A century of wild bee sampling: historical data and neural network analysis reveal ecological traits associated with species loss

Abstract

We analysed the wild bee community sampled from 1921 to 2018 at a nature preserve in southern Michigan, USA, to study long-term community shifts in a protected area. During an intensive survey in 1972 and 1973, Francis C. Evans detected 135 bee species. In the most recent intensive surveys conducted in 2017 and 2018, we recorded 90 species. Only 58 species were recorded in both sampling periods, indicating a significant shift in the bee community. We found that the bee community diversity, species richness and evenness were all lower in recent samples. Additionally, 64% of the more common species exhibited a more than 30% decline in relative abundance. Neural network analysis of species traits revealed that extirpation from the reserve was most likely for oligolectic ground-nesting bees and kleptoparasitic bees, whereas polylectic cavity-nesting bees were more likely to persist. Having longer phenological ranges also increased the chance of persistence in polylectic species. Further analysis suggests a climate response as bees in the contemporary sampling period had a more southerly overall distribution compared to the historic community. Results exhibit the utility of both long-term data and machine learning in disentangling complex indicators of bee population trajectories.

Keywords: bee decline; community analysis; conservation; nature preserve; neural network; pollinator.

Figure 1.

Community composition. Visual representation of dissimilarity between the bee community species composition (a) and trait composition (b) in the ESGR during historical (1972/1973) and contemporary (2017/2018) collections, based on Bray–Curtis dissimilarity. Plot in (b) uses trait variables from the neural network analysis with the following two differences: (i) instances of each categorical trait were counted as percentage of community composition, and (ii) latitude and longitude ranges instead of maximum and minimum values were used to avoid negative numbers (necessary for Bray–Curtis). Individual dots represent data collection months, and ellipses indicate the 95% confidence intervals. Arrows indicate the most significant bee species (a) or bee traits (b) (p < 0.01) driving the ordination of points. Arrow lengths correspond to the strength of the correlation between the species/traits and the ordination. Trait abbreviations in (b) are as follows (all per month community): oligoPerc, percentage of oligolectic bees in community; polyPerc, percentage of polylectic bees in community; avgLatR, average latitudinal range; avgPhR, average phenological range; avgPhM, average phenological mean date; avgITD, average ITD. (c) Proportions of species identified in the wild bee communities at the ESGR during two sampling periods, the historical (1972/1973) and the contemporary (2017/2018) samples. Species that represented more than 2% of the community are listed and coloured.

Figure 2.

Trait influence on all bee species extirpations between sample periods. Ridgeline plot representing the distribution of rescaled Olden importance values for each trait variable (input feature) across 1000 iterations of neural networks trained on the full trait/extirpation dataset. Neural networks used a single hidden layer of six neurons and were trained with a decay rate of 5 × 10⁻³ across 1500 maximum iterations. Trait data were prepared and used as neural network input features to predict bee species extirpation between the 1972/1973 historical period and the 2017/2018 contemporary period. Mean training prediction accuracy on extirpation/persistence was approximately 99%. Positive RSO values indicate positive association with extinction, while negative values indicate the increased association with persistence. Colours track the RSO values.

Figure 3.

Categorical trait changes between sample periods. (a–c) Percentage persistence of species separated by categorical traits across raw species counts. (d–f) Percentage of raw species richness across categorical traits grouped into species found only in the historical period (1972/1973), in both the historical and contemporary (2017/2018) periods, or only in the contemporary period. Note, sociality was not included in the neural network analysis (see Methods).