TITLE:
Multi-Task Gaussian Process for Imputing Missing Daily Rainfall Data Using Nearby Stations: Case of Burkina Faso
AUTHORS:
Souleymane Zio, Dazangwende Emmanuel Poan, Yoda Adaman, Kima Bénéwendé Yves Arnaud Noe
KEYWORDS:
Rainfall, Missing Data, Multi-Task Gaussian Process, Correlation
JOURNAL NAME:
Journal of Sensor Technology,
Vol.15 No.1,
January
24,
2025
ABSTRACT: Precipitation is a critical meteorological factor that significantly impacts agriculture in the sub-Saharan and Sahelian regions of Africa. Accurate knowledge of precipitation levels aids in planning effective agricultural strategies. However, these regions often face challenges with missing rainfall data at numerous gauges. This issue arises due to various factors, including socio-political instability (e.g., terrorism), economic constraints (e.g., insufficient station density due to limited resources), and human factors such as a shortage of qualified personnel. This study evaluates the effectiveness of the multi-task Gaussian process (MTGP) based on the linear model of coregionalization (LMC) for imputing missing daily rainfall data in Burkina Faso, leveraging the correlations among nearby stations. The proposed method is compared with commonly used statistical and machine learning techniques, including mean imputation (ME), K-nearest neighbors (KNN), Multivariate Imputation by Chained Equations (MICE), and Last Observation Carried Forward (LOCF). The results demonstrate that the MTGP approach outperforms MICE, KNN, LOCF, and ME. Additionally, when compared to the independent Gaussian process (IGP), which does not account for correlations between stations, MTGP shows a performance improvement of 50%.