Assessing spatiotemporal resolution of variables in landscape-scale species distribution models

The strategic importance of biodiversity conservation is increasing all over the world to face the threats that the global change bring to forest ecosystems. To accomplish that, Species Distribution Models (SDMs) stand as the most employed statistic models in ecological conservation. Nevertheless, explanatory predictors in these correlative models usually consist in free-air climate variables with coarse spatial (>1 km) and temporal (average of several years) resolution. This approach neglects the real habitat conditions experienced by most of the organisms on their life span. Hence, improving the reliability of these ecological models is crucial for biologists, land managers, and policymakers.

Our aim was to compare microclimate temperatures derived from 13 years (2010-2022) of below-canopy hourly data loggers to free-air macroclimate derived from a mechanistic downscaling of global reanalysis data at 30 arcsec (CHELSA). We also tested the role of LiDAR-derived vegetation metrics to improve fine-scale SDMs. We developed three sets of predictors based on their temporal resolution: 1) an average across the years of observation based on the general presence or absence of the species, 2) an ensemble of year models (i.e., the average of the probability for each year weighted on its accuracy), and 3) a random year of observation.

Using Bayesian Additive Regression Tree (BART) algorithms, we built SDMs for different species of birds, plants, insects, and mammals in a temperate rainforest landscape of the Pacific Northwest (HJ Andrew Experimental Forest, Oregon, United States). We built 12 different modeling frameworks based on the combination between climate input data (microclimate vs macroclimate), vegetation (with vs without), and temporal resolution (average vs ensemble vs random).

We measured the distance between the probability distribution obtained from the different combinations using the Kolmogorov-Smirnov distance. We evaluated and compared three accuracy metrics of the models (AUC, TSS, MCS) through a 5-fold spatial block cross-validation. We tested for differences in distance and accuracy both at the taxa level and on ecological traits of the different species (mobility, prevalence, specialization).

Preliminary results for bird species showed that temporal resolution was more important than climate datasets when including vegetation variables in the models. Some insect species showed a greater improvement in accuracy for models trained with microclimate compared to models trained with CHELSA.