Keywords

18.1 Introduction

Biodiversity loss is widely recognized as a crucial survival issue in society. This became clear when most countries of the international community signed the Convention on Biological Diversity (CBD) at the Earth Summit in Rio de Janeiro in 1992 (Brooks et al. 2002; McKee et al. 2004; CBD 2006). There is vast evidence that the loss of biodiversity is not only a biological and an ethical problem but it is also substantially a financial problem because important ecosystem services are lost on a global level (Mainka et al. 2005; Stern 2006). Many countries have recognized these facts and have implemented national biodiversity monitoring strategies to detect changes in biodiversity, usually substituted by monitoring species and species richness (Wilson 1992; Nakashizuka and Stork 2002). Some of these protocols recently started to accept a loss of precision for single species by assessing multiple species at the same time (rapid multispecies assessments). This is because resources to implement one monitoring system per species often are simply not available and multiple-species monitoring on a landscape level is much more resource efficient (Franklin 1993; Manley et al. 2004, 2005). The hope is that these approaches can help to provide information and manage well for the entire global systems. These systems are also more resilient against sudden changes in the focus of research interest, which may render more specific monitoring systems useless before they are even fully implemented (Watson and Novelly 2004). Some regional examples of such monitoring systems are the Multiple Species Inventory and Monitoring Protocol (MSIM) for National Forest System lands in the USA (Manley and Horne 2006), the Alberta Biodiversity Monitoring Program (ABMP) in Canada (ABMP 2003; Stadt et al. 2006), the Biodiversity Monitoring Switzerland (Küttel 2007), or the International Biodiversity Observation Year in Western Pacific and Asia (IBOY-DIWPA, Nakashizuka and Stork 2002).

These types of national monitoring systems can have at least two problems: Firstly, they are highly specialized for the area within the borders of the country they were developed for. Very often, these protocols can only be used in specific environments (for example, temperate forests and mountainous areas) but they are usually not applicable to ecosystems which might occur outside of the country of origin. Important political borders are (usually) clear and precise on a map. However, they tend to be artificial from an ecological perspective, while changes of biodiversity and biogeography are subtle and usually occur as gradients, often connected globally. Most of today’s threats to biodiversity, for example, global climate change , have influences which do not stop at political and administrative borders, neither do migrating animals nor ecological processes. The second problem is that even if the ecosystems in different countries are similar enough, very often the details of data collection and/or processing methods differ too much to compare the monitoring results from different countries. As a result, they are not allowing then for proper generalizations and create fragmented efforts that are inefficient and even can fail, whereas the state of the globe warrants just the best remaining options.

In the long-term view, the intention behind our pilot project is to develop a globally valid biodiversity monitoring system which delivers comparable results at achievable costs in every ecosystem and every country of the world, to be used for an efficient management. To this date, most habitats have not even been assessed once, and there is a considerable shortage of biodiversity monitoring on a global scale (Green et al. 2005; Dobson 2005). The core of these efforts is the sampling of taxonomic groups that are potentially living in almost every terrestrial ecosystem: plants, birds, and ground-living insects. Add-on protocols for other species groups with more limited distribution are also in need to be developed. As an example, this study here includes testing of one possible add-on protocol for butterflies, amphibians, and reptiles. It should be stated that such monitoring schemes are also in need of computational tools, as well as data management concepts that embrace the internet. Such workflows should be done in a smooth and combined fashion.

The main idea of this project here is to develop a relatively simple low-cost rapid biodiversity assessment and monitoring system, which is repeatable over time and aims at multiple species. Further, it offers multiple ways of analysis but with a rather high accuracy to avoid errors for good management advice. Furthermore, this system is supposed to be globally applicable and compatible with current data standards, so that the data from this project may readily contribute to ongoing global biodiversity initiatives (Group on Earth Observations (GEO) 2007; International Polar Year (IPY) 2008) and data pools like GBIF.org.

As a pilot study, data were collected with a set of methods in six diverse regions in the world in form of a biodiversity GRID. This chapter summarizes the experiences and results gained by testing a biodiversity GRID sampling methodology developed by the authors. The results are a valuable starting point for more detailed geo-referenced taxonomic and quantitative studies based on a valid underlying database. They provide crucially needed information for setting up sampling schemes with higher accuracy. For that reason, data are made fully available to the public and for investigators for their own assessment and for improvement as deemed necessary. Full metadata for the datasets have been created and uploaded to globally accessible NBII (National Biological Information Infrastructure) Clearinghouse website and can be found online at http://mercdev3.ornl.gov/nbii/. In the long-term view, such data are expected to be easily visualized and connectable to other datasets in public domains (Guralnick et al. 2007).

