Generating informative and accurate descriptions of natural hazards and phenomena using large transformer-based models
Abstract
Earth observing satellites continuously monitor our planet's natural environment. They observe a wide variety of natural phenomena that can be occurring anywhere at any time - tropical cyclones, wildfires, volcanic eruptions, floods, phytoplankton blooms, dust storms, and icebergs, to name a few. These events can be seen by means of browsing remote sensing imagery captured from these satellites and, of course, many types of events are captured and recorded by national and regional science organizations, e.g., the U.S. National Weather Service. However, these two mechanisms for observing natural events are not conducive to being consumed by the wider public or general popular media. To meet the needs of these wider audiences, a narrative reporting approach would provide an easier-to-consume package for users. For this research, we have pulled together remote sensing imagery, natural event metadata, and other ancillary descriptive material to construct a brief narrative describing the natural events that are shown in the imagery. Our data comes from the NASA Earth Observatory, Planet Labs Gallery, AGU publications, and other sources. We present promising preliminary results and an early view of our pipeline, which leverages techniques such as text synthesis based on summarization, Named Entity Recognition (NER), and our training of large transformer-based models. Based on our results we encourage the community to use our models and datasets for automatic text generation to produce informative as well as accurate descriptions of natural hazards and phenomena. We hope that our work will contribute to rapid progress in the automated text generation domain.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMIN31A..06T