3.2 Combining the Power of Crowdsourcing and Machine Learning to Explore Cloud Classification

Tuesday, 24 January 2017: 8:45 AM
310 (Washington State Convention Center )
Tianle Yuan, GSFC, Greenbelt, MD; and L. Oreopoulos, K. Meyer, S. Platnick, and H. Wang

Marine low clouds are the largest source of uncertainty for future climate change projections. Climate change is expected to affect everyone in the US. Better understanding of these clouds will reduce the uncertainty. These clouds are results of a complex set of intricate processes and they display rich behaviors. In particular, they organize into different forms, or different types of cells, which have distinct impacts on the Earth's energy balance. NASA satellites have amassed vast amount of imagery for these clouds over the globe. Automatic computer algorithms to classify clouds into different types of cells still do not exist, which hinders our ability to understand these clouds. Humans, however, are fully capable of this task after relatively easy training. Here we present preliminary results of combining machine learning algorithm with crowdsourcing techniques to explore low cloud classification using satellite images. We attempt to achieve such classifications without feature engineering. Instead, we let the algorithm directly take full images and classify them based on its internal learning mechanism. Susccessful development of such algorithms will enhance our ability to analyze and understand low clouds and its climate feedback.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner