The increase in Earth science observation instruments and platforms throughout the world has resulted in an exponential rate of data growth. This presents the Earth science community with a rich and ever expanding archive of useful data. Machine learning (ML) algorithms offer the potential for innovative new data analysis and the generation of valuable insights from this massive archive of Earth observation data. However, despite the popularity of applying machine learning to problems in other fields, there is only limited adoption of machine learning in the Earth science community. This lack of machine learning utilization in the Earth sciences has largely been credited to a lack of labeled training data.

The Machine Learning project encompasses the following key focus areas:

Building tools, pipelines and labeled datasets

The application of ML algorithms to Earth science phenomena entails the development of ML tools and pipelines. These tools also require an archive of labeled data for scientists to leverage in building ML applications. IMPACT’s ML team develops needed resources and promotes an open approach to machine learning in the Earth sciences.

Adoption of machine learning algorithms using Earth science data

The ML team works to encourage the use of applied artificial intelligence and ML to answer questions in Earth science. Members of the team advertise and support the usability of ML in NASA’s Earth Science division. To accomplish this goal, the team incorporates machine learning into different stages of the data lifecycle to improve functionality and operations.

Explore avenues to support machine learning efforts in ESDS

The ML team tracks new developments in the areas of artificial intelligence and ML to contribute to advancing NASA’s Earth Science Data Systems (ESDS). As part of this focus, the team applies ML perspectives to analytics architecture designs. The team also strives to utilize the full capacity of machine learning advances in high-end computing and cloud computing in order to achieve NASA Science’s research goals. Along the way, the team collaborates with academia, industry and other government agencies.

Tools and Portals

Phenomena Portal

Phenomena portal is an online platform for visualizing the detection of multiple phenomena by machine learning models.

Hurricane Portal

Hurricane Portal is an online platform that demonstrates the use of a machine learning model for estimating the wind speed of hurricanes

Image Labeler

Image Labeler is a web-based tool used to facilitate the labeling of Earth science images for use in training machine learning models.

Presentations and Publications

• “Employing Deep Learning to Enable Visual Exploration of Earth Science Events”
• “ES2Vec: Earth Science Metadata Keyword Assignment using Domain-Specific Word Embeddings”
• “Machine Learning Lifecycle for Earth Science Application: A Practical Insight into Production Deployment”
• “Pixel level smoke detection model with deep neural network”
• “Surveying the Machine Learning Landscape in Earth Sciences”
• “Image Labeler: A Web Interface to Catalog Earth Science Events”
• “Satellite Smoke Detection Using Pixel-Based Machine Learning”
• “Pixel Based Model For High Latitude Dust Detection”
• “Deep Learning-based Tropical Cyclone Intensity Estimation Portal”
• “End-to-end Machine Learning Applications Framework for Earth Science”
• “ES2Vec: Earth Science Metadata Keyword Assignment using Domain-Specific Word Embeddings”