To accurately predict where trees will thrive in the next 50 to 100 years, we must first understand where they thrive today. Current species range data and ecosystem classifications, however, suffer from limitations in accuracy, adaptability, and scalability. This project aims to overcome these challenges by producing a species distribution model capable of adapting to changing climates, accurately informing regeneration strategies and forest inventories.

Early efforts relying on pre-existing ecosystem classifications resulted in significant errors, prompting the use of trained machine learning models to predict species distributions based on historical data and geo-climatic variables. For this project I have developed a deep learning approach using deep neural networks. This approach yields improved forecasts of species ranges for key important forest tree species in North America under historical climate conditions.

Forest inventory and ecological plot data from the U.S. and Canada were combined to develop a robust dataset of species occurrences. To reduce the influence of misidentifications and out-of-range records, observations were filtered using historical species range maps with a 200 km spatial buffer. Additionally, plot data were weighted according to modeled land cover probabilities to better represent actual forest cover conditions.

Predictor variables included 24 bioclimatic variables generated by ClimateNA (1951–1980 normal period), 15 topographic variables derived from digital elevation models, and 19 land cover class probabilities produced by a separate deep neural network trained on MODIS remote sensing data using the same set of predictor variables. These inputs were used to train species frequency models across North America, which were then applied to generate continent-wide species distribution maps based on geo-climatic habitat suitability.

The accuracy of species distribution predictions was evaluated through visual comparison with historical species range maps and statistical validation using known species occurrences from forest inventory plots.

These models have undergone multiple iterations, with refined input variables to improve mapping accuracy. The final framework demonstrates strong capacity to predict species habitat suitability and distribution using a diverse suite of environmental predictors. It supports species selection for both commercial and ecological regeneration, accommodates future climate projections, and can be paired with MODIS-derived land cover data to inform spatially explicit forest inventory and management planning.