Machine Learning-Based Bias Correction of Precipitation Measurements at High Altitude

• Published in: Remote sensing (Basel, Switzerland) Switzerland), 2023-04, Vol.15 (8), p.2180
• Main Authors: Li, Hongyi, Zhang, Yang, Lei, Huajin, Hao, Xiaohua
• Format: Article
• Language: English
• Subjects: Algorithms; Altitude; Analysis; Bias; bias correction of precipitation measurement; Datasets; Environmental hazards; Frequency distribution; Gauges; High altitude; High-altitude environments; Hydrology; Learning algorithms; Machine learning; Methods; Precipitation; Precipitation (Meteorology); Remote observing; Remote sensing; remotely sensed precipitation; Sensors; Statistical analysis; Statistical methods; Statistics; Tibetan plateau; Wind speed; XGBoost
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs15082180

Accurate precipitation measurements are essential for understanding hydrological processes in high-altitude regions. Conventional gauge measurements often yield large underestimations of actual precipitation, prompting the development of statistical methods to correct the measurement bias. However, the complex conditions at high altitudes pose additional challenges to the statistical methods. To improve the correction of precipitation measurements in high-altitude areas, we selected the Yakou station, situated at an altitude of 4147 m on the Tibetan plateau, as the study site. In this study, we employed the machine learning method XGBoost regression to correct precipitation measurements using meteorological variables and remote sensing data, including Global Satellite Mapping of Precipitation (GSMaP), Integrated Multi-satellitE Retrievals for GPM (IMERG) and Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS). Additionally, we examined the transferability of this method between different stations in our study site, Norway, and the United States. Our results show that the Yakou station experiences a large underestimation of precipitation, with a magnitude of 51.4%. This is significantly higher than similar measurements taken in the Arctic or lower altitudes. Furthermore, the remote sensing precipitation datasets underestimated precipitation when compared to the Double Fence Intercomparison Reference (DFIR) precipitation observation. Our findings suggest that the machine learning method outperformed the traditional statistical method in accuracy metrics and frequency distribution. Introducing remote sensing data, especially the GSMaP precipitation, could potentially replace the role of in situ wind speed in precipitation correction, highlighting the potential of remote sensing data for correcting precipitation rather than in situ meteorological observation. Moreover, our results indicate that the machine learning method with remote sensing data demonstrated better transferability than the traditional statistical method when we cross-validated the method with sites located in different countries. This study offers a promising strategy for obtaining more accurate precipitation measurements in high-altitude regions.