Reconstructing historical habitat data with predictive models

Abstract

Historical vegetation data are important to ecological studies, as many structuring processes operate at long time scales, from decades to centuries. Capturing the pattern of variability within a system (enough to declare a significant change from past to present) relies on correct assumptions about the temporal scale of the processes involved. Sufficient long-term data are often lacking, and current techniques have their weaknesses. To address this concern, we constructed multistate and artificial neural network models (ANN) to provide fore- and hindcast vegetation communities considered critical foraging habitat for an endangered bird, the Florida Snail Kite (Rostrhamus sociabilis). Multistate models were not able to hindcast due to our data not satisfying a detailed balance requirement for time reversibility in Markovian dynamics. Multistate models were useful for forecasting and providing environmental variables for the ANN. Results from our ANN hindcast closely mirrored the population collapse of the Snail Kite population using only environmental data to inform the model. The parallel between the two gives us confidence in the hindcasting results and their use in future demographic models.