Ecological Indicators 87 (2018) 209–260
Contents lists available at ScienceDirect
Ecological Indicators journal homepage: www.elsevier.com/locate/ecolind
Research paper
Using dimension reduction PCA to identify ecosystem service bundles ⁎
T
Cedric Marsboom , Dirk Vrebos, Jan Staes, Patrick Meire University of Antwerp, Department of Biology, Ecosystem Management Research Group, Universiteitsplein 1C, B-2610 Wilrijk, Belgium
A R T I C L E I N F O
A B S T R A C T
Keywords: Ecosystem services Dimension reduction Principal component analysis Ecosystem service bundling
The concept of ecosystem services (ES) has facilitated the identification, mapping and communication about the many non-marketable benefits of green infrastructure. These benefits are important to consider during a spatial planning process. For spatial prioritisation of sites with a high societal importance, there is need to filter this information to insightful spatial indicators. The mapping of ES-hotspots and identification of ES-bundles have been put forward as promising methods for spatial prioritisation and the assessment of multifunctionality. While “hotspot mapping” and “ES-bundles” speak to the imagination of many, it is open to many different interpretations. In addition, there is a risk that the commonly applied hotspot mapping of single services and subsequent overlay analysis does not capture true hotspots of multifunctionality, where we expect multiple services to co-occur, but at lower intensities. Therefore, hotspot mapping should be applied on ES-bundles, rather than single ES. Yet, there are few methods to objectively identify and map such bundles of co-occurring services. In this research we propose dimension reduction principal component analysis (PCA), as a solution to identify and map bundles of ES. This technique is an established technique in remote sensing, where it is used to reduce unnecessary clutter in a data set. This research shows that if the methods for quantification and mapping of ES are sufficiently independent and biophysically sound, the PCA method can reveal multifunctionality between services and lead to (new) insights that can be used for better informed decisions on management and planning. The PCA graphs, ES-bundle maps and the integrated RGB-visualisation are objective and factual outputs of a statistical analysis that can be used for communication and discussion with stakeholders. It gives insight in cooccurrence of services and challenges to look for answers to why things are the way they are. Although scale effects did not play an important role in the results of this study, we advise to use this method on relatively small scales and repeat analysis rather than generalizing large scale results to the local scale or transfer findings between study sites as land-use patterns (and its interplay with abiotic conditions) are the result of many different socio-ecological developments throughout history, which can obviously differ from region to region.
1. Introduction Landscape multifunctionality is an important objective in modern spatial planning. The concept of ecosystem services has increasingly been adopted as narrative to point-out that the societal relevance of many “clustered” services outweighs the market value of few, mostly provisioning services (Jones et al., 2013; Plieninger et al., 2014; Termorshuizen and Opdam, 2009; Vallés-Planells et al., 2014). But translating this narrative to spatial explicit assessments remains a challenge. So far, there have been only a few studies that encompass a broad range of services in a comprehensive, quantitative and spatially explicit manner (Boerema et al., 2016). But with an increasing availability of tools and methods, we can expect a trend towards integrated, high resolution assessments that address many ecosystem services. Such studies generate a vast amount of spatial explicit data and there is a need to filter this information to insightful spatial indicators. ⁎
The identification and mapping of ES-hotspots and ES-bundles have been put forward as promising methods forspatial prioritisation and the assessment of multifunctionality. While “hotspot mapping” and “ESbundles” speak to the imagination of many, it is open to many different interpretations. Since a multitude of definitions and interpretations exist, the concepts are prone to misuse (Schröter and Remme, 2015). The most common definition of single ecosystem services hotspots is the definition of Egoh et al. (2008), which identifies hotspots as “areas which provide large proportions of a particular service”. The existing techniques for mapping this type of hotspot are well established and relatively easy to implement and areas which signify high delivery (Bai et al., 2011; Beverly et al., 2008; Crossman and Bryan, 2009; Egoh et al., 2009, 2008; Forouzangohar et al., 2014; Gos and Lavorel, 2012; Locatelli et al., 2014; O'Farrell et al., 2010; Onaindia et al., 2013; Plieninger et al., 2013; Schulp et al., 2014; Timilsina et al., 2013; Willaarts et al., 2012; Willemen et al., 2010). These types of hotspots
Corresponding author at: UA – Campus Drie Eiken, Ecosystem Management research group, Universiteitsplein 1, Building C, C1.20 B-2610 Wilrijk, Belgium. E-mail address:
[email protected] (C. Marsboom).
