Collected bee specimens were identified to genus both by author KC and with correspondence by bee expert Robbin Thorp , and additionally, in consultation with personnel and comparisons with collections at the UC Davis Bohart Museum . Plant identification was aided by correspondence with former UC Davis Arboretum Director of Horticulture, Ellen Zagory.Data collection was conducted weekly, but compiled into monthly data aggregations to minimize the possibility of sampling omission errors, such as variable detectability . Using this method, it was more likely that a greater magnitude of rarer associations were observed , which is advantageous for a study such as this, seeking to explore the relationships between bees and the plants they utilize. This data compilation method was selected after reviewing bee and other pollinator field research methods as well as much personal trial and error in field and personal correspondence with expert Robbin Thorp. From previous experience we determined that monthly walks produced significantly less association data than compiled weekly walks. Thus, we found that weekly data collection was best for observing ephemeral bee-to-flower foraging associations and monthly aggregations were most effective in understanding bee foraging and flower bloom times . A compiled monthly time step was primarily used for this study as it is a common standard protocol in both bee foraging and plant phenological records, flower harvest buckets such as field guides. As an example, Andrena was seen foraging in the Mary Wattis Brown garden on Ceanothus two weeks in a row.
This association was counted once for the month, not twice, when recording monthly association data. We determined two criteria for measuring a plant’s successful performance, including how many bee genera were attracted to a plant and also, the strength of a bee-to-plant association, with demonstrated repeat foraging events representing stronger associations. Additionally, we sought to determine forage plants utilized by bees which were not included in Table 1. Furthermore, we sought to determine characteristic trends among utilized forage plants, for example, whether they were native or not, and if not, what region of the world they originated from. The analyses were completed in MS Excel in an effort to identify plants missing from the current literature that hold potential for hosting bee foraging and, thus, provide habitat value.Initial investigation into potential pollinator plants revealed that 96 of the 134 of plant genera from Table 1 were included in the Arboretum’s plant collection maps. This indicated that the Arboretum’s records included many of the predictive plants for bees, allowing us to test the majority of the bee-to-plant associations from Table 1. We sought to evaluate to what degree the Arboretum geodatabase plant presence or absence was accurate. We were uncertain about the absolute accuracy of the Arboretum maps, as some new planting projects had taken place since the geodatabase had been completed. We were also interested in studying which weedy plants were used for pollination which were not included on the maps.We used two approaches to assess accuracy of the existing information foraging matrix. Model success is defined here as correct prediction of plants utilized by bees for foraging. First, we looked at how well Table 1 correctly predicted bee foraging overall in aggregate.
This relatively coarse method examines which plants, regardless of bee genus, were successfully both predicted and observed as floral resources. Next, a more precise investigation into the 1:1 association relationships between bees and their forage plants was performed. This statistical testing approach determines the accuracy to which the predictive Table 1 plants were utilized. The existing literature foraging matrix constructed in section 2.2 was validated by compiling field observations made in section 2.5 to determine its efficacy. There are three potential outcomes from this assessment: a correctly predicted presence , an omission error , and a commission error ; however, it should be noted that no correctly predicted absences are possible to assess in this study because the existing literature lists do not designate known absences . This makes traditional assessments of model accuracy using test statistics from a confusion matrix impossible, such as the Kappa statistic .Despite this limitation, it is possible to assess “sensitivity,” also known as the “true positive fraction” from the correctly predicted observations and the “omission rate,” or also known as the “false negative fraction”; these two measures are inversely related and sum to 1 . The third possible outcome is a metric of commission error which assesses the false positive rate. Each of these three metrics will be further described below. In the first case, if a known bee genus is observed in the field that is using a known plant genus this is considered a “correctly predicted” occurrence . For each bee genus this metric is calculated by dividing the count of literature plant genera correctly predicted by the count of all plant genera observed to be used in the Arboretum by that respective bee genus. This yields a “sensitivity score” or true positive fraction. In the second case, if a known bee genus is observed using a plant genus not on the literature list, this is an omission error. The omission rate is calculated by dividing the count of all additional plant genera observed to be used in the Arboretum by the count of all plant genera observed to be used in the Arboretum by that respective bee genus.
