Broad Genus Associations Can Inform a Forest Type Classification Framework

Forest types provide a grouping variable for analyses of large, tree-based databases (e.g., forest structure applications). However, there is not a single universally standard method for classifying forest types. Here, we investigated the potential of translating tree genus association rules discovered in association rule mining (ARM) into a forest type classification across a large spatial scale, leveraging an N = 1844 forest inventory plot dataset selected from Mexico’s national forest inventory database (Inventario Nacional Forestal y de Suelos) in a proof-of-concept study. We hypothesized that a set of association rules discovered across different transaction construction methods and geographic partitions could produce a set of stable itemsets suitable for translation into a forest type classification. Of the 167 distinct genera in the dataset used for rule mining, 4 genera—Quercus, Pinus, Arbutus, and Juniperus—and a recurrent set of 5 associations among them were translated into 7 “overlap-aware” forest types, with 74.8% of plots assigned to a forest type and the remainder retained as unclassified. We then described variation in post-classification forest structure in the context of potentially confounding environmental and plot spacing effects. Promising, but not conclusive, results reveal broad variation in forest structure described by the classification, with groups of low, medium, and high median basal area (m2/ha) and density (trees/ha) emerging across the classes. The spatial distributions of the classes tended to be visually clustered, whereas unclassified plots had a more visually even distribution, suggesting some correspondence of the classification with environmental and spatial context within the analyzed dataset. We discuss the strengths and limitations of the studied ARM-informed forest type classification framework, including its potential utility in the study region of Mexico, while noting that testing in other regions is warranted for external geographic validation. Finally, we offer suggestions for adapting and refining the framework for robustness and potential utility for similar ecological applications.