https://doi.org/10.1016/j.ecolind.2017.10.049 Received 25 January 2017; Received in revised form 15 August 2017; Accepted 23 October 2017 1470-160X/ © 2017 Elsevier Ltd. All rights reserved.
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
extensive dataset of ecosystem service maps. To reduce the dataset this technique statistically groups highly correlated variables, in our case ecosystem services maps, on principal component (PC) axes. These PC axes provide two main results. First, each axis provides a statistical grouping of correlated ecosystem services. Secondly, these axes can be presented as maps which signify the multi-service hotspots for that bundle of ecosystem services. In this paper we apply this method on a study site and discus the applicability and differences with the existing method. Further, we hypothesise that composition and robustness of ES-bundles depends on the scale and context of the study site and illustrate the scale effects by applying the developed methodology on different spatial extends with the same resolution. Finally we suggest guidelines for application and interpretation.
mapping techniques are suitable when single or few ecosystem services are used for the identification and prioritisation of sites for conservation (Schröter and Remme, 2015). The methods to identify and map multiservice hotspots are much less established. There, the aim is to map zones, which are high in delivering a multitude of ecosystem services (Schröter and Remme, 2015). The methods to identify and map multi-service hotspots are much less established. There, the aim is to map zones, which are high in delivering a multitude of ecosystem services (Schröter and Remme, 2015). The concept of multi-service hotspots is closely related to the idea of ES-bundles. According to Raudsepp-Hearne et al. (2010), ESbundles are “sets of ecosystem services that repeatedly appear together across space or time for a given area”. Multi-service hotspots therefore should signify locations that are high in delivering a certain bundle of services. These multi service areas are not only of great importance for conservation purposes (Egoh et al., 2009), but also on a policy level they commend special attention in larger planning processes. Although the concept of ecosystem service-bundles (ES-bundles) is relatively well-established (Crouzat et al., 2015; Qiu and Turner, 2013; Raudsepp-Hearne et al., 2010; Van der Biest et al., 2014), its practical application is not straightforward. Identifying ES-bundles encounters several conceptual and technical problems. Currently, the top richest cells method is often applied on individual services and then used as a basis to identify the multiservice hotspots (Eigenbrod et al., 2010; Qiu and Turner, 2013; Rodriguez et al., 2015; Wu et al., 2013). The spatial overlap of single service hotspot maps is used as a criterion to identify multi-service hotspots. But simply adding up single service hotspots ignores the notion that ecosystem services often occur in bundles due to physical (e.g. moisture gradients, slope etc.), anthropogenic (accessibility, population density) and ecological factors and constraints. Ecosystem services supply and demand interactions rely on many interacting biotic and abiotic drivers. Depending on spatial context and configuration, identical land use may deliver different services. A patch of forest in an urban environment will provide many other services (e.g. air quality regulation, recreation, noise attenuation, health effects) then an identical patch of forest in a rural environment (e.g. carbon sequestration, timber production, pollination). This makes that some ecosystem services will unlikely occur at the same locations and they will cancel each other out when simply adding up maps (Assessment, 2005; Bennett et al., 2009; Daily et al., 2009; Raudsepp-Hearne et al., 2010; Setala et al., 2014; Turner et al., 2014; Zhang et al., 2007). Therefore, the convergence of e.g. 3 ES in a specific type of ES-bundle can be as important as the convergence of 6 other ES in another type of bundle. When used for multiservice hotspots, this technique encounters these problems with positive and negative correlations making