Deep learning for identifying bee species from images of wings and pinned specimens.

One of the most challenging aspects of bee ecology and conservation is species-level identification, which is costly, time consuming, and requires taxonomic expertise. Recent advances in the application of deep learning and computer vision have shown promise for identifying large bumble bee (Bombus) species. However, most bees, such as sweat bees in the genus Lasioglossum, are much smaller and can be difficult, even for trained taxonomists, to identify. For this reason, the great majority of bees are poorly represented in the crowdsourced image datasets often used to train computer vision models. But even larger bees, such as bumble bees from the B. vagans complex, can be difficult to separate morphologically. Using images of specimens from our research collections, we assessed how deep learning classification models perform on these more challenging taxa, qualitatively comparing models trained on images of whole pinned specimens or on images of bee forewings. The pinned specimen and wing image datasets represent 20 and 18 species from 6 and 4 genera, respectively, and were used to train the EfficientNetV2L convolutional neural network. Mean test precision was 94.9% and 98.1% for pinned and wing images respectively. Results show that computer vision holds great promise for classifying smaller, more difficult to identify bees that are poorly represented in crowdsourced datasets. Images from research and museum collections will be valuable for expanding classification models to include additional species, which will be essential for large scale conservation monitoring efforts.

Native Flowering Border Crops Attract High Pollinator Abundance and Diversity, Providing Growers the Opportunity to Enhance Pollination Services.

Over the past century, habitat loss from agricultural intensification has contributed to pollinator decline. One way to mitigate the harmful effects of agricultural intensification is through the re-introduction of native flowering plants as border strips that provide supplemental floral and nesting resources to pollinators. However, border crop species vary in bloom period and flower densities, and are thus likely to attract different suites of pollinator species. Resulting differences in pollinator community composition are likely to affect their ability to provide pollination services to adjacent crop habitat. To address these issues, we implemented a two-year study on the impact of different flowering border crops on pollinator abundance, richness, and community composition. We also examined which crop features (bloom duration, number of flowers, floral area) were most likely to affect pollinator densities. We found that native flowering plant border crops of diverse prairie mix and monocultures of silflower (Silphium integrifolium Michx.) and cup plant (Silphium perfoliatum L.) attracted the highest abundance and species richness of bees and pollinator groups combined, while alfalfa (Medicago sativa L.) attracted the highest lepidopteran abundance and species richness. We also found a significant, positive relationship between pollinator abundance and floral resource amount and bloom duration. These findings offer valuable insight into the impacts of different land management strategies on different pollinator groups, and thus provide landowners with management options for attracting specific pollinator groups and species.

Persistent soil carbon enhanced in Mollisols by well-managed grasslands but not annual grain or dairy forage cropping systems.

Intensive crop production on grassland-derived Mollisols has liberated massive amounts of carbon (C) to the atmosphere. Whether minimizing soil disturbance, diversifying crop rotations, or re-establishing perennial grasslands and integrating livestock can slow or reverse this trend remains highly uncertain. We investigated how these management practices affected soil organic carbon (SOC) accrual and distribution between particulate (POM) and mineral-associated (MAOM) organic matter in a 29-y-old field experiment in the North Central United States and assessed how soil microbial traits were related to these changes. Compared to conventional continuous maize monocropping with annual tillage, systems with reduced tillage, diversified crop rotations with cover crops and legumes, or manure addition did not increase total SOC storage or MAOM-C, whereas perennial pastures managed with rotational grazing accumulated more SOC and MAOM-C (18 to 29% higher) than all annual cropping systems after 29 y of management. These results align with a meta-analysis of data from published studies comparing the efficacy of soil health management practices in annual cropping systems on Mollisols worldwide. Incorporating legumes and manure into annual cropping systems enhanced POM-C, microbial biomass, and microbial C-use efficiency but did not significantly increase microbial necromass accumulation, MAOM-C, or total SOC storage. Diverse, rotationally grazed pasture management has the potential to increase persistent soil C on Mollisols, highlighting the key role of well-managed grasslands in climate-smart agriculture.

Assessing the potential for deep learning and computer vision to identify bumble bee species from images.

Pollinators are undergoing a global decline. Although vital to pollinator conservation and ecological research, species-level identification is expensive, time consuming, and requires specialized taxonomic training. However, deep learning and computer vision are providing ways to open this methodological bottleneck through automated identification from images. Focusing on bumble bees, we compare four convolutional neural network classification models to evaluate prediction speed, accuracy, and the potential of this technology for automated bee identification. We gathered over 89,000 images of bumble bees, representing 36 species in North America, to train the ResNet, Wide ResNet, InceptionV3, and MnasNet models. Among these models, InceptionV3 presented a good balance of accuracy (91.6%) and average speed (3.34 ms). Species-level error rates were generally smaller for species represented by more training images. However, error rates also depended on the level of morphological variability among individuals within a species and similarity to other species. Continued development of this technology for automatic species identification and monitoring has the potential to be transformative for the fields of ecology and conservation. To this end, we present BeeMachine, a web application that allows anyone to use our classification model to identify bumble bees in their own images.

