Correcting Monthly TRMM 3B43 Precipitation with Quantile Regression Model in the Urmia Lake Basin (2001-2019)

Extended Abstract



Introduction

The enhancement of rainfall data precision is vital for hydrological modeling, water resources management, and drought monitoring. Traditional rain gauge station data often fall short of capturing the spatial variations of precipitation. In contrast, remote sensing, such as the TRMM 3B43, can provide more accurate spatial distribution of rainfall, especially in regions lacking rain gauge stations. The TRMM 3B43 algorithm is designed to offer the most accurate rainfall estimates by leveraging a combination of multi-sensor measurements gathered from various satellites. This study focuses on the application of quantile regression models to correct TRMM 3B43 precipitation in the Urmia Lake basin for the period 2001-2019. The Urmia Lake basin, characterized by its ecological significance and susceptibility to climate variability, serves as a critical focal point for this study. By refining TRMM 3B43 data specifically in this region, our aim is to contribute valuable insights for more informed water resource planning and management. This research not only addresses the limitations of traditional rain gauge stations but also underscores the importance of tailored approaches for regions with unique hydroclimatic conditions.



Methods

Monthly precipitation data from TRMM 3B43 was calibrated using quantile regression models, along with Ordinary Least Squares regression, based on data from 12 synoptic stations in the Urmia Lake basin. The calibration process utilized 70% of the data for the period 2001-2019 for model training, with the remaining 30% reserved for validation. Quantile regression was applied across various quantiles (τ = 0.05 to τ = 0.99) to capture a comprehensive range of conditions. The resulting regression equations were evaluated based on their ability to correct TRMM monthly precipitation data, and the accuracy was assessed using established metrics such as RMSE, MAE, PBIAS, and KGE.



Results and Discussion

The scatter plots comparing monthly observations and TRMM precipitation data revealed a consistent overestimation by the TRMM satellite across all months. The conducted studies confirmed this overestimation pattern across various time scales. Attempts to correct the TRMM data using Linear Regression proved inadequate, as the corrected data still exhibited both underestimation and overestimation. However, the application of quantile regression in different quantiles successfully corrected TRMM 3B43 data. The results showed that the TRMM satellite data were almost 100 percent similar to the observed data in all months. The Root Mean Squared Error (RMSE) was highest in January, December, February, March, April, November, and May, with values of 22.64, 20.76, 20.04, 17.99, 17.35, 15.07, and 13.67, respectively. After correction by quantile regression, the RMSE values dropped to less than 2.35 in all months. Furthermore, the Mean Absolute Error (MAE) significantly decreased after quantile regression correction, showcasing a closer alignment between TRMM data and observed rainfall. PBIAS indicated that TRMM overestimated rainfall in all months from January to December by varying percentages. However, after quantile regression correction, the PBIAS values were generally less than 2.10 in all the studied months. The Kling-Gupta efficiency (KGE) values demonstrated improvement in all months when using the quantile regression method compared to the original TRMM data and the linear regression method. KGE values above 0.97 indicated the superior performance of the corrected TRMM data using the quantile regression model. In the calibration stage, 30% of randomly selected data were corrected using quantile regression equations obtained from the remaining 70% of the data. This process resulted in almost all points aligning closely with the ideal line of 1:1 after applying quantile regression. Comparison of monthly rainfall data across all studied stations showed a similar trend between TRMM precipitation data and observed rainfall, with TRMM often estimating higher values than the actual observed rainfall. Notably, in stations like Sarab, Salmas, and Sahand, the TRMM satellite data initially differed significantly from the observed data. However, after calibration with quantile regression, the satellite rainfall data closely matched the observed data. This underscores the effectiveness of the quantile regression method in correcting TRMM data, making it suitable for various studies, including water resources management and drought monitoring.



Conclusion

In conclusion, the application of quantile regression effectively corrected TRMM 3B43 precipitation data, addressing persistent overestimation issues observed in the Urmia Lake basin from 2001 to 2019. The method significantly improved accuracy, demonstrated by reduced RMSE and MAE values, enhanced PBIAS alignment, and superior KGE performance. These findings affirm the utility of quantile regression-corrected TRMM data for reliable use in diverse studies, particularly in water resources management and drought monitoring applications.



These compelling findings not only highlight the success of the quantile regression approach but also underscore its potential impact on refining precipitation data in other hydroclimatically sensitive regions. The corrected TRMM data, validated through rigorous metrics, emerges as a reliable and valuable resource for diverse applications, particularly in the realms of water resources management and drought monitoring. This study contributes to the evolving field of precipitation correction methodologies and emphasizes the significance of tailored approaches for optimizing the accuracy of remote sensing-derived data in hydrological studies.