Data-driven research in a nonstationary world: a top-down approach to understanding critical interactions and scales from the catchment to the planet

Seminar | March 8 | 4-5 p.m. | 1011 Evans Hall

 Laurel Larsen, Berkeley Geography Department

 Department of Statistics

One consequence of earth systems moving out of a regime of stationarity is that statistical models based on past behavior may no longer be useful for predicting the future. Rather, an understanding of the mechanisms driving dynamic earth systems is needed. The mechanisms responsible for nonlinear—even surprising—behavior often involve feedbacks between biotic and abiotic processes. Examples of these feedback processes abound in aquatic sciences, where flow, vegetation, sediment, topographic, and biogeochemical processes tend to exhibit strong coupling. Effects of biotic-abiotic feedback processes may be examined using numerical models (a bottom-up approach), but because multiple feedback processes may produce the same outcome, these studies may not be conclusive. Likewise, although correlative field studies may be useful for generating hypotheses about system drivers, they are often not sufficient to resolve causally important processes in complex hydrologic systems with multiple limiting factors or nonlinearities.

An emerging alternative to the bottom-up method is a top-down approach in which causal interactions and their critical spatiotemporal scales are delineated directly from data using emerging frameworks of inference. Challenges of not knowing the functional form of the relationship between drivers or which of several potential drivers is limiting a response at any time are dealt with by performing analyses in the framework of uncertainty reduction rather than prediction. If a variable (such as discharge) uniquely and independently reduces the uncertainty of another variable (such as ecosystem respiration) over particular time lags, we conclude that it exerts a causal influence. In this manner, and using a stream gaging station data record, we show that fine sediment deposition by intermediate-size storms boosts stream metabolism over timescales of 100+ days in an urban stream that is limited by particulate carbon. Conversely, large storms depress stream respiration over timescales of one month in streams with close coupling between gross primary productivity and respiration but over 1-2 day timescales in streams where these processes remain uncoupled. Application of these methods over larger spatial scales can reveal critical spatial scales and pathways of rainfall recycling in transitional forests of Brazil, or the most prominent drivers of global temperature changes during historic transitions from glaciated to unglaciated states. These case studies highlight the potential for application of these emerging causal inference techniques to the vast library of hydroecological data already collected, provide insight into optimal strategies for data collection in new sensor networks, and showcase how data-driven research may improve mechanistic model generation.