A New Multiple Imputation Approach Using Machine Learning to Enhance Climate Databases in Senegal

Abstract

Abstract This study aims at enhancing climate data in Senegal using information from the Global Surface Summary of the Day (GSOD). It uses data from 1991 to 2022 from major secondary synoptic stations in Senegal. These data are subject to missing values (data gaps). To address these gaps, multiple imputation was used based on three machine learning models: PMM (Predictive Mean Matching), RF (Random Forest), and NORM (Bayesian Linear Regression). The PMM model relies on averages of similar data, the RF model handles complex relationships between variables, even on an intra-seasonal scale, while the NORM model captures seasonal variations and extreme values. The results highlight the higher performance of the RF model in terms of accuracy and variance explanation compared to the others. The findings of this study open new avenues for informed decision-making in sectors such as agriculture and urban planning, where accurate climate data play a crucial role. However, while this study lays the groundwork for better utilization of climate data in Senegal, challenges persist, including the ongoing need to collect high-quality data and adapt models to data intricacies.