Utilizing the density of inventory samples to define a hybrid lattice for species distribution models: DISTRIB‐II for 135 eastern U.S. trees

Abstract

Species distribution models (SDMs) provide useful information about potential presence or absence, and environmental conditions suitable for a species; and high‐resolution models across large extents are desirable. A primary feature of SDMs is the underlying spatial resolution, which can be chosen for many reasons, though we propose that a hybrid lattice, in which grid cell sizes vary with the density of forest inventory plots, provides benefits over uniform grids. We examine how the spatial grain size affected overall model performance for the Random Forest‐based SDM, DISTRIB, which was updated with recent forest inventories, climate, and soil data, and used a hybrid lattice derived from inventory densities. Modeled habitat suitability was compared between a uniform grid of 10 × 10 and a hybrid lattice of 10 × 10 and 20 × 20 km grids to assess potential improvements. The resulting DISTRIB‐II models for 125 eastern U.S. tree species provide information on individual habitat suitability that can be mapped and statistically analyzed to understand current and potential changes. Model performance metrics were comparable among the hybrid lattice and 10‐km grids; however, the hybrid lattice models generally had higher overall model reliability scores and were likely more representative of the inventory data. Our efforts to update DISTRIB models with current information aims to produce a more representative depiction of recent conditions by accounting for the spatial density of forest inventory data and using the latest climate data. Additionally, we developed an approach that leverages a hybrid lattice to maximize the spatial information within the models and recommend that similar modeling efforts be used to evaluate the spatial density of response and predictor data and derive a modeling grid that best represents the environment.

Facilitating Adaptive Forest Management under Climate Change: A Spatially Specific Synthesis of 125 Species for Habitat Changes and Assisted Migration over the Eastern United States

“…The overall synthesizing approach uses the results of the previously mentioned DISTRIB-II, SHIFT, and ModFacs model components; details on these components are provided in an online supplement as well as previous papers (Figure 2, Supplementary File 1-Climate Change Atlas description) [11,[19][20][21]. DISTRIB-II provides estimates of habitat suitability (or habitat quality, HQ), by species, to century's end according to various scenarios of climate change.…”

“…Range indicates whether species is wide or narrow (W/N) in distribution, dense or sparse (D/S) in frequency of presence within FIA plots, and high or low (H/L) in importance value when it is found. MR refers to model reliability (see [21] for explanation). %Cell refers to the percentage of DISTRIB-II cells with the species present.…”

“…The species is also categorized as found commonly (Dense if ≥40% of FIA plots among grid cells with IV >0) or rarely (Sparse if <40% of FIA plots have the species) within its overall distribution among eastern US FIA records. Finally, the ecological importance or abundance for each species can be indicated by the average importance value among the plots where the species is present, with High if average IV ≥6.0 (the median of mean IV across all species) and Low for values <6.0 [21]. We also assigned three levels of model reliability (MR) [21].…”

“…Finally, the ecological importance or abundance for each species can be indicated by the average importance value among the plots where the species is present, with High if average IV ≥6.0 (the median of mean IV across all species) and Low for values <6.0 [21]. We also assigned three levels of model reliability (MR) [21]. To do so, we evaluated and combined five model performance variables into a single rating: (1) a pseudo-R 2 obtained from the RandomForest (RF) model; (2) a Fuzzy Kappa comparing the imputed RF map to the FIA-derived map [22]; (3) a tree skill statistic of the imputed RF, after removing records with very high coefficient of variables (CV); (4) the deviance of the CV among 30 regression trees via bagging [23,24]; and (5) the stability of the top five variables from 30 regression trees [25].…”

“…By ratioing suitable habitat (IV) in future to suitable habitat at present, we can generate five classes of change (noted as ChngCl45 or ChngCl85, depending on representative concentration pathways (RCP) 4.5 and RCP 8.5; Figure 3, [21]) according to the following break points in future:actual IV ratio: large decrease (<0.5), small decrease (0.5-0.8), no change (0.8-1.2), small increase (1.2-2.0), large increase (>2.0). For rare species, those occupying <10% of the region, the following break points were used: large decrease (<0.2), small decrease (0.2-0.5), no change (0.5-4.0), small increase (4.0-8.0), and large increase (>8.0).…”