it less reliable and more difficult to interpret the results. More advanced methods, such as Self-Organizing Maps (SOM) (Crouzat et al., 2015; Kohonen, 2001; Mouchet et al., 2017; van der Zanden et al., 2016) give promising results, but its application in the domain of ES-research remains limited. Its practical application may be hampered by the complexity of the method and high sensitivity to data quality and issues with the occurrence of no-data zones in input maps (Rustum and Adebayo, 2007). Some types of ecosystem services typically have large areas of no-data or zero values when they are mapped. Other techniques like Bayesian Belief Networks (BBN) require a weighing of the different ecosystem services (Van der Biest et al., 2014), which is not only very difficult to do but it relies heavily on expert judgment which introduces extra subjectivity in the analysis (Gos and Lavorel, 2012). The correlations between ecosystem services can be difficult to interpret with an increasing number of ecosystem services, especially for large study sites (Carpenter et al., 2009; Pataki et al., 2011; Raudsepp-Hearne et al., 2010; Setala et al., 2014). Therefor there is need for an objective procedure to identify and map ecosystem service bundles, especially for studies that encompass many services. This paper applies the principle of dimension reduction on an
2. Materials and methods 2.1. Study area 2.1.1. Small scale The small-scale study area is a 12 km by 12 km square near the city of Turnhout, Belgium (Fig. 1). The area is characterised by two main focal points: a mid-sized city (Turnhout, 43.460 inhabitants) and the E34 highway which crosses trough the study area. Nature reserves are located north, south and easth of the city. Woodland and agriculture are also common in the area. A full description of the land use in the study area can be found in Table 1. 2.1.2. Larger scale As the larger scale study area we opted for the province of Antwerp in which the small scale study area is situated. The province has 1.8 million inhabitants on an area of 2867 km2. It includes both rural areas as well as one of the most urbanized areas of Belgium. A description of the land use in the area can be found in Table 1. 2.2. Input data All 15 maps used, were developed within the ECOPLAN project, which mapped and modelled ecosystem services for the Flemish Region (part of Belgium), using input data of high thematic and spatial resolution (5m) (Ecoplan, 2016). The ecosystem service maps used in the analysis are presented in Table 2. A full description of the ECOPLAN project and the used ecosystem services can be found in the supplementary materials. These quantitative maps result from biophysical and statistical models, using a large set of biophysical variables. The use of quantitative maps has the advantage to be less subjective than maps based on qualitative indicators. 2.3. Dimension reduction PCA We applied techniques for dimension reduction on a set of ecosystem services maps. Due to the spatially explicit nature of remote sensing techniques it can provide solutions for the spatial related issues involved in ecosystem services mapping such as extent and specific locations (Feng et al., 2010). Because remote sensing imagery often results in large datasets consisting of many different bands, dimension reduction is often used in remote sensing as a pre-processing step before classification (Li et al., 2012). A dimension reduction technique reduces these large datasets into more manageable datasets by removing redundant information and reducing the variability between the bands to a limited number of components (Plaza et al., 2005). A PCA was used for dimension reduction. PCA was developed by Pearson in 1901 and developed independently by Hotelling, 1933; Jolliffe, 1986. Abson et al. (2012) already stated the potential of PCA to aggregate spatially explicit variables. PCA is a well-known technique and, in remote sensing, one of the most commonly used dimension 210
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Fig. 1. An overview of the study area: the smaller study area is marked by the black square, the larger study area is marked as the filled in area on the map of Belgium.