18.2 Methods: GRID Sampling Pilot Study and Methodological Advantages

18.2.1 Study Area

To document the validity of the method across a diverse set of habitats, data from six different study areas were collected in Costa Rica, Nicaragua, Papua New-Guinea, Russia (Sakhalin Island), and two locations in the USA (Alaska); see Fig. 18.1. The diversity of the ecosystems includes lowland and mountain tropical rainforest, tropical dry forest, boreal forest, and arctic tundra. The data were originally collected for a master’s thesis (Nemitz 2008), with substantive support and under the supervision of the second author.

Fig. 18.1
figure 1

Selected locations of GRID sampling plots carried out by the EWHALE lab

Full size image

18.2.2 Sampling GRID

For efficiency reasons, a systematic sampling approach was chosen (Cochran 1946; Olea 1984). First of all, an equally spaced GRID was implemented: 25 points were arranged in five rows and five columns in order to cover a consistent area but also to have a known spatial neighbor relationship among all plots, which is consistent with recommendations given by Ricklefs (2004). The distance between plots was 100 m, resulting in a total GRID size of 400 × 400 m. While the final GRID system ideally covers the globe systematically without intentional placement, for these initial studies the GRIDs were placed in a way so that roughly half to two thirds of the plots fell inside a forested area, and the remaining plots fell at the forest edge zone or inside the cultural landscape. This survey setup enables other studies on the same dataset to make realistic and representative statements about fragmentation effects and diverse habitats (instead of just homogeneous habitats). The only exception is the GRID in Barrow, northern Alaska, where naturally only one habitat type occurs (arctic tundra). Additionally, five points were randomly placed within the GRID to be able to model the influence of random patterns on the results and their spatial relations (Fig. 18.2).

Fig. 18.2
figure 2

Structure of the biodiversity GRID with 25 systematically selected plots, 5 randomly selected plots, and trapping webs installed at 4 plots (underlined)

Full size image

The coordinates of each plot were obtained from a regular handheld Global Positioning System (GPS) receiver. All plots as well as the access paths between them were marked with decomposing flagging tape to facilitate field recognition. A simple schematic map was drawn by hand for each fieldwork participant to ensure that plots are found when the GPS does not receive signals, as was the case in many dense forest settings with rugged terrain and steep valleys.

Trained taxonomists were not available, as they rarely are for many ecosystems worldwide and without larger budgets. Instead, here we followed a parataxonomic approach: Species details were documented, and all notes regarding the observed species were made as precisely as possible, although most of the observers were not specifically trained in tropical ornithology or entomology (where species identification tends to create the biggest problems). Data collection followed the motto “the more detail the better,” but it was not intended to exclude any data because of lacking taxonomic details. If the observer did not readily know the correct scientific name of a specimen, a common name or, in a lack of knowledge of a common name, a short description was noted. Photos were often taken if possible. This original field note is referred to as the “narrative name” of an observation respectively of a species. This resulted in good abundance and occupancy estimates for individual grid plots, but in less detailed taxonomic data. Such is the characteristic in rapid biodiversity assessments on shoestring budgets. However, it allows for a first impression and provides detailed information for deeper investigation, if desired, for precious wilderness areas. This type of rapid assessment additionally serves as a pilot study for further assessments and fine-tuning by the global public. In the present study, the focus lies on a spatial global coverage, instead of a local detail. This remains an efficient concept because all plots can be revisited due to the geo-referencing and inquired for more detail.

18.2.3 DISTANCE Sampling

In the ideal case, the applied monitoring protocol should result not only in information about the presence or absence of species but also in an estimate of population size explicit in space and time. The DISTANCE sampling approach uses the concept of a detection function based on distance of the observed object from the observer to estimate population density (Buckland et al. 1993, 2001). It plays a central role in this study and is used in a number of ways.

