Deep Learning with Geospatial Data | SciPy 2017 | Shane Grigsby

Deep learning is all the rage these days, but what do you do when your data isn’t handwritten digits, or pictures of cats? Geospatial data comes with it’s own unique challenges—huge high dimensional datasets in weird file formats, irregular and often mismatched grids, and a pervasive lack of labeled training data… to name just a few! In this talk, we’ll explore cutting up and resampling giant remote sensing rasters using modern python tools like rasterio, georasters, and GDAL; detrending, extracting, and storing information with pandas; sensible dataset and dimensionality reduction through scipy transforms; and a few places to go in Keras and other deep learning libraries. Come learn how to jump from satellite or airborne data to your own “geoMINST” database that’s ingestible to your favorite deep learning technique!

[ramav]

This is really nice work - is it published? Also, did you think about autoencoders for reducing dimensionality, so that it can become a regression problem?