Investigating the explanatory and predictive power of geodiversity metrics on Swedish insect diversity

Understanding and predicting biodiversity patterns at fine spatial scales is a key challenge in biological modeling. Commonly used coarse-resolution variables often lack sufficient resolution for fine-scale modeling, or the grid size of fine-resolution variables does not align with the scale of biodiversity sampling. This study investigates whether geodiversity metrics—quantitative measures of abiotic heterogeneity—can improve both the explanatory and predictive performance of biodiversity models, using insect metabarcoding data collected across Sweden. Three major insect orders—Coleoptera, Diptera, and Hymenoptera—are analyzed based on 190 Malaise trap sampling sites from the Insect Biome Atlas Project. The study tests three geodiversity metrics: mean square roughness (MSR) for continuous environmental features (temperature, soil moisture, etc.), and landscape Shannon’s diversity index (SHDI) and landscape shape index (LSI) for categorical ones (soil type and forest presence). To examine spatial differences in explanatory power, the modeling area is divided into north and south Sweden, reflecting their contrasting climatic conditions and geomorphological structures. Model-based inference using AICc and Lasso regression is applied for assessing the explanatory and predictive power of individual features, and also for feature selection. Supportive vector regression (SVR) is applied to assess predictive accuracy. Results show that geodiversity variables improved the AICc optimized model in both north and south Sweden, and to a larger extent in the south. While AICc optimization for explanatory excludes many geodiversity variables in northern Sweden, Lasso regression that focuses on prediction tends to retain them, indicating their practical value in predictive modeling. With proper feature selection, geodiversity variables improve predictive performance. Models with geodiversity also better reflect observed biodiversity and ecological expectations than baseline environmental models. Biodiversity datasets are usually small due to their cost in systematic sampling; large datasets are hardly available, and feature dimensionality is limited. Assessment of feature explanatory power has shown that many redundant variables are commonly used, and they have both less explanatory and predictive power in comparison to geodiversity variables. This not only suggests that geodiversity variables are potentially good candidate variables but also highlights the importance of exploration for stronger features. In conclusion, this study demonstrates that integrating geodiversity into biodiversity models can significantly enhance their explanatory and predictive capabilities. These findings support the broader use of geodiversity in biodiversity modeling, especially for fine-scale ecological assessments and conservation planning.