A recent study provides insight into how well climate reanalysis data matches real world observations, which is crucial for understanding the growing impact of climate change. Scientists produce reanalysis data sets by combining past weather observations from various sources with modern climate models to make a continuous record over time. The study, published in the Journal of Geophysical Research: Machine Learning and Computation, evaluated the accuracy of four major reanalysis data sets in representing daily and extreme temperature across the United States. The researchers found that all reanalysis data sets were least reliable in the mountainous western U.S., likely due to the complex terrain. The European Centre for Medium-Range Weather Forecasts’ Reanalysis dataset, version 5 (ERA5), was the most accurate data set overall. ERA5 was best at predicting both daily and extreme temperatures, making it a valuable tool for future climate research.
The researchers also used machine learning techniques and a more traditional analysis method to compare data from over 300 weather stations with reanalysis outputs. They found that the machine learning technique known as a Feed-Forward Neural Network (FFNN) captured complex relationships in the data better than other analysis methods. The researchers suggest that future studies should continue exploring machine learning to enhance our ability to represent complex variables. The Climate Program Office’s Modeling, Analysis, Predictions, and Projections (MAPP) Program supported this work through a grant to develop new model-based monitoring products addressing key climate impact areas. Lead author Chibuike Ibebuchi, a postdoctoral researcher at Kent State University, works with MAPP-funded scientist Cameron Lee on a project aimed at monitoring hazardous temperature conditions in North America.
For more information, contact Amanda Chiachi.