UC San Diego

Climate variations on seasonal to decadal time scales can have enormous social, economical and environmental impacts. As such, the ability to make skilful and reliable climate predictions at these time scales offers many benefits for climate preparedness, adaptation and resilience. In the recent years, major progress has been made in the development of such predictions with the advent of simulations with global climate models that are initialized from the current climate state. However, many challenges remain including an understanding of the underlying physical mechanisms for skilful predictions and whether such predictions could be improved. The purpose of this thesis is to establish new benchmarks for seasonal to decadal predictions in diverse components of the climate system and to provide some pieces of evidence that help to understand what are the drivers for these predictable patterns. Specifically, we use a suite of empirical models to perform predictions of oceanic and atmospheric variables together with initialized climate predictions to: 1. Understand the contribution of remote and local factors to the predictability of North and Tropical Pacific Oceans Sea Surface Temperature and Land Surface Temperature over Western North America; 2. Provide a higher baseline level skill for the state-of-art global prediction systems, from seasonal to decadal time scales; 3. Explore possible sources of errors in the global climate model simulations using statistical predictive models.

First, we isolate contributions to the forecast skill from different spatial and time scales in the Pacific Ocean using a Liner Inverse Modelling (LIM) approach, showing the importance of temporal scale interactions in improving the predictions on decadal time scales. Specifically, we show that the Extratropical North Pacific is a source of predictability for the tropics on seasonal to interannual time scales, while the tropics enhance the forecast skill for the decadal component. We then show that the skill for an empirically-built LIM is comparable to and sometimes better than that from two state-of-art global prediction systems, from seasonal to decadal timescales and for several regions around the globe. These results indicate that the evolution of the system in those areas may not be not fully driven by unpredictable dynamics and that there may be some room for improvement in the dynamical models predictions, given that a low-dimensional linear model is able to generate better skill than the fully-coupled nonlinear model. Bearing that in mind, we use the LIM linear feedback matrix to explore possible sources of errors in the dynamical model simulations and we find that some of the simulated atmospheric and oceanic local and remote feedbacks differ in several key regions from that obtained with observations. These results may indicate sources of error in the dynamical models and therefore in its prediction skill that merit focused attention.

We then investigate the role of remote and local predictors in seasonal predictors of minimum and maximum air temperatures over the Western North America, using a Canonical Correlation Analysis approach. We show that remote predictors, in the form of Pacific climate modes, provide the best predictive skill for temperature over land, particularly during wintertime. Lastly, considering that persistence is the widely-used measure when evaluating the predictive skill for dynamical models, we suggest the use of CCA as a much higher benchmark for seasonal predictions of land surface air temperatures.