reduction techniques (Fodor, 2002; Plaza et al., 2005). In addition, it has the best mean-square error (Fodor, 2002) of all linear dimension reduction techniques available. PCA is perfect for datasets in which there is a (approximate) linear relationship between the different services and has a proven track record to determine ecosystem service relationships (Birkhofer et al., 2015). The aim of the PCA is to generate a reduced number of bands with concentrated information. This ordination based technique captures the variability in our data by transforming spatial correlated ecosystem services into uncorrelated bundles of ecosystem services. This technique captures the underlying patterns within a multi variable dataset (Abson et al., 2012). These bundles are then presented as PC axes. These axes have as little correlation between them as possible. An overview of the workflow is given in Fig. 2. The various ecosystem services maps are first combined into a raster stack as different bands, to facilitate calculations, although this is not a necessity (not shown in Fig. 2). The PCA technique itself follows the five classic steps of a PCA (green squares). In the first step a mean is calculated over each layer and then subtracted from that layer. In step two, a correlation matrix is calculated, which is then used in step three to calculate the eigenvector and eigenvalues. In step four the eigenvalues are ranked in order of significance (highest to lowest value). Step three and four are combined into a single step in Fig. 2. The output tables of step two and three are merged and exported as a table (yellow circle). In the last step the new reduced data set is calculated by transposing the original ecosystem service maps. To limit processing time for the large study site, a random sampling was performed to end up with 25% of the total amount of cells. This sample is then analysed by the dimension reduction algorithm. The PCA script is written in R (version 3.1.3, (R Development Core Team, 2016)), the raster, maptools, rgdal and Rgl packages were used. A major concern for any statistical analysis is how zero values and no data pixels may affect the results of the analysis. No data signifies in our analysis
Table 1 An overview of the land use in the study area. Land use
Area Small study area
Nature Agriculture Grassland Orchard Floriculture Recreation Industrial Built-up area others Military domain
Large study area
[km²]
[%]
[km²]
[%]
12.65 7.93 52.22 1.03 0.3 4.67 5.12 30.62 29.45 0.01
8.78 5.51 36.26 0.72 0.21 3.24 3.56 21.26 20.45 0.01
118.52 130.92 902.87 5.16 13.02 83.3 157.51 742.21 711.06 2.43
4.13 4.57 31.49 0.18 0.45 2.91 5.49 25.89 24.8 0.08
Table 2 An overview of the used ecosystem services in this research. A full description of the used ecosystem services can be found on http://ecosysteemdiensten.be/cms/en/. Regulating services
Producing services
Cultural services
Noise attenuation Denitrification
Wood production Food production
Infiltration
Water provisioning
Health effects of nature Added value to houses due to a green environment Cooling effects from green infrastructure
Carbon in soil Carbon in biomass Air quality regulation Nitrogen in the soil Phosphorous in the soil Water retention
211
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
ecological research (Schröter and Remme, 2015). The first step in the top richest cells algorithm is the ranking of all cells with descending values. A top class is derived from this ranking comprising of the top 10% of cells. This percent is a predefined choice and can be adapted if desired. We opted for the (commonly used) 10% top class and therefore refer to this method as the top 10% method in de rest of this paper.
2.5. Overlay analysis An overlay analysis was used to compare the results of a top 10% approach on individual ES maps to the results of a top 10 richest cells approach on the PC-maps. The top 10% richest cells selects the top 10% area (pixels) with the highest values for each map (n = 15). These top 10% maps are then overlaid on top of each other and the overlap is calculated. This overlap is then compared to the PCA maps to provide indications on the (dis-) similarities in the spatial patterns generated by both approaches.
