From occurrences to predictors: modeling species distribution and environmental drivers of Nymphalidae

Abstract

Anthropogenic climate change poses growing threats to global biodiversity. These threats put pollinators at risk, which affects their ecological and agricultural roles. Butterflies in the Nymphalidae family act as both pollinators and indicators of environmental change, making them valuable models for ecological research. The reliability of occurrence data--from professional or participatory--significantly shapes species distribution modeling outcomes. Chapter 1 evaluates the comparative performance of professional, participatory, and combined occurrence datasets in modeling Nymphalidae distributions across two climatically sensitive North American ecoregions, the Western Cordillera and South Central Semi-Arid Prairies. Using MaxEnt and Random Forest algorithms across four temporal bins (2008-2022), we demonstrate that Random Forest consistently outperforms MaxEnt in predictive performance. Spatial autocorrelation analyses revealed fundamental differences between data sources: participatory records exhibited strong spatial clustering (Moran's I = 0.36-0.65), while professional records showed weak to moderate clustering (Moran's I = 0.13-0.32), reflecting systematic differences in sampling efforts. Random Forest models built with professional datasets achieved the highest performance (AUC = 0.984; TSS = 0.944), while combined datasets offered the best trade-off between spatial coverage and predictive strength. Chapter 2 applies these methodological insights to Vanessa cardui (Nymphalidae), a migratory butterfly and ecological generalist, examining species distribution across the same ecoregions. Random Forest models again achieved strong predictive performance (AUC = 0.968; TSS = 0.895). Variable importance analyses identified precipitation seasonality, maximum temperature of the warmest month, and mean diurnal range as the strongest predictors of species distribution. Residual diagnostics revealed systematic deviations at low and high suitability values, highlighting challenges in predicting rare outcomes. However, removing abundance-related bias shifted environmental relationships: population change was more strongly associated with temperature seasonality and diurnal variation, while consistently warm conditions showed negative correlations. Spatial predictions revealed persistent but patchy areas of suitability from 2008-2022, concentrated in both urban-adjacent and remote landscapes, with overall suitable habitat comprising a small proportion of the total study area. Together, this research demonstrates both the potential of integrating diverse datasets with machine learning to assess species distributions and the importance of selecting appropriate datasets, as outcomes and interpretations highly depend on data source and quality. Simultaneously, it emphasizes the methodological and interpretive challenges posed by sampling bias and residual patterns. By identifying key environmental drivers and spatial hotspots, these models inform future monitoring and conservation strategies for butterflies in the face of accelerating environmental change.