Abstract
AbstractArchetypes of land- and socio-ecological systems, generated using unsupervised classification methods, enable the assimilation of complex environmental and socio-economic information. Such simplification has considerable potential to feed into decision support systems for sustainability planning. But, the usefulness of archetypes depends on how well they relate to sustainability criteria, such as ecosystem service (ES) delivery, that are external to the input datasets employed for archetype generation. Sensitivities in such post-hoc association analyses, and the associated utility of the archetype framework in a decision support context, remain unexplored. Here we emulated post-hoc association analysis procedures using simulated socio-ecological datasets and ES response variables. Our simulations revealed a substantial influence on analysis performance from (1) the number of variables used as inputs in archetype generation, (2) the correlation structure of input datasets, (3) the type and distribution of input variables, and (4) the functional form (linear or non-linear) characterising the relationship between ES variables and their predictors. We observed near-identical performance when archetypes were generated using K-means clustering and Self-Organising Maps (SOMs) – two commonly used archetype classification methods. Further, better archetype classifier performance did not guarantee better discrimination of ES value distributions between archetypes. Our results suggest that designing a framework to generate archetypes for sustainability planning, and the selected methodological choices therein, should place greater emphasis on what the archetypes will be used for in downstream analyses, and not focus solely on archetype classifier performance. This would better ensure the identification of archetypes adaptable to a diverse array of sustainability indicators and sufficiently robust for monitoring decision outcomes over time.
Publisher
Cold Spring Harbor Laboratory