Precipitation Retrieval from Fengyun-3D Microwave Humidity and Temperature Sounder Data Using Machine Learning

Kangwen Liu,Jieying He,Haonan Chen

doi:10.3390/rs14040848

Abstract

As an important component of the Earth system, precipitation plays a vital role in regional and global water cycles. Based on Microwave Humidity and Temperature Sounder (MWHTS) onboard FY-3D satellite, four machine learning models, random forest regression (RFR), support vector machine (SVM), multilayer perceptron (MLP), and gradient boosting regression tree (GBRT), are implemented to retrieve precipitation rate, and verified with Integrated Multi-satellite Retrievals for GPM (IMERG). This paper determines the optimal hyperparameters of the machine models and proposes three linear combinations of MWHTS channels (183.31 ± 1.0–183.31 ± 3.0 GHz, 183.31 ± 1.0–183.31 ± 7.0 GHz, and 183.31 ± 3.0–183.31 ± 7.0 GHz), which can better characterize precipitation of different intensities. With the inclusion of three linear combinations, the performances of all four machine learning models are significantly improved. It is concluded that the RFR and GBRT have the best retrieval accuracy. Over ocean, the MSE, MAE, and R2 values of precipitation estimates using RFR are 1.75 mm/h, 0.44 mm/h, and 0.80, respectively, and are 1.80 mm/h, 0.45 mm/h, and 0.78 for GBRT. Simultaneously, this paper analyzes the retrieval results from the perspective of the different rain rates and temporal matching difference between MWHTS and IMERG data. The RFR and GBRT also maintain the best retrieval accuracy under the condition of Gaussian noise, indicating the relatively strong robustness and antinoise performance of ensemble learning models for precipitation retrieval.

Highlights

Precipitation is of great significance in various fields of meteorology and hydrology, such as regional and global water resources, climate change, and numerical weather modeling research [1,2,3]
The results indicate that the use of the dynamic emissivity retrieved from the 89 GHz channel of Microwave Humidity and Temperature Sounder (MWHTS)/FY-3C apparently increases the amount of assimilated data and improves the initial fields and the 24 h forecasts of precipitation distribution and intensity
At different MWHTS channels, three linear TB combinations are proposed as inputs for precipitation retrievals, which can better characterize precipitation at different intensities and further improve the retrieval accuracy; (2) Using grid search and cross-validation methods to explore the optimal hyperparameters of machine models, this paper compares the performance and explores the feasibility and rationality of four machine learning models (RFR, support vector machine (SVM), multilayer perceptron (MLP), and gradient boosting regression tree (GBRT)) for precipitation retrieval; (3) This article quantifies the retrieval advantages with the addition of linear combinations as inputs and verifies the robustness of the ensemble learning models

Summary

Introduction

Precipitation is of great significance in various fields of meteorology and hydrology, such as regional and global water resources, climate change, and numerical weather modeling research [1,2,3]. The application of satellite observations is an vital method to obtain precipitation information [1,4,5,6,7]. The microwave wavelength can be flexibly selected according to practical applications, and the influence of ice clouds and other particles can be ignored or effectively utilized. Since the operation of satellite series Fengyun-3 (FY-3, including FY-3A, 3B, 3C, 3D, 3E), the satellites have obtained rich data for weather, climate, and environmental research [8]

Results

Discussion

Conclusion