APR 4, 2016 GENEVA GRAY
There is a lot of data out there. It seems like every agency has produced their own downscaled dataset using different methods, training data, and a hodge-podge of global climate models. They are all unique, but none of them are the “best.” This blog post will not give you tips in working downscaled data or picking what is right for your project; my colleague already wrote that post awhile back. Instead, I’m sharing how my project uses downscaled data and how I am developing a new analytical way to whittle down the number of models needed.
I work on a project centered around the New River Estuary in North Carolina. This location is home to Camp Lejeune, a Marine Corps base. Like many other Department of Defense facilities, Camp Lejeune plans for their future land, species, and environment management strategies, which all include anticipating the effects of climate change. My role is to evaluate the performance of downscaled climate projections in Eastern North Carolina and to create an ensemble of climate models to be used in terrestrial, coastal, and estuarine models based on my evaluation results.
In an ideal world, everyone would use an ensemble of downscaled climate models in their analysis. An ensemble is simply a collection of multiple models, which woefully lacks a woodwinds section. Your ensemble can contain either one downscaled set or span across several downscaling methods. It provides many benefits including a measure of climate change uncertainty, a range of possible futures, and a “wisdom of a crowd” feature where the average of all the models actually out-performs any individual model. All of this leads up to the suggestion that adding more models will give you a better dataset. However, this isn’t exactly the case.
First, there are issues on the climate data user’s end. Some models take too long, or are too complicated, to run 30 times over. This means a larger ensemble may not necessarily be the best dataset for your needs. Next, there are issues on the data creation end. Recent literature in the Global Climate Model (GCM) community has identified a few dependency issues between different GCMs. While each GCM is created by a different organization, the creators may use the same sub-model components or the same coding logic. This creates similar models and similar model error which affect uncertainty calculations and many other useful statistics calculated from an ensemble of climate models.
Luckily, there is a way to find dependent global climate models and reformulate the ensemble to create a product that is statistically rigorous. The catch: this hasn’t been done with downscaled climate models… yet.
In my research for Eastern North Carolina military base managers, I test methodology to reduce the number of ensemble members while still maintaining the spread of possible futures. I start with seven downscaled (dynamic, statistical, and constructed analogue) datasets totaling to 83 individual ensemble members. Below are the (simplified) steps I take in order to find any model dependence issues:
- Evaluate the model baseline against an observation dataset.
- Compare the residual error of each model and find sources of similarities.
- Identify the similarities and cluster those models together.
- Produce a reduced ensemble using the calculated clusters.
Most ecosystem modelers do not have the ability to run their models 83 times to see the complete range of how climate change may affect their model output. However, by identifying the similarities and clustering dependent models, I hope to present the users with a reduced ensemble tailored for the New River Estuary in North Carolina. This ensemble with be both analytically independent and smaller to increase ease of use.
So yes, there are a lot of climate data choices out there and not everyone has the time to find and account for sources of dependence. My future work includes analyzing larger and different areas to help identify some rigorous and reduced ensembles of downscaled climate projects for a wider number of users. Until then, happy data hunting!
“Addressing interdependency in a multi-model ensemble by interpolation of model properties” By Sanderson B., Knutti R., and Caldwell P.
“Addressing interdependency in a multi-model ensemble . Part 2 – A Representative Democracy” By Sanderson B., and Caldwell P.
“On the effective number of climate models” By Pennell C., and Reichler T.
Geneva is a graduate research assistant at the Southeast Climate Science Center.