At each of the 30 plots (25 systematic and 5 random), a 5-min-point transect DISTANCE sampling counts for birds were conducted within a 360° angle (Buckland et al. 2008). A short settle-in period of 1 min was granted prior to the counting in order to allow for the snapshot character of the DISTANCE sampling, especially meeting the assumption that the presence of the observer does not introduce bias by causing responsive movements of animals. Following common practice, the point counts took place only in the morning between 5:30 and 10 am. Birds are known to show higher activity at this time, which generally increases detectability and maximizes inventory accuracy. Each bird seen or heard was noted, including an estimate of the radial distance from the observer. Double counts were avoided by the observer’s attention and by the relatively short counting period.

The second application of DISTANCE sampling used in a trapping web (Parmenter and MacMahon 1989). Seventeen pitfall traps with a diameter of 9 cm each were arranged in a DISTANCE sampling trapping web design to estimate ground-living insects (as described for instance in Buckland et al. 2001, p. 216 ff.). This sampling method is very labor-intensive and could not be implemented at all 30 plots. Thus, four of the plots were systematically selected in the GRID area to capture the general patterns of species and abundances without edge effects; they were located at grid points B2, D2, B4, and D4 (underlined in Fig. 18.2). Trapping webs were usually checked every 24 h, and records were taken every 48 h. Between check dates the cups were emptied without recording to avoid correlation in time between trapping events, and to obtain spatially independent results.

The third application of DISTANCE sampling was an add-on sampling protocol using the DISTANCE sampling line transects, conducted at each of the 30 plots. Transects with a length of 10 m and traversing the individual plot at its center were surveyed to estimate numbers of butterflies, amphibians, and reptiles. These line transects allow for an alternative assessment of the point transects.

18.2.4 PRESENCE

DISTANCE sampling point counts for birds and trapping webs for ground-living insects were repeated three times. These repetitive visits further allow for an analysis with the software PRESENCE, which gives an estimate of general occurrence of a species in the area in a point-based sense for the time period covered. PRESENCE generates a detection function based on multiple visits under the assumption that the population is closed, meaning that no animals leave or enter the area of interest between several visits. Repetitions were not realized for the add-on protocol for the DISTANCE sampling line transects and consequently such data were not used for PRESENCE.

Abundance and population density are probably the most important basic parameters in inventories and for population dynamics (Krebs 2001). Whenever possible it is intended to get true abundance estimates for each species, corrected for imperfect detection of individuals with different methods (Buckland et al. 2001; MacKenzie 2005a). However, with a standardized rapid multiple-species protocol this is not always possible. Especially species with large territories are often difficult to monitor on a relevant eco-regional scale (Manley and Horne 2006). Therefore, the probability that an area is occupied by a species, known as occupancy or Psi, is also estimated. Occupancy is the simplest level of interest (Hill et al. 2006), which gives much less information than abundance or density, but still has implications for wildlife management. At the same level of precision it can usually be obtained at lower costs than abundance estimates (MacKenzie 2005b; Bailey et al. 2007). Clearly, occupancy is not the prime goal in this project, but it is accepted as better-than-nothing baseline information widely used elsewhere. It also allows matching up with DISTANCE sampling and other studies underway elsewhere worldwide and can contribute to a global database.

18.2.5 Environmental and Vegetation Covariates

GPS coordinates and a short description of the ecosystem were taken at each plot as well (for example, pasture, forest interior, and forest edge with a verbal description). Further, a plot photo and a canopy photo were taken with a digital camera to allow for a general impression of the area and also to allow for an analysis of the basic vegetation structure and light conditions in other studies on the same dataset, e.g., remote sensing investigations and leave area index (LAI). Vegetation height and diameter at breast height (DBH) were recorded for all trees within 5 m of the plot center; in most cases, nearest neighborhood measures (center and four nearest trees) were also taken. Estimates were recorded regarding canopy cover percentage, understory cover percentage, shrub cover percentage (at 1.35 m height), bare soil percentage, duff coverage percentage, leaf browsing percentage, and number of flowers visible. The thickness of epiphytes , hemiepiphytes, mosses, and lichen was reported as categories (none, low, medium, high). Presence/absence of identified plant species or plant families was also written down, as well as any detected animal tracks (e.g., land crab holes, large mammal tracks, etc.). Detailed species lists and the full protocol are available from the authors.

