AI for Good

Chapter 4Deep Learning with Geospatial Data

—Caleb Robinson, Anthony Ortiz, Simone Fobi, Amrita Gupta, Girmaw Abebe Tadesse, Akram Zaytar, and Gilles Hacheme

Executive Summary

The availability of satellite data has radically transformed how we can monitor and influence numerous activities on Earth, including agriculture, urban planning, disaster monitoring and response, and climate change research. However, different satellites have different data collection methodologies and resolutions, which makes the application of deep learning methodologies challenging. Here, to realize the potential of deep learning for remote sensing applications like satellite imagery, we developed a library of processing code called TorchGeo that provides ways to load data from a variety of benchmark datasets, allows for sampling of geospatial data, and provides transformers that can be used with multispectral imagery—that is, imagery with channels beyond red, green, and blue—which can be used to reveal the presence of gas, biological activity, water quality, and pollution, for instance. TorchGeo is the first library that released pre-trained models for multispectral satellite imagery that can be used in downstream sensing tasks when there is limited labeled data. To promote further research, we used existing datasets and benchmarks to create reproducible benchmark results that can be used to accelerate satellite imagery–based research efforts.

Why Is This Important?

While human activity is contributing ...