Performance evaluation, error decomposition and Tree-based Machine Learning error correction of GPM IMERG and TRMM 3B42 products in the Three Gorges Reservoir Area

Abstract

Satellite precipitation products (SPPs) have excellent development prospects in hydrometeorology research for their high spatial-temporal resolution. Understanding the performance of SPPs is essential before their application in estimating multiple processes. This paper evaluated the accuracy and error composition of GPM IMERG Final-run V06 and TRMM 3B42V7 in the Three Gorges Reservoir Area (TGRA) based on the China Gauge-based Daily Precipitation Analysis dataset (CGDPA) during 2001–2019. Afterwards, the precipitation correction by merging satellite and gauge observation precipitation was conducted with three Tree-based Machine Learning models, Decision Tree (DT), Adaptive Boosting Decision Trees (Adaboost), and Random Forest (RF). The results show that: (1) Two SPPs can well reflect the spatial pattern and in-year variation of precipitation in the TGRA. However, they are limited in estimating precipitation amount for seriously underestimated the light rain or overestimated torrential rain. Under the influence of climate and variety terrains, the SPPs accuracy in the middle-elevation relief mountains of downstream is significantly higher than that in the low-elevation relief mountains of upstream. Moreover, that on the monthly and annual scales are higher than on the daily scale. Comparing two SPPs, the area proportions of IMERG-F better than 3B42V7 are 64.0%, 86.0%, and 81.6% on daily, monthly and annual scales, respectively. (2) Both the systematic error components of two SPPs raise with the increase of precipitation amount, and the error spatial pattern of 3B42V7 is more affected by precipitation. The random error component of light rain and daily precipitation is higher than other grades and scales. Despite this, the random error component decreases significantly when daily precipitation is accumulated. (3) The three primary factors affecting error correction of SPPs are precipitation amount, longitude, and DEM in TGRA. Among the three Tree-based methods, the RF showed the best error correction effect, and it can significantly improve the monthly and annual precipitation accuracy and better reproduce probability density function (PDF). Cumulative precipitation can reduce the random error component of daily precipitation and have high correction accuracy when cumulative time exceeds 5-days.