3. Result 3.1. PCA results 3.1.1. Small scale The PCA on the small study area resulted in 12 principal components (PC) (PC 12 explained 100% of the variation) of which the first 6 PC’s explained 87.45% of the variation. Consequently PC 7–12 only explained 12.55% of the variation. These PC’s seemed erratic and didn’t contribute to any logical interpretation of spatial patterns (based on expert judgment). The PC graphs were built for the first six PC’s to identify the bundles and defining PC axes, the first 3 are shown in Fig. 3. All the principal components graphs can be found in Appendix A. In the left side of Table 3 we list the bundles that could be distinguished from the graphs. The first 3 PC explain the most variation, nevertheless, higher order PC’s can potentially be relevant, depending on the application (Townshend, 1984). For example both PC 4 and PC5 have the cultural ecosystem services grouped on its axis. Where PC 5 shows a positive correlation, the opposite can be observed for PC 4. PC 5 thus clearly shows a clear negative correlation between “health effects of nature” and “ Added value to houses due to a green environment” on the one hand and “noise attenuation” and “Cooling effects from green infrastructure” on the other hand. On PC 4these are distinctly grouped together. Although the modelling concept and variables to quantify these four ecosystem services is largely similar, the input variables are treated differently in the respective models. All four ES are strongly driven by a moderately high population density in combination with local green infrastructure. But noise attenuation and urban heat island cooling effects usually occur in extremely high population density areas (cities), while health effects and added value to real estate relate to population density in general.
Fig. 2. An overview of the workflow followed in this research. Blue prisms represent input parameters and output results. The green squares are calculating steps, while orange circles represent relevant intermediate results. The first three steps represent the PCA and the last step transposes the ecosystem services (EES) maps to PC maps. The output generates an area covering map with the multi service ESS area’s and an output table displaying the relevant statistics. This output table is then used to build the PC graphs and the map legend.
missing data for the ES models due to missing information in the input variables, which is different from a modelled/predicted absence of the ES (zero values). The problem of zero-inflated datasets (true zeros and no data) is a well-studied, but complex problem in ecology (Zuur et al., 2010). However the different R-packages used in the presented method can process no data zones. Nevertheless, the presence of large amounts of no data or zero value pixels within a specific study area can still affect the results. A large proportion of zeros will shift the centering of the data in the PCA method. While large no data areas in relation to small ‘actual data’ area can lead to no data values in the correlation matrix. Therefore, one should carefully consider the impact of no data areas and zero values on the analysis. No data pixels also influenced the number of cells in the top richest cell approach, as fewer value pixels in total also resulted in fewer pixels of the top 10% class. Besides the different PC maps, both the eigenvalues and eigenvectors of step three can be drawn from the analysis. The latter are used to create PCA graphs. The PC maps can be further processed and visualised depending on the aim of the analysis. Three PC axes were combined into a RGB image by assigning either a red, green or blue colour to each axis (Chitwong et al., 2002; Pandey et al., 2014) as an example of how to visualise the results.
3.1.2. Large scale In order to indicate scale effects we ran the analysis with the same set of ecosystem services on a larger scale (province of Antwerp). The application of the dimension reduction technique on the larger study area and resulted in 13 principal components (PC) of which the first 7 PC’s, which explained 87.72% of the variation, were most useful. PC 8–13 seemed erratic and didn’t contribute to any logical interpretation of spatial patterns (based on expert judgment). The PC graphs were built for the first 7 PC’s to identify the bundles and defining PC axes, the first 3 are shown in Fig. 4. All the principal components graphs can be found in Appendix B. In the right side of Table 3 we list the bundles that could be distinguished from the graphs.
