow the instructions in https://www.postgresql.org/download/.

Install also libpq-dev.

Create a database. Here are a few tips :

• to perform administrative tasks on the database, log in as the postgres user su postgres
• to create a database user: createuser --interactive
• to create the database solar: createdb solar
• to edit the access policies : nano /etc/postgres/9.6/main/pg_hba.conf (add a line local solar solar md5) then reload systemctl reload postgresql
• in the psql command prompt: (use \c to connect, \l to list databases and \d to list tables)

grant all privileges on database solar to solar;
alter user solar with password '***';
\c solar
create extension postgis;

• test the connection: psql solar solar
• set up a tunnel if your database is not on the same system than the one which runs your python code:

ssh -M -S ~/my-ctrl-socket-5432 -fnNT -L 5432:localhost:5432 username@database-machine
ssh -S ~/my-ctrl-socket-5432 -O exit username@database-machine

You can install tensorflow from pip and it is the fastest and easiest option.

If you plan to compute the CNN features using a CPU, the packaged version of tensorflow may not take advantage of advanced instruction sets supported by your CPU. In that case tensorflow can be built from source. Tested on CPUs supporting SSE4.1, SSE4.2, AVX, AVX2, FMA (Intel Xeon E5-2643 v3 @ 3.40GHz), the optimized version of tensorflow achieves a 2x speedup compared to the packaged version.

Tensorflow works best on GPUs. Ran on a nVidia GeForce GTX 1060 6Go, a mid to high-end graphic card launched in 2016, tensorflow is 10x faster than on 2 Intel Xeon E5-2643 v3 with the optimiser version. It applies VGG16 on 35.000 images of dimension 96x96 per second. Graphic cards are by far the cheapest and the most efficient option to perform deep learning on large amounts of data.

Make an editable installation.

pip install -e .

You need access to a WMS server to download iages of roofs. If you are not familiar with WMS, see this issue to get started: opensolarmap#4

Copy config.ini.example to config.ini and adapt it.

You will need a file buildings_labelled.csv and a file contribs.csv that you can find in the directory data.

Geographic coordinates are a means to label locations on the earth with numbers. The most used geographic coordinates system nowadays is WGS84. It approximates the surface of the earth as a speroid, and defines a 2-dimensional coordinates system on that theoretical spheroid, composed of a latitude (lat, theta) and a longitude (lon, long, lambda). Although complex, it is possible to compute distances and surfaces using WGS84.

Cartography is the art of representing the surface of the earth on planes, using a "projection". Such a projection cannot be satisfactory as it necessarily violates either angles, distances, surfaces or bearings. Among the huge number of invented projections, Web Mercator is commonly used. Lambert93 is also commonly used in France. Both Web Mercator and Lambert93 use a 2-dimensional planar system whose axis are called X and Y. Distances and surfaces are fast and easy to compute using Web Mercator or Lambert93 but they are very imprecise.

The Geodetic Parameter Registry assigns a identifier to each of these systems:

• WGS84 : 4326
• Web Mercator : 3857
• Lambert93 : 2154