18.2.6 Predictive Modeling and Data Mining

Despite the fact that ecological data often are nonlinear, not normal, interrelated, and multidimensional in nature, lack of computing power (hard- and software) and science culture in the past has limited predictive modeling in ecology to generalized linear models (GLMs). Recent developments in computer technology and steep price declines of equipment and communication are now relaxing these limitations (Bauldock et al. 2001; Drew et al. 2011). Current research using machine learning algorithms in ecology are rather promising and seem to clearly outweigh the traditional GLMs in convenience, speed, inference approach, and accuracy (e.g., Huettmann 1999; Prasad et al. 2006; Magness et al. 2008). While machine learning proves very valuable to mine data, predictive modeling is a tool to achieve global and generalizable information about biodiversity distribution conveniently (Breiman 2001; Elith et al. 2006). It has the ability to process all available environmental data to analyze the effect on general biodiversity patterns (Faith 2005). It has also been shown in numerous cases that well-constructed models often show a much better performance and higher consistency in population estimations and habitat modeling than do expert opinions (Pearce et al. 2001; Yamada et al. 2003). Predictive modeling using data mining can also handle a large variety of data since there are virtually no requirements of parametric assumptions to be met (this makes it different from classical parametric applications). Additionally, these machine learning algorithms have less problems interpreting much, noisy or sparse information (Elith et al. 2006), partly because interactions between variables are included in the analysis (Craig and Huettmann 2008; Magness et al. 2008). The predictions have the advantage that they can be tested for generalizations and in a quantitative form. That way, the research emphasis is shifted away from p values and Akaike’s information criterion (AIC) to receiver operating characteristic (ROC) and similar metrics (e.g., from the confusion matrix; Drew et al. 2011).

18.3 Results: Successful Integration of Methods

One table each was prepared for observations, plot information, and visit information. This can be done in a spreadsheet such as with OpenOffice or MS Excel. It was imported into MS Access database (Microsoft Office Access 2003 SP3). Additionally, an attempt was made to derive as much taxonomic information about the narrative names noted in the field as possible. All information was cross-checked and taxonomic validity was verified with online information provided by (Integrated Taxonomic Information System on-line database) and imported into the same MS Access project database.

All files for data analysis with other software were derived through individual queries from this central project database. According to the method of plot selection, the bird count and DISTANCE transect data were separated into a systematic set (25 plots) and a random set (5 plots). The random set was run as an assessment/test dataset for the systematic set. Although small sample sizes were expected to increase standard errors, we found for our grids that 16 % (5 out of 30 plots) of data for testing can be a feasible number to express variances, or at least for indicative trends in the data with an alternative test set. To increase the number of observations for analysis, these two sets were also pooled (25 + 5 = 30 plots) and additionally analyzed together (systematic set and random set) for a more representative picture of the GIRD area. Bird count data were further split into aural and visual detections because the type of detection is known to greatly influence detection probability (Marques et al. 2007). If an observation was detected both visually and aurally, then it was allocated to the visual dataset. This resulted for instance in a total of nine datasets for each of the six study areas: five bird count datasets (systematic, random, pooled, visual detection, and aural detection); three DISTANCE transect datasets (systematic, random, and pooled); and one trapping web dataset.

Following the machine learning paradigm, each dataset was analyzed with Salford Systems Random Forest, version 1.0 (Breiman and Cutler 2005). Random Forest is a machine learning algorithm using sets of classification and regression trees for data mining and to build powerful predictive models (Breiman 2001). The value of such analysis methods for numerous ecological applications is increasingly recognized (Elith et al 2006; Drew et al. 2011). Random Forest settings all remained as “default” (500 decision trees); only the number of predictors considered for each node was set to the square root of the number of used covariates (rounded up). Different models were constructed using plot vegetation data and presence of other narrative name observations as covariates. The best model was selected by the highest ROC integral (Fawcett 2006).

A full DISTANCE sampling analysis was run for all narrative names with at least 20 observations using DISTANCE 5.0 Release 2 (Thomas et al. 2006). This is considerably lower than the 60–80 observations usually recommended, so inconsistencies resulting from small sample size and with limited detections have to be considered especially for those narratives with less than 60 observations. Among all models for a narrative name, the best ones for conventional DISTANCE sampling and multiple covariate DISTANCE sampling were selected by visual assessment of model fit and AIC.

Occupancy estimations were derived by the software PRESENCE (Hines 2006) for the same narrative names as selected for the DISTANCE analysis. For each narrative species name one model was run assuming constant detection probability and second assuming different detection probabilities for each visit. Additional runs were conducted adding each of the site and visit specific covariates used in DISTANCE. Observation specific covariates were left out because of the different structure of analysis in PRESENCE, which takes only site- and visit-specific covariates into account. The best model was selected using AIC.