Harvesting effects on wild bee communities in bioenergy grasslands depend on nesting guild.

Conversion of annual crops to native perennial grasslands for bioenergy production may help conserve wild bees by enhancing nest and food resources. However, bee response to the disturbance of biomass harvesting may depend on their nesting location, thus their vulnerability to nest destruction, and the response of the forb community on which they forage. Moreover, because bees have long foraging ranges, effects of local harvesting may depend on the amount of natural habitat in the surrounding landscape. We performed a large-scale one- and two-year experiment in Michigan and Wisconsin, USA, respectively, to examine how grassland harvesting, landscape context, and study year affect the forb community, above- and belowground-nesting bee species richness, community composition, trap nest emergence, and visitation rate. In Wisconsin, harvesting increased forb richness, cover, and evenness compared to unharvested control sites. Harvesting negatively affected aboveground-nesting bee richness and emergence from trap nests, possibly because of nest destruction during the previous harvest. By contrast, harvesting positively affected belowground-nesting bee richness, possibly because of the greater food resource availability and reduced thatch allowing greater access to nesting sites in the soil. Harvesting also affected bee community composition, reflecting the increase in belowground-nesting species at harvested sites. Despite harvesting effects on forb and bee communities, there was no effect on flower visitation rate, indicating little effect on pollination function. We did not find a harvest by landscape context interaction, which, in combination with the negative harvesting effect on trap nest emergence, suggests that harvesting can affect local population growth rather than simply affecting forager aggregation in different resource environments. For bees, there was no harvest by study year interaction, indicating a consistent response over a short timescale. Similarly, in Michigan, belowground-nesting species also responded positively to harvesting, which was more pronounced in sandier soils that are preferred for nesting. However, other components of the Michigan bee and forb communities were not significantly affected by biomass harvesting. Overall, our study demonstrates that harvesting grasslands can positively affect the ~80% of bee species that nest belowground by enhancing nest and/or forage resources, but that conserving aboveground nesters may require leaving some area unharvested.

Carbon storage potential increases with increasing ratio of C4 to C3 grass cover and soil productivity in restored tallgrass prairies.

Long-term soil carbon (C) storage is essential for reducing CO2 in the atmosphere. Converting unproductive and environmentally sensitive agricultural lands to grasslands for bioenergy production may enhance C storage. However, a better understanding of the interacting effects of grass functional composition (i.e., relative abundance of C4 and C3 grass cover) and soil productivity on C storage will help guide sustainable grassland management. Our objective was to examine the relationship between grass functional composition and potential C storage and how it varies with potential soil productivity. We estimated C inputs from above- and belowground net primary productivity (ANPP and BNPP), and heterotrophic respiration (R H) to calculate net ecosystem production (NEP), a measure of potential soil C storage, in grassland plots of relatively high- and low-productivity soils spanning a gradient in the ratio of C4 to C3 grass cover (C4:C3). NEP increased with increasing C4:C3, but only in potentially productive soils. The positive relationship likely stemmed from increased ANPP, rather than BNPP, which was possibly related to efficient resource-use and physiological/anatomical advantages of C4 plants. R H was negatively correlated with C4:C3, possibly because of changes in microclimate or plant-microbe interactions. It is possible that in potentially productive soils, C storage can be enhanced by favoring C4 over C3 grasses through increased ANPP and BNPP and reduced R H. Results also suggest that potential C storage gains from C4 productivity would not be undermined by a corresponding increase in R H.

Flexible foraging shapes the topology of plant-pollinator interaction networks.