2.4. Top richest cells technique The top 10% richest cell method was chosen as it is the most commonly used method for creating single service hotspots in 212
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Fig. 3. The principal component (PC) graphs of the first 3 principal components of the small study area. All the principal components graphs can be found in Appendix A. With ecosystem service (ES) 1: Noise attenuation; ES 2: Denitrification; ES 3: Infiltration; ES 4: Carbon in soil; ES 5: Carbon in biomass; ES 6: Air quality regulation; ES 7: Nitrogen in the soil; ES 8: Phosphorous in the soil; ES 9: Water retention; ES 10: Wood production; ES 11: Food production; ES 12: Water provisioning; ES 13: Health effects of nature; ES 14: Added value to houses due to a green environment; ES 15: Cooling effects from green infrastructure. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
do largely match up. Two axes form exactly the same bundles for both the large and small scale, Axis 3 bundles retention and denitrification on the negative side of the axis and infiltration and extracted infiltration on the positive side. Also axis 4 matches up for all the cultural services between both study areas. Other axes match closely but not completely, axis six for example has food production on the positive side of the axis in the smaller study area which matches up in the larger study area but there is also extracted infiltration explained on the
3.2. Difference in scale If we compare the analysis of the small scale and large scale analysis we can see important similarities, but also some differences. The scale effects are not profound in this case. The first three PC of the larger study area explain a bit less of the total variation (respectively 24.8, 18.0 and 14.5%) than those of the smaller study area (respectively 24.3, 20.0 and 18.5%). The bundles which are formed on each PC (Table 3) 213
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Table 3 The bundling of ecosystem services on the principal component in both the small and the larger study area. Ecosystem services marked in bold are bundled on the positive side of the principal component axis, the other ones are bundled on the negative side of the axis. The cumulative percentage of variation explained is indicated between brackets. PC axis
Small study area
Large study area
1
Nitrogen in the soil, Phosphorous in the soil, carbon in the soil, infiltration, Water production and food production (24%) Wood production, Air quality regulation, carbon in wood and carbon in soil (44%) Denitrification, retention, infiltration and Water production (62%) Cooling effects from green infrastructure, noise attenuation, health effects of nature and added value to houses due to a green environment (74%) Cooling effects from green infrastructure, noise attenuation, health effects of nature and added value to houses due to a green environment (83%) Food production (88%) No longer relevant
Nitrogen in the soil, phosphorous in the soil, carbon in the soil & air quality regulation (25%) Carbon in wood, air quality regulation & wood production (43%) Infiltration, water production, water retention & denitrification (57%) Cooling effects from green infrastructure, Noise attenuation, health effects of nature & added value to houses due to a green environment (68%) Cooling effects from green infrastructure & added value to houses due to a green environment (75%) Food production & water production (82%) Noise attenuation (88%)
2 3 4 5 6 7
services. Denitrification doesn’t have a large overlap with any of the other services. This might be due to the heavy spatial fragmentation of this particular ES. In contrast, added value to houses due to a green environment, health effects of nature and cooling effects from green infrastructure on the other hand have a strong overlap. This may point to a common input parameter for their models (in this case population density). A similar reasoning can be made for wood production, carbon in biomass and air quality regulation. These ES are strongly related to a woodland land cover. If all the top 10% maps are overlaid with each other to look for bundles with more than 2 ES, they cover a total selection of 55.8% of the study area. This accounts for such a large proportion that we can hardly speak of hotspots anymore. A possible answer for large selection in area can be found in Fig. 6, which represents the number of ES in the selected areas of this overlay map. The actual map can be found in Appendix D.
negative side. 3.3. Visualisation and interpretation: PCA mapping PC axes can be converted back into maps. This allows to study the spatial patterns of the bundles. As is clear from Table 3, some ecosystem services can be explained by multiple PC axes. For some very specific ecosystem services (e.g. noise attenuation), the explaining axes can be found in the higher order principal components. However in general, the higher order principal components can be ignored, since they explain only a very small part of the variance and encompass mostly random variation. We can also observe that for our results, some ecosystem services (such as carbon in the soil), are explained by multiple lower axes. To visualise this higher level of multifunctionality, the maps of the first three principal component axes can be combined in a RGB image (Fig. 5). The three individual PC-maps, respectively in red, green and blue, can be found in Appendix C. This composite image visualises the overlapping principal component maps and assigns it a unique colour. The legend of the map is presented as a 3D principal component plot of the first 3 axes. The colour of the squares represents the colour of that ecosystem service on the map. The other colours on the map are formed by overlapping colours. These overlapping colours represent hotspots of ecosystem services. The overlapping colour signifies the ecosystem services that are present in that hotspot.