The amount of information gathered through each GRID was enormous given the relatively short sampling time of 3–14 days for the GRIDs. Table 18.1 displays the number of narratives for which valid predictive models were retained through Random Forests analysis (ROC value > 0.5); as well as the number of narratives for which abundance estimates through DISTANCE sampling and occupancy estimates through PRESENCE analysis were gained. One-hundred-sixteen valid predictive models were gained through only 12 weeks of sampling. Forty-five of those also delivered abundance and occupancy estimates.

Table 18.1 Total number of narratives analyzed by the study area, sampling method, and analysis method
Full size table

18.4 Discussion: The Pros and Cons of the Systematic Multiple-Species Biodiversity Sampling GRID

An overview of all identified challenges and potential ways to address them is included in Table 18.2 below. While the quantity of data was so immense and allowed to run a number of informative and sophisticated models, the actual quality of the data was in some cases not very good nor exhaustive; especially confidence interval ranges for DISTANCE sampling could become big. It is assumed that larger field datasets will enable stratification of analysis by habitat, etc. and help solving this problem. Larger amounts of data would also open the possibility to split the data into different sets and analyze those separately if this is assumed to be beneficial for analysis precision, as is probably the case for aural and visual detection of birds. PRESENCE analysis showed a high tendency for perfect probability of occupancy (Psi = 1.0), especially for trapping web results. It is expected that this problem can be solved with more field data, especially through larger numbers of plots and more GRIDs per study area. The effective PRESENCE sampling size for each narrative in each study area was 30 per visit, because this is the number of plots, and because detection/nondetection is entering the analysis only once per plot and species. DISTANCE analysis on the other hand benefited from spatial repeats because all distances of all narrative detections at each plot were used for the modeling of the detection curve within the study area (Fig. 18.3).

Fig. 18.3
figure 3

Habitat picture of plot D3 in La Suerte Costa Rica

Full size image
Table 18.2 Table of identified challenges and potential ways to address them
Full size table

GIS Geographical Information System Violation of assumptions necessary for DISTANCE and PRESENCE analysis has not been tested explicitly in this study done for rapid assessments. Repeats of sampling at each plot have been done within few days, so that extensive movement of animals is unlikely (but not impossible). Movement of animals in and out of the sampled area would be problematic for PRESENCE analysis, which assumes a constant population. However, obtained results did not imply that this is a problem in the study at hand. Results of models with constant detection probability were generally very close to those with survey-specific detection probability. DISTANCE assumptions were a bit trickier. Recent research suggests that the assumption of perfect detectability close to the observer, at a distance of 0 m (g(0) = 1), should be vigorously tested in each study (Bächler and Liechti 2007). This is often impossible in many cases given the number of study sites and different ecosystems, the short time and the budget constraints, and the reality of biodiversity sampling (similar to what Bächler and Liechti found). Another assumption crucial for DISTANCE analysis is the required high precision of distance estimates. In this study, all distances for point and line transect observations were estimated by the observers without technical distance measurement tools. This is because technical devices were not available and would have failed in some of the environments, especially for aural detections and in remote rainforest detections. Besides, the GRID system with distances of 100 m between plots and additional markers at 50 m between plots proofed to be extremely helpful for the observers to validate their estimates. Because we only deal with detection distance estimates of 0–50 m and because DISTANCE allows to “bin” estimates for a valid analysis, the distance estimation problem in the field and when using GRIDs remains small.

Generally, both PRESENCE and DISTANCE analysis are supposed to estimate trends in animal populations. Joseph et al. (2006), for example, directly compare the two methods and give a recommendation for which types of species which type of sampling can be used. They come to the conclusion that “Abundance surveys were best if the species was expected to be recorded more than 16 times/year; otherwise, presence-absence surveys were best” (Joseph et al. 2006). Support in the study at hand for the regular interchangeability of these two methods is ambiguous. The probability of detection as estimated by DISTANCE correlated is slightly negative with the probability of detection as estimated by PRESENCE. The two might be mathematically different in fundamental ways while only sharing the same label, but this hypothesis was untested in detail. One should keep in mind that PRESENCE is essentially based on a one-dimensional point estimate, whereas DISTANCE is inherently two dimensional. It remains unclear how such estimates across dimensions can be done in a geographically meaningful way, e.g., not addressed by Joseph et al. (2006), nor by the providers of either analysis software. Detailed comparison of these two different p estimates and their methodological differences is beyond the scope of this work. It is also imaginable that the strict GRID design, which is not tailored to gain high-precision results for any of the two methods, does in fact work in favor for one of the two. We found though that coarse and informative estimates can be gained when rapid assessment is the goal.

