This project was done as part of a bachelor's thesis at HEIG-VD. It is a collection of notebooks that were used to explore different types of geographical data. They also enable the user to train a convolutional neural network that predicts the population density of a region with a high level of detail (250-1000m) from nightime satellite imagery.
- python 3.7
On Windows, some scientific packages have some compatibility issues. Those are shapely, gdal, fiona, and rasterio.
Download them from Christoph Gohlke's website and manually install them
(use pip install packagename.whl).
Then, on both Windows and Linux, use the command line pip install -r requirements.txt from the project's root folder.
(optional) In order to enable training on the GPU, follow the TensorFlow tutorial of your platform under "Requirements to run TensorFlow with GPU support".
Use the command python -m notebook to browse notebooks on Windows, or jupyter notebook on Linux.
Each notebook can be found in the folder with the notebook's name.
This set of scripts computes the sum of light perceived from space for each country, and compares these values with other miscellaneous per country data. This was only useful for data exploration.
This notebook preprocesses a raster with two bands : nighttime satellite picture and population density raster. It is then used to train a convolutional neural network that makes high resolution predictions for population density from another satellite picture.
This notebook compares two rasters to either compile a scatterplots of the values of each pixel, or compile the difference between two rasters in order to visualize the evolution from one year to the next for example.
This notebook downloads and assembles tiles from NASA's GIBS API. There really is no point in using this now.
This notebook answers to the specific need to extract the huge quantity of nighttime data from the NOAA and compile yearly rasters from monthly rasters.