The 2019 ESIP Winter Meeting has passed. See session descriptions to access meeting content, including presentations, recordings, and key takeaways. See here for info on upcoming meetings.
Session Abstract: The Machine Learning Cluster was formed following the ESIP 2018 Summer Meeting. Since then, members have been meeting via telecon to establish areas of focus from across a broad range of possibilities. Overall, we seek to serve the ESIP Community in its efforts to apply machine learning. We continue to seek community input on achieving that, and hope to hear more as a result of this session.
This session will start with an introduction to the cluster, a brief discussion regarding options for areas of focus, and also the results to date of a community survey. People interested in the topic will be invited to take the survey and/or join the cluster.
Following this, in the spirit of learning what ESIP members are doing in this arena, we hear from three cluster members involved in machine learning. The first presentation describes the application of machine learning techniques to a specific science problem, and the following two presentations describe tools for machine learning. The following are abbreviated presentation abstracts:
Estimating Daily Surface Air Temperature Using Satellite Land Surface Temperature and Top-of-Atmosphere Radiation Products over the Tibetan Plateau Yuhan Rao, Department of Geographical Sciences, University of Maryland In this study, we developed a new method based on the rule-based Cubist regression model to estimate daily surface air temperature under all-sky conditions with temporally gap-filled LST, incident solar radiation at the surface, top-of-atmosphere (TOA) albedo and outgoing longwave radiation products as the input. Two different model strategies are compared in this study: 1) building a unified all-sky model or 2) building individual models for clear-sky and cloudy-sky separately.
Marvin: An Automated Machine Learning as a Service for Primitive Annotation and Execution Brian Wilson, Sujen Shah, Chris Mattmann, Ryan McGranaghan, Giuseppe Totaro, Mark Hoffmann, Vishal Lall, Alice Yepremyan, Srinidhi Nandakumar, Wayne Burke, Paul Ramirez, Jet Propulsion Laboratory A NASA team at JPL has been participating in the DARPA Data Driven Discovery of Models (D3M) program, whose goal is to create a system that can automate the composition of ML Pipelines to simultaneously solve a 1000 ML problems, ranging from classification or regression on a simple “feature matrix” to image recognition, audio segmentation, EKG/EEG classification, video processing, graph clustering, time-series forecasting, etc.
Raster Vision: Deep Learning for Aerial and Satellite Imagery Rob Emanuele, Azavea Raster Vision is an open source framework for doing deep learning on satellite, aerial and other large raster data. It allows engineers performing deep learning on raster data a way to quickly and repeatably prepare training data, train models, make predictions, and evaluate models. It works with local GPU machines as well as AWS services.
Session Takeaways (post-meeting): 1) Bias can easily be built into these ML models. Therefore, you need to have domain experts or somebody who knows something about the ingoing data to deal with bias or problems with input data. 2) There are a lot of questions about what the ML cluster should focus on and how much depth vs breadth the cluster should address in such a broad field 3) Its easy to get lost in the proverbial ML weeds so either you need expertise in the ML side of things (data/computer science) and the earth science applications or bring in a few people with different expertise.