The comparison of Random Forest ROC values for data from systematic and random plots showed relatively small differences. In most cases, the ROC values for data from systematic plots were very close to the pooled ones, which is to be expected as survey effort for this data had a share of roughly 80 % at each study site (25 plots compared to 5 plots). In a few cases, the ROC values of data from random plots were surprisingly high and strong. In these cases, the most likely explanation is that the input data were very clean, allowing for powerful models. It raises some interesting questions about apparently very good data qualities in GRIDs free from much of the traditional problems such as lack of independence, interactions, and autocorrelation, for instance. Random Forest has no tendency to overfit data with larger numbers of covariates for small sample sizes (Breiman 2001). In the DISTANCE and PRESENCE analysis, the differences between the estimates from random and systematic plots were generally low; with a few exceptions the results usually were very close and similar.

The comparison of results from visual and aural datasets implied that for a detailed analysis the individual species’ biology has to be taken into account. In many cases, the analysis results were similar, but there were also many narratives for which the proportion of observations by one of the two means of detection (aural/visual) as well as the gained results differed considerably. In some cases, the larger number of observations was within one type of detection, while the ROC value derived through Random Forest analysis was larger using the other type of detection. There were also a number of narratives for which the pooled data resulted in considerably higher DISTANCE abundance estimates than the separated datasets, which seems to be implausible at first sight. A multitude of biological reasons can explain such differences though, for example, clustering, gender dimorphism in singing behavior (affecting aural detectability), gender dimorphism in coloration of plumage (affecting visual detectability), differences in general behavior/secrecy by age or gender, different effect of environmental covariates on (especially visual) detectability, etc. It is highly recommended to separate between aurally and visually collected bird data, although this was not possible continuously in the study at hand because of relatively low sample size. Generally speaking, many biological aspects can also hinder the identification when observers are lacking specific experience, for example, species’ gender and age dimorphism in appearance and behavior. On the other hand, it is well known also for more traditional survey approaches that “misidentifications at the species level are common” (Guralnick et al. 2007). It can be argued that lower precision in taxonomic identification is a minor issue compared to the analysis methods and results offered by the GRID, much as lower estimate precision discussed above. In addition, it has been found that a feedback system to integrate observers experience with the sampling methods in different ecosystems is recommended, so that unforeseen ad hoc changes of the sampling in the field can be reported back and considered for the analysis.

A major issue with the data as collected for this pilot study is the large number of observations with subjective species descriptions. This is a common problem in the tropics and when following a parataxonomy approach. An observation noted as “tiny ant” may be a “small ant” for the next observer, thus real monitoring of trends in time by visiting the same plot several times might prove difficult, especially when different observers are surveying. There is no immediate solution for this problem caused by the parataxonomic approach, because even if time and financial resources allowed for an extensive training period prior to sampling, for many regions and ecosystems qualified trainers, parataxonomists and literature would still not be available, especially not for more than one taxonomic group. (Keep in mind that classic taxonomy is getting decreasingly funded these days, nor that a good and sustainable business concept really is in place how to address the taxonomic problem; see also in Nemitz et al. 2012.) However, when considering that no other relevant data exist, in the interim, such approaches will help to further fine-tune sampling efforts in the future. Overall, we consider less the detailed taxonomic identification a problem when we are facing wholesale losses of entire ecosystems and species; instead, the global management problem should be addressed here first and foremost (see, for instance, Earth Stewardship http://www.esa.org/esa/?page_id=2157), and how robust monitoring schemes can support and feed into best-possible management actions. And at least for aural detections of bird species automated identification methods are under development (Brandes 2008). These add other technological and financial challenges, but those are expected to be smaller than those from providing adequate training for a large number of observers. It is far easier and more reliable to teach a bird song to a computer and scale up digitally than to train human observers one by one and with a consistent accuracy (as computers can provide). Thus, it would still be a costly technology; but it is also expected to be a cost-effective and rapid method with a consistent outcome over the long term. Similarly, automated approaches for identification of insect and amphibian species and visual bird detections would be extremely helpful for further development of the biodiversity GRID approach, but are to the authors’ knowledge not (yet) feasible and really in sight.

