Downscaling Satellite and Reanalysis Precipitation Products Using Attention-Based Deep Convolutional Neural Nets

Alexander Y Sun,Guoqiang Tang

doi:10.3389/frwa.2020.536743

Abstract

High-quality and high-resolution precipitation products are critically important to many hydrological applications. Advances in satellite remote sensing instruments and data retrieval algorithms continue to improve the quality of the operational precipitation products. However, most satellite products existing today are still too coarse to be ingested for local water management and planning purposes. Recent advances in deep learning algorithms enable the fusion of multi-source, high-dimensional data for statistical learning. In this study, we investigated the efficacy of an attention-based, deep convolutional neural network (AU-Net) for learning spatial and temporal mappings from coarse-resolution to fine-resolution precipitation products. The skills of AU-Net models, developed using combinations of static and dynamic predictors, were evaluated over a 3 × 3° study area in Central Texas, U.S., a region known for its complex precipitation patterns and low predictability. Three coarse-resolution satellite/reanalysis precipitation products, ERA5-Land (0.1°), TRMM (0.25°), and IMERG (0.1°), are used as part of the inputs, while the predictand is the 1-km PRISM data. Auxiliary predictors include elevation, vegetation index, and air temperature. The study period includes 18 years of data (2001–2018) at the monthly scale for training, validation, and testing. Results show that the trained AU-Net models achieve different degrees of success in downscaling the baseline coarse-resolution products, depending on the total precipitation, the accuracy of large-scale patterns captured by the baseline products, and the amount of information transferable from predictors. Higher precipitation rate tends to affect AU-Net model performance negatively. Use of the attention mechanism in the AU-Net models allows for infilling of multiscale features and generation of sharper images. Correction using gauge data, if there is any, can further improve the results significantly.

Highlights

Precipitation is a primary driver of water and energy cycle (Trenberth et al, 2007), providing essential inputs to many water, food, and energy applications including, but not limited to, global and regional climate variability assessments, land surface-atmosphere interactions, natural hazard prevention, crop yield management, hydrological forecasting, and surface and groundwaterPrecipitation Downscaling Using Attention-Based U-Net resources planning (Hong et al, 2007; Seneviratne et al, 2010; Becker et al, 2013; Schewe et al, 2014)
For each coarse-resolution precipitation product (i.e., ERA5Land, Tropical Rainfall Measuring Mission (TRMM), and Integrated Multi-satellite Retrievals for GPM (IMERG)), the performance of four groups of predictors (M1 − M4) that are defined under section 3.3 were evaluated, leading to a total of 12 different AU-Net models
The correlation maps of both TRMM and Global Precipitation Measurement (GPM) exhibit similar spatial patterns, which tend to be higher in the eastern part and lower in the northwest high-elevation areas; all correlation values are above 0.8 (Note: despite the similarity, TRMM and IMERG processing are different in a number of ways, e.g., the Level2 and Level-3 algorithms, the infrared data used for gap filling and replacement, as well as the spatiotemporal resolutions)

Summary

Introduction

Precipitation is a primary driver of water and energy cycle (Trenberth et al, 2007), providing essential inputs to many water, food, and energy applications including, but not limited to, global and regional climate variability assessments, land surface-atmosphere interactions, natural hazard prevention, crop yield management, hydrological forecasting, and surface and groundwaterPrecipitation Downscaling Using Attention-Based U-Net resources planning (Hong et al, 2007; Seneviratne et al, 2010; Becker et al, 2013; Schewe et al, 2014). Available precipitation products may be classified into ground-based, satellite-based, reanalysis, and hybrid multisource/multi-sensor products. Ground-based products are derived from rain gauges and weather radar. The spatial coverage of rain gauge networks is often limited, varying significantly across different countries owing to temporal sampling resolutions, periods of operation, data latency, and data access (Kidd et al, 2017). High-resolution gridded products are only available in a few developed counties that have extensive gauge network coverage. In the U.S, the Parameter-elevation Regressions on Independent Slopes Model (PRISM) gauge-based product (4-km resolution, 1895– present), developed by the Oregon State University (Daly et al, 1997), is widely used for operational planning and validation of satellite products. The Stage IV radar-based, gaugeadjusted precipitation data (4 km, 2002–present) available from the National Center for Environmental Prediction (NCEP) is commonly used as a reference dataset in many conterminous U.S (CONUS) precipitation product comparisons (Lin and Mitchell, 2005)

Methods

Results

Discussion

Conclusion