4. Discussion We start this discussion section by providing arguments why the presented approach to mapping ES-bundles and ES-hotspots could be preferred over other methods. We deepen the reflection on the comparison between the popular, but rather simple top 10% richest cell method and the PCA approach. We use this analysis to point out that the top 10% especially captures monofunctionality rather than multifunctionality. We also discuss the visualisation challenge for ES bundles and the interpretation issues for the integrated RGB-visualisation. Finally, we raise a few points to consider before applying the approach to other study sites. We first point out that the co-occurrence of particular ES indeed depends much on model architecture and its key variables. Preferably, the approach is applied on the results of complex multiple services modelling approaches. Finally, we elaborate briefly on the impact of scale effects on the PCA results. Secondly, Our aim was to demonstrate that the PCA method performs well in capturing multifunctionality of ES. Yet, there are few existing procedures to compare to. Usually, studies simply overlay services and count co-occurrence. To apply such a method, single services need to be reclassified to comparable units (scores). Such a classification is a very subjective step in any analysis (Greenland and O'rourke, 2001). Any choice for a particular classification method (e.g. equal area vs. equal interval) will affect the analysis. Since, the top 10% method is traditionally the most used method to identify single service hotspots in ecological research (Schröter and Remme, 2015). We expanded this to an overlay analysis of the top 10% maps, being a logical and rather objective step to analyse multifunctionality. In some cases, the top 10% overlap analysis (Table 5) confirms the bundling of particular ecosystem services (e.g. wood production and air quality regulation). It matches the correlation that can be deducted from the PCA analysis results (Fig. 3). It also illustrates the negative correlation between particular ecosystem services (e.g. water retention
3.4. Comparing the PCA method to the top 10 method In order to evaluate the PCA method, we applied the top 10% richest cells also to the relevant PC-maps (PC’s 1–6). We then analysed spatial overlap between the top 10 PC-maps and the top 10 single service maps. This then resulted in two cross tables. The first cross table (Table 4) shows the overlap between the principal components and the individual ecosystem services. This table shows that large portions of the top 10% percent of each ecosystem service falls inside the top 10% of the principal component map on which they are bundled. In Table 4 the values of the first two principal components were inverted as all the ecosystem services were bundled on the negative side of these axes. This table confirms the bundles that are derived from the PCA analysis but also clearly shows that the PCA hotspots don’t include all the top 10% percent pixels. For some services, a significant part of the top 10% is captured in the top 10% of the bundle (e.g. air quality regulation in PC2). Other services have their top 10% captured by many bundles (e.g. infiltration). We can also see examples where the top 10% is not captured by any of the bundles (e.g. water retention). This is due to the fact that the top 10% percent method includes all the extreme values. These high values are often not very representative, having exceptional conditions that have very little correlation with other ecosystem services. The second cross-table (Table 5) shows the overlap between the top 10% maps of the single ecosystem services. This table is used to determine the bundling of ecosystem services based on the top richest cells method. Overlap signifies high occurrence of both ecosystem 214
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Fig. 4. The principal component (PC) graphs of the first 3 principal components of the larger study area. All the principal components graphs can be found in Appendix B. With ecosystem service (ES) 1: Noise attenuation; ES 2: Denitrification; ES 3: Infiltration; ES 4: Carbon in soil; ES 5: Carbon in biomass; ES 6: Air quality regulation; ES 7: Nitrogen in the soil; ES 8: Phosphorous in the soil; ES 9: Water retention; ES 10: Wood production; ES 11: Food production; ES 12: Water provisioning; ES 13: Health effects of nature; ES 14: Added value to houses due to a green environment; ES 15: Cooling effects from green infrastructure. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
overlaying all the top 10% maps. Fig. 6 clearly shows that when looking for multi-service hotspots, this technique still maps mostly single service hotspots. If we overlay all top 10% single ES-maps, 51% of the ‘hotspot area’ was selected on the basis of one single top 10% map. However, this does not mean that other services cannot co-occur in lower supply ranges than their top 10%. This means that the identification of the multifunctional hotspots (with more than 2 ecosystem services) requires some puzzling, which again introduces subjective
and infiltration). This is important, as these services will cancel each other out if ecosystem services are simply overlaid and summed. Yet, because the top 10% method uses (only) the extreme values, it can give false negatives as well. It is not because there is little or no overlap in the top 10% range, that they cancel each other out completely (e.g. carbon in the soil and nitrogen in the soil). They can be correlated, but maybe just not in their extreme values, which makes the interpretation of the top 10% less straightforward. This is further elaborated by 215
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Fig. 5. Up to 3 principal components can be combined into a RGB image. The legend of the map is presented as a 3D plot of those 3 principal components. The colour of the squares is the same colours of the map. The remaining colours on the map are formed due to overlap of the different axes.
