A new image segmentation model for precipitation estimation using meteorological satellite infrared images and geographic information

Abstract

Satellite products are expected to play important roles in water-related management and public welfare, particularly in developing countries. Higher-resolution precipitation products are required to cope with increasingly severe water-related disasters. In this study, we propose a new satellite precipitation estimation algorithm based on deep learning that uses data from multiple satellite infrared (IR) bands and geographic information (e.g. elevation, latitude, and longitude) as input. For the deep learning model component, we use various image segmentation models, including U-Net, PSPNet, and DeepLabv3+. Cosine similarity and correlation coefficients for precipitation rate were used to determine the IR bands of the input data; five bands were used as IR. Four input datasets were constructed: IR alone; IR and elevation data; IR and latitude/longitude; and IR, elevation data, and latitude/longitude. When PSPNet or DeepLabv3+ was used as the deep learning model, and elevation and latitude/longitude were added to IR as input data, the mean square error and fraction skill score showed improved accuracy over GSMaP_MVKv7 and PERSIANN-CCS; precipitation overestimation was eliminated. These results indicate that deep learning models can be used to estimate precipitation from satellite IR observations with high resolution and accuracy.