In our work, we found a considerable difference between analysis at the narrative level and the analysis at higher taxonomic levels. Pooling at a higher taxonomic level provided a larger number of observations, but did not automatically imply a high ROC value. This indicates that pooling by taxonomic classes, like a family or an order, is not a recommended way to receive better datasets. Details depend on the wanted inference from the rapid assessment. Differences on the biological level, like habitat requirements, can be huge between two species belonging to the same taxonomic class, especially when looking at such diverse ones as Passeriformes or insects. The benefits gained through the analysis of higher taxonomic classes were rather low; the range of DISTANCE confidence intervals was not much reduced. In the PRESENCE analysis, the pooling caused similar problems for point transect data, as did the small number of plots for trapping web data: low variance between plots, because some bird belonging to the order Passeriformes, was detected at almost every plot, and therefore it had universally high occupancy estimates (Psi = 1.0 for all analyses at the biological order level). The assumption that higher taxonomic categories can be used as valid surrogates for rapid assessment and monitoring of species diversity gains no support from this study. This requires more attention, still.

The overall biodiversity GRID approach has some promising aspects, especially regarding global coverage, cost–benefit, open access database, workflow, and avoidance of sampling bias common in many population studies (e.g., access and roadside bias, Kadmon et al. 2004). However, some aspects of the sampling methods can be discussed: one being the differences in taxonomic knowledge and identification skills between different observers. This was already observable in this relatively small pilot study and will probably grow to a major challenge for a truly global GRID system (designed, or opportunistically). Another important question is the importance of time of the year when the survey is actually carried out. Buckland et al. (2008) recommend for instance the breeding season for bird surveys. The opportunistically selected study sites in this project were not always ideal to meet this criterion and for the Arctic versus the tropics and with longitudinal gradients, for example. Globally, one has probably to settle for presence/absence patterns instead of biological seasons. To conduct studies exactly at the best time of the year will provide a considerable planning challenge in a global project along a latitudinal and elevation gradient. Primarily, goals and objectives need to be decided upon.

Another issue with the covariates is, for instance, an extremely low variation of the study site such as in Barrow, arctic tundra. Most of the covariates gathered in the field can be expected to correlate with the presence/absence and abundance of trees (highest tree, canopy cover, duff cover, plant species, presence/absence, etc.). This worked well in environments where the landscape was diverse in the GRID, including pasture, wetland, and different kinds of forests. It proved to be more problematic though in a relatively homogenous ecosystem like arctic tundra. Other estimates which would not be visible in dense forests, like diameter of trees or distance to the next lake, were added to the protocol in this case. All in all, the predictive modeling still worked well at this study site, but the recommendation resulting for global monitoring is to gather additional data with higher small-scale variations at each plot, especially pedological data available in all terrestrial ecosystems. Another idea is to use data from other surveys or other publicly available geo-referenced data, for example, distance to roads from available maps or slope gradient and aspect from high-resolution digital elevation models (DEMs). The possibilities this approach involving proxies opens up are amazing. The only thing to be kept in mind is the GRID spacing: A satellite picture resolution with pixel sizes of 2 × 2 km will not provide adequate covariates for modeling of a GRID with much smaller spacing (e.g., 100 m, as in this study). DEMs with a 25-m pixel size are suggested. The future will likely see much higher-resolution data, and here we provided a first approach to start from.

18.5 Conclusions

This study demonstrates three analysis methods which carefully designed biodiversity GRIDs can offer to ecological research and assessment. The available analysis options from these types of data are by a magnitude more. Especially autocorrelation and interaction issues between plots and questions regarding fragmentation and change over time come to mind. Adding that all sort of spatial data, especially also those resulting from remote sensing, can be connected to these GRID data and that continuous efforts exist to make research data publicly available, the possibilities to conduct relevant analysis from GRID data are quite enormous and still not fully explored and used.

