LANDSAT-landstats 1.1

You can use the files in this repository to predict population from satellite images.

LANDSAT lanstats is a supervised learning model to predict socio-economic characteristics from satellite data. You can easily modify this code to predict other socio-economic characteristics or use other satellite images. In addition to the convnet, I've provide files to construct the data.

The model works pretty well. The current version results in a test R² of 0.74. Version 1.0 resulted in a test set R² of 0.26. The main difference is the improvement in satellite image resolution from 1000 meters per pixel to 180 meters per pixel.

I ran the convnet on AWS's G2 machine and constructed the database on AWS's M10 machine. Get in touch if you'd like a link to the AMI. I also tried this on a mid-2009 MacBookPro and can recommend you don't.

The files of interest are:

• cnn.py: runs the convolutional neural network
• data_cleaning.py: merges satellite images with population databases (shapefiles)
• do.py: a testing file for data_cleaning.py
• ee_data.py: downloads the satellite images

I downloaded population data from the US census bureau and the Australian Bureau of Statistics. Provided you have .tif satellite images and shapefiles for your socio-economic data, it should be easy to tweak the files. I am happy to share any constructed data and the model weights.

In addition to the usual, the analysis relies on these packages

• Keras
• Google Earth Engine
• Shapely
• GeoPandas
• Rtree
• GDAL
• pyproj
• parmap
• h5py