A spatio-temporal statistical downscaling model for combining spatially misaligned maximum temperature data using R-INLA
Format: CPS Abstract
Climate data are essential for analysing and modelling climate variability and trends and their impacts on different health and socio-economic activities. Though, Africa is one of the most vulnerable regions to climate change, such climate studies and applications are very scarce in Africa due to the limited availability and access to climate data. The weather stations are sparse and unevenly distributed across many parts of Africa and suffer from large proportions of missingness over space and time. Thus, the in-situ geostatistical climate data measured directly from monitoring weather stations are assumed to be accurate within its measurement error, but sparse in space and time. Alternatively, physical climate model output provides another source of climate gridded data that cover large and dense spatial and temporal domains at a certain resolution but not at smaller scales. Physical climate model data do not account for the uncertainty in the data and hence tend to exhibit bias compared to the in-situ observations. To enhance the accuracy of climate model outputs and the spatial and temporal coverage of in-situ data, the simulated climate model output data can be calibrated against the real measurements from monitoring weather stations. However, integrating such climate model output data with in-situ observed data for improved interpolation is not trivial as it involves misalignment in space and time which can lead to biases in predictions. In this study, our aim is to present a statistical downscaling framework for combining the monthly maximum temperature observations from 52 monitoring weather stations across the Nile Basin countries throughout the 10-year period (2011 – 2020) with the 44 x 44 km gridded data simulated from a regional climate model (RCM) across the same study region and study period. To accurately account for uncertainty from the different data sources and propagate it to predictions, a spatio-temporal coregionalization model that assumes a joint distribution between the covariate (simulated model output) and the response (in-situ observations) is employed. The proposed spatio-temporal coregionalization model is fitted under a hierarchical Bayesian framework using integrated nested Laplace approximation (INLA) coupled with stochastic partial differential equation (SPDE) approach. This spatio-temporal model assumes that the true underlying process for both sources of data is a Gaussian process. This allows us to flexibly adjust the spatio-temporal latent field of the simulated climate model data, improve the prediction of the monthly maximum temperature data along the entire spatial domain at finer resolution and forecast the monthly maximum temperature for the future. By accounting for the temporal structures, the model is able to detect extreme weather events like El Niño and La Niña. The model predictions are compared against the results of a spatio-temporal generalized additive model (GAM) fitted using only one source of data (in-situ observations) and proved to be relatively better by lowering the root mean square error.
Keywords: Africa, Coregionalization model, Gaussian process, INLA, Regional climate model, Spatio-temporal, Statistical downscaling.