Climate change is resulting in regional increases in the frequency of extreme climatic events, including multi-year wet or dry periods. Recent observations from the Chihuahuan Desert show a remarkable perennial grass response in desertified shrublands that has been sustained (2010-present) following a wet period (2004-2008) that differs from the prior 15-year record with little or no grasses (1989-2003). This greening of the desert suggests a shift to an alternative state where production responses cannot be predicted from historic relationships between rainfall and primary production. Our objective was to develop an artificial intelligence (AI) system that could: first, identify discrete changes in patterns in rainfall, such as multi-year wet, dry or recovery periods based on rainfall amount by year; second, develop relationships between production, rainfall, and other explanatory variables for each period; and third, apply this learned behavior to other locations. To develop our system, we analyzed long-term datasets from the Jornada USDA-LTER site in southern New Mexico on primary production, rainfall, and soil properties for 5 locations spatially distributed across the landscape. We then tested our AI system for an additional 10 spatially-distributed locations at the Jornada.
Results/Conclusions
Desert greening was maintained for at least ten years following the wet period for all locations. However, the magnitude of perennial grass responses was unevenly distributed across the Jornada landscape. Largest responses occurred on sandy soils of mesquite shrublands and smallest responses occurred on soils dominated by creosotebush. Both rainfall amount and soil properties were needed in our AI system to predict grass production through time. Our AI system was able to identify thresholds or discrete changes in rainfall patterns and to develop relationships between rainfall and explanatory variables for each rainfall period. Our AI system also learned through time as new rainfall or production amounts were experienced. Ultimately, our goal is to improve prediction of responses in ecosystems where nonlinear dynamics with thresholds are important in time and space as climate continues to change, thus making application of static relationships from long-term data questionable. AI methods provide one dynamic approach to developing these predictive relationships.