In plant-pollinator networks, foraging choices by pollinators help form the connecting links between species. Flexible foraging should therefore play an important role in defining network topology. Factors such as morphological trait complementarity limit a pollinator's pool of potential floral resources, but which potential resource species are actually utilized at a location depends on local environmental and ecological context. Pollinators can be highly flexible foragers, but the effect of this flexibility on network topology remains unclear. To understand how flexible foraging affects network topology, we examined differences between sets of locally realized interactions and corresponding sets of potential interactions within 25 weighted plant-pollinator networks in two different regions of the United States. We examined two possible mechanisms for flexible foraging effects on realized networks: (1) preferential targeting of higher-density plant resources, which should increase network nestedness, and (2) context-dependent resource partitioning driven by interspecific competition, which should increase modularity and complementary specialization. We found that flexible foraging has strong effects on realized network topology. Realized connectance was much lower than connectance based on potential interactions, indicating a local narrowing of diet breadth. Moreover, the foraging choices pollinators made, which particular plant species to visit and at what rates, resulted in networks that were significantly less nested and significantly more modular and specialized than their corresponding networks of potential interactions. Preferentially foraging on locally abundant resources was not a strong driver of the realization of potential interactions. However, the degree of modularity and complementary specialization both increased with the number of competing pollinator species and with niche availability. We therefore conclude that flexible foraging affects realized network topology more strongly through resource partitioning than through focusing on high-density resources.

Contrasting Foraging Patterns: Testing Resource-Concentration and Dilution Effects with Pollinators and Seed Predators.

Resource concentration effects occur when high resource density patches attract and support more foragers than low density patches. In contrast, resource dilution effects can occur if high density patches support fewer consumers. In this study, we examined the foraging rates of pollinators and seed predators on two perennial plant species (Rudbeckia triloba and Verbena stricta) as functions of resource density. Specifically, we examined whether resource-dense patches (densities of flower and seeds on individual plants) resulted in greater visitation and seed removal rates, respectively. We also examined whether foraging rates were context-dependent by conducting the study in two sites that varied in resource densities. For pollinators, we found negative relationships between the density of flowers per plant and visitation rates, suggesting dilution effects. For seed predators, we found positive relationships consistent with concentration effects. Saturation effects and differences in foraging behaviors might explain the opposite relationships; most of the seed predators were ants (recruitment-based foragers), and pollinators were mostly solitary foragers. We also found that foraging rates were site-dependent, possibly due to site-level differences in resource abundance and consumer densities. These results suggest that these two plant species may benefit from producing as many flowers as possible, given high levels of pollination and low seed predation.

Taking the trophic bypass: aquatic-terrestrial linkage reduces methylmercury in a terrestrial food web.

Ecosystems can be linked by the movement of matter and nutrients across habitat boundaries via aquatic insect emergence. Aquatic organisms tend to have higher concentrations of certain toxic contaminants such as methylmercury (MeHg) compared to their terrestrial counterparts. If aquatic organisms come to land, terrestrial organisms that consume them are expected to have elevated MeHg concentrations. But emergent aquatic insects could have other impacts as well, such as altering consumer trophic position or increasing ecosystem productivity as a result of nutrient inputs from insect carcasses. We measure MeHg in terrestrial arthropods at two lakes in northeastern Iceland and use carbon and nitrogen stable isotopes to quantify aquatic reliance and trophic position. Across all terrestrial focal arthropod taxa (Lycosidae, Linyphiidae, Acari, Opiliones), aquatic reliance had significant direct and indirect (via changes in trophic position) effects on terrestrial consumer MeHg. However, contrary to our expectations, terrestrial consumers that consumed aquatic prey had lower MeHg concentrations than consumers that ate mostly terrestrial prey. We hypothesize that this is due to the lower trophic position of consumers feeding directly on midges relative to those that fed mostly on terrestrial prey and that had, on average, higher trophic positions. Thus, direct consumption of aquatic inputs results in a trophic bypass that creates a shorter terrestrial food web and reduced biomagnification of MeHg across the food web. Our finding that MeHg was lower at terrestrial sites with aquatic inputs runs counter to the conventional wisdom that aquatic systems are a source of MeHg contamination to surrounding terrestrial ecosystems.

Habitat loss alters the architecture of plant--pollinator interaction networks.

Habitat loss can have a negative effect on the number, abundance, and composition of species in plant-pollinator communities. Although we have a general understanding of the negative consequences of habitat loss for biodiversity, much less is known about the resulting effects on the pattern of interactions in mutualistic networks. Ecological networks formed by mutualistic interactions often exhibit a highly nested architecture with low modularity, especially in comparison with antagonistic networks. These patterns of interaction are thought to confer stability on mutualistic communities. With the growing threat of environmental change, it is important to expand our understanding of the factors that affect biodiversity and the stability of the communities that provide critical ecosystem functions and services. We studied the effects of habitat loss on plant--pollinator network architecture and found that regional habitat loss contributes directly to species loss and indirectly to the reorganization of interspecific interactions in a local community. Networks became more highly connected and more modular with habitat loss. Species richness and abundance were the primary drivers of variation in network architecture, though species compositi n affected modularity. Theory suggests that an increase in modularity with habitat loss will threaten community stability, which may contribute to an extinction debt in communities already affected by habitat loss.