The results shown so far are more than promising. Three of the most sophisticated current methods in ecology are already involved: Random Forest as a powerful data-mining tool to construct predictive models that generalize well, DISTANCE sampling for the estimation of population abundance, and occupancy estimation with PRESENCE to gain information for species with low sample sizes. All of them are globally used by now. Many results lack high precision for each single species but are promising regarding first snapshot assessments of multiple species anywhere and in concert. Holistic approaches are urgently needed to improve cost-effectiveness of ecological research and effectiveness of ecosystem management (Magness et al. 2008), while at the same time more precise study designs have their place in evaluation of known risk species for which more detailed population estimates are necessary. Challenges have been faced by the current study (see also Table 18.2), but those are to be expected when working on a global scale. A number of recommendations could be given to improve the involved methods: The most pressing next step is to sample more study sites and build a stronger database for more globally representative cases. It is unlikely that the biodiversity GRID approach will be accepted and implemented by many country governments within a short time frame, so another urgent point of development is the connection with other and existing data sources. Additionally, the project would gain from development of a meta-software with the ability to batch several other software solutions for a workflow, and from a closer investigation of the comparability of DISTANCE and PRESENCE results as well as detection probabilities estimated by those two programs and with an underlying GIS.

The large dataset produced through this type of survey resulted in a treasure for data-mining approaches and pilot study data for the more detailed study of species of special interest. A promising idea, and that to the authors’ knowledge has not been tested, would be to use predictive modeling to identify study sites for adaptive sampling of species of special interest, optimizing precision per study effort (as described by Pollard and Buckland 2004, for instance). The management link remains equally untested, yet it should be explored more for a good implementation of data, buy-in, and in an adaptive management framework. This could prove especially useful for the monitoring of endemic species with small regional distributions, which might be sampled inadequately by a global biodiversity GRID system, depending on plot density. Distances of 100 m between plots seem to be ideal and feasible in the field, but can probably not be achieved on a global scale yet. Assuming a global land area of about 130 million km2 (without Antarctica), this would result in roughly 13 billion plots. Increasing the distance to 500 m would still add up to 2.6 billion plots, while 5 km distances as have been used by Magness et al. (2008) in Alaska would result in 260 million plots! This sounds huge at first, but political will built on economic and social considerations clearly decided to protect biodiversity , while so far the necessary actions to do so are lacking. Optimized sampling of just well-defined strata is another method to limit meaningful sampling to less plots and effort, and within a reasonable budget and reliable outcome. Information is essential for conservation and protection, while “the extent of global data gathering underway is inadequate to meet the challenge set out at the WSSD in Johannesburg” (Green et al. 2005) World Summit on Sustainable Development (WSSD). To act accordingly is certainly costly but so is the cost of doing nothing and of restoration, with the latter one often being even higher than the combined costs of monitoring and conservation (Dobson 2005). The decision ultimately boils down to another question: How valuable is reliable knowledge (MacKenzie 2005b)—and what is it used for, and by whom?

Despite the global relevance and scope of the project it was not funded by relevant funding bodies. This resulted in very few sampling sites which in addition had been selected opportunistically: GRIDs were usually installed in areas of ongoing other research. The coverage and diversity were still extremely high so that problems of this way of selecting are not expected. But with better funding a more careful design could have been implemented, based on this work completed here. Probably more important would be a higher number of study sites, since effectively the sample size for the whole globe is currently six pilot study areas. Another very important area definitely needing attention is the development of a similar approach for aquatic or partly aquatic ecosystems, which so far, have been ignored completely in this work despite their importance for biodiversity. Including atmospheric metrics would even be more complete.

One reason for the lack of funding so far could be the visionary approach taken. The intention to have a globally applicable multiple-species monitoring and rapid biodiversity assessment scheme is contrary to the recommendation of many scientists to aim for micro-interpretation and maximum precision by designing each survey individually for each species of interest (Pollock et al. 2002; Mackenzie and Royle 2005; Bailey et al. 2007). The argument against this point of view is probably cultural and that it is ultimately more cost-effective and useful to aim for several dozens of species estimates with a precision of ± 80 % (or similar) than to aim for only one species estimate with a precision of ± 5 %. It has also been shown that the assumption that information of single species can serve as surrogate information for biodiversity in general is often not valid (cf. also van Jaarsveld et al. 1998). Manley et al. (2004) make a point that history of ecological research is rather dubious in some cases, which can also result in favoring multiple-species surveys: “Any effort that relies solely on a small set of indicator species will be subject to skepticism given the history of misuse, overuse, and poor performance of the indicator concept.”

In conclusion, biodiversity GRIDs are an important step into the direction to fill holes in global biodiversity information for conservation and management in a cost-efficient and sustainable way. The challenge to make this approach work is to move political decision makers and for a culture to act according to their declarations of intent and public needs to protect and improve biodiversity and ecological processes. Ideally, this should be done proactively and with supporting institutions and their agents.