Two Tasks Are Better Than One: Learning If and How Much It’s Raining Improves Deep Models

Research

View list of research

坂内さんEN（1200_800）.jpg

To measure precipitation from satellite data, researchers from the Institute of Industrial Science, The University of Tokyo, developed a deep-learning-based precipitation retrieval framework. Rather than identifying rain areas through classification and then determining the rate of rain through regression, multi-task learning was used to perform classification and regression simultaneously, improving performance by sharing knowledge across tasks. Compared with existing methods, this approach yielded comparable performance for rain classification and improved inferences of precipitation intensity.

2023.04.06

Press Release

Two Tasks Are Better Than One: Learning If and How Much It’s Raining Improves Deep Models

Researchers from the Institute of Industrial Science, The University of Tokyo, find that a deep-learning model that simultaneously learns to classify satellite data as rain/no-rain while determining rain intensity outperforms other models.