This yields an “omission score” or false negative fraction . It should be noted that the sensitivity and omission scores have the same denominator. In the third case, if a known bee genus is not observed to use a known plant genus that is present in the Arboretum, this is a commission error. In other words, the list predicts the bee genus to use the plant, but it is not observed. The commission rate in this study is calculated as a percentage by dividing the count of literature plant genera not observed to be used in the Arboretum by the count of all plant genera from the literature list in common with the Arboretum and multiplying by 100. There are 38 plant genera on the literature list that are not present in the Arboretum and therefore those plant genera are excluded from the error assessment. Finally, to assess the significance or model independence for each bee genus for all observations, a chi-square test was performed on each respective bee genus model result to assess observed versus expected values. This model independence test was conducted using CHISQ.TEST function in Microsoft 365 Excel. This test returns the probability of whether the model could attain the value of the chi-square statistic by chance alone under the assumption of independence. Values for p range from 0-1 and low values of the test statistic indicate independence. The degrees of freedom were calculated by subtracting 1 from the total number of columns used in each respective bee genus model.The completed presence-only bee-to-plant foraging matrix , derived from the literature, contains 23 bee genera and 134 plant genera. Of the 23 bee genera on listed on Table 1, 22 were observed in the Arboretum as well as five additional native bee genera that were not on the list. The only predicted native bee genus not observed in the Arboretum was Colletes, which had a singular association with just one plant genus, Solidago, round flower buckets which is found in Arboretum. In this case most likely either Colletes populations are too disjunct to access the floral resource or there are other lacking resource attributes which prohibited Colletes from using the Arboretum as habitat.The completed observed results of the bee-to-plant foraging matrix contains 27 observed bee genera and 297 observed forage plant genera. Table 2 differs from Table 1 in that results recorded the redundancy of the weekly foraging associations, demonstrating the relative strength of each bee-to-flower association throughout the year. A significant finding of this research is that more than three times the unique bee-to-plant foraging associations were observed than predicted . However, it is clear from Table 2 that plants varied considerably in terms of relative attraction . Appendix 2 shows a complete record of all bee genera predicted versus observed foraging.Observation data were summarized to show the annual pattern of association activity by garden . Bee foraging was well supported by the novel Arboretum plant communities. A full distribution of bee-to-plant associations by garden and month can be seen in Table 3.
Two gardens substantially out-performed all the rest: the Mary Wattis Brown native plant garden and the All-Stars in the Ruth Risdon Storer garden. The plants in each garden supported large numbers of bees, but there were notable differences in function over time. While native plant garden bee foraging peaked in May, the non-native garden peaked in August . Floral resource timing differences accommodate different seasons of bees, who also exhibit staggered emergence and activity months. Additionally, as plants in the native garden often desiccated and rested for the hottest summer months, many of the non-native plants continued to bloom, persisting to provide plentiful floral resources through the hottest months and even fall for summer and fall bees.We examined if Table 1 bee plants in the Arboretum’s map records correctly predicted foraging by bees. As stated in section 2.6, 96 of the 134 predicted plants were included in the Arboretum’s plant record maps. Of the 96 predicted matrix plants which were also in the plant maps 70 were actually used for forage. Wholistically, predicted foraging plant presence was highly correlated with a successful foraging utilization. Within the 84 of 96, or 87.5%, beeto-plant matrix association plants found in the Arboretum plant collections were foraged on by bees, thus, the success rate of the aggregate model indicated high correspondence. The majority of the plants stated to be in the Arboretum geodatabase maps were still present and also used by foraging bees. In total, 84 of the 134 predicted plants, were utilized by bees for foraging. In other words, though 70 predicted plants were also used for foraging and also confirmed on the maps, 14 additional predicted, but unmapped plants , were utilized for bee forage. The majority of Table 1 plants expected for bees were on the maps and 70 out of 84 plants used for forage . This high indication of map accuracy combined with the confirmation of bees foraging on the expected plant list seemed quite promising. Overall, it seemed the habitat relationship model, combined with existing habitat maps, were quite accurate to aid in making predictions in bee foraging habitat use as a whole . Meanwhile, bees were found to forage on many plants not predicted per the Table 1 matrix. Of the “unexpected novel” observed plants, 258 were on the Arboretum maps ,while only 39 forage plants were not on the maps. Interestingly, this is very similar to accuracy percentages of results above . This infers that map records were consistently accurate at providing foraging plant locations and subsequent pollinator association. While the bee-to-plant matrix is predictive of bee foraging the majority of the time, there are a variety of ways to analyze the matrix’s success. The Arboretum mapping accuracy omitted new or weedy plants and therefore some associations seen in Table 2. While these initial results above seem promising, when a more precise analysis is done below, it becomes clear that the individual bee genus models were not as predictive for foraging associations.We analyzed each bee genera by their predicted versus actual foraging data. Each bee genus was compared to the predicted plants it ought to have foraged on versus the observed data. The error analysis matrix and model independence tests presented in Table 4 show the results for each respective bee genus observed in the Arboretum. Error results are reported only for those bee genera listed in Table 1 . Table 4 breaks down the counts relevant to calculating the three aspects of error assessment for this study. The overall average true positive fraction for correctly predicting bee genera in the Arboretum was found to be 0.14 and likewise, the overall omission error rate, or false negative fraction, was found to be 0.86 . The overall average commission error rate for all bee genera was 47.8%, meaning that nearly half of the plant genera reported in the literature that bees are reported to use were not observed to be used in this study.