Combining precipitation data from observed and numerical models to forecast precipitation characteristics in sparsely-gauged watersheds: an application to the Amazon River basin.

Abstract:

Forecasting rainfall in areas with sparse monitoring efforts is critical to making inferences about the health of ecosystems and built environments. Recent advances in scientific computing have allowed forecasting and climate models to increase their spatial and temporal resolution. Combined with observed point precipitation from monitoring stations, these models can be used to inform dynamic spatial statistical models for precipitation using methods from geostatistics and machine learning.

To prove the feasibility, process, and capabilities of these statistical models, we present a case study of two statistical models of precipitation for the Amazon River basin from 2003-2010 that can infer a spatial process at a point using areal data from numerical model output. We investigate the seasonality and accumulation of rainfall, and the occurrence of no-rainfall and large-rainfall events. These parameters are used since they provide valuable information on possible model biases when using climate models for forecasts of the future process of precipitation in the Amazon basin. This information can be vital for ecosystem, agriculture, and water-resource management.

We use observed precipitation data from weather stations, three areal datasets derived from observed precipitation (CFSR, CMORPH-CRT, GPCC) and three climate model precipitation datasets from CMIP5 (MIROC4h, HadGEM2-CC, and GISS-E2H) to construct the models. The observational data in the model domain is sparse, with 195 stations in the approximate 7×10⁶ square kilometers of the Amazon basin, and therefore requires the areal data to create a more robust model.

The first model uses the method of Bayesian melding to combine and make inferences from the included data sets, and the second uses a regression model with spatially and temporally-varying coefficients. The models of precipitation are fitted using the areal products and a subset of the point data, while another subset of point data is held out for verification. The statistical models are evaluated based on the predictive performance using metrics such as mean absolute errors and coverage of predictive intervals. Precipitation estimates from the climate models are also evaluated on the basis of their performance in analyzing the accumulation, seasonality, and occurrence of rainfall.