the output of the PCA approach.
choices. All of these drawbacks are avoided by the PCA method. Recently, a number of studies make use of Self-Organising Maps (SOM) to identify ES-bundles (Crouzat et al., 2015; van der Zanden et al., 2016). This method is also a dimension reduction method and can be seen as a non-linear generalisation of our PCA approach (Gorban et al., 2008). The output of the SOM is a classified map consisting of spatial clusters where SOM class depicts spatial clusters with similar co-occurrence of ES. This kind of output can be very relevant, but difficult to compare to
4.1. RGB visualisation: seeing the wood for the trees? With the RGB visualisation we compact the information from15 ESmaps into a single map. In essence this is an integrated indicator of multifunctionality. Developing indicators for complex systems is challenging (Hekstra, 1983) and can be misleading or biased. In most cases, 216
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
4.2. Understanding how models and data direct multifunctionality: ESbundles by model design or finding new insights?
Table 4 The percent of overlap between the top 10% of each individual ecosystem service and the top 10% percent of each principal component (PC). PC 1 and 2 were inverted because their top values are situated on the negative side of the principal component axis. Ecosystem service
PC 1 inverted
PC 2 inverted
PC 3
PC 4
PC 5
PC 6
Noise attenuation Wood production Food production Water provisioning Carbon in biomass Carbon in soil Air quality regulation Nitrogen in soil Phosphorus in soil Infiltration Water retention Denitrification Added value to houses due to a green environment Health effects of nature Cooling effects from green infrastructure
19.08 3.01 54.42 16.48 2.74 10.38 3.33 30.78 30.78 27.13 5.5 7.8 82.55
13.62 78.96 1.54 0.39 79.65 25.49 84.61 3.44 3.44 1.31 3.07 3.5 31.28
0.28 0.08 0 19.39 0.08 2.36 0.32 0.11 0.11 32.51 0.01 1.96 1.57
28.86 11.46 17.51 6.9 11.04 14.13 9.85 17.09 17.09 14.38 11.4 16.61 100
27.25 10.35 17.03 7.03 10.02 13.97 9.06 17.14 17.14 14.71 10.94 16.33 100
0.57 14.4 80.38 15.36 14.39 0.55 25.13 22.42 22.42 22.36 2.74 2.17 0.67
42.02 7.71
18.44 7.71
2.57 0
52.03 7.71
52.28 7.71
0.82 0
In the highly cited theoretical paper by Bennett et al. (2009), they state that ecosystem services can interact directly or appear to interact through the impact of a shared driver, but we have limited theory or general rules about these relationships and their implications for management of ecosystem services. This has been confirmed by many studies. The most evident approach is to link ES-supply to particular land-uses (Raudsepp-Hearne et al., 2010). But also soil and water conditions are key drivers. Exemplary for this is the study by Pan et al. (2013), which concludes that for their large scale study site, the total ecosystem services supply is mainly driven by precipitation. The interplay of abiotic conditions, societal factors and ecosystem functioning that determines such bundles of ES is difficult to untangle (Duncan et al., 2015). In our case, most of the ecosystem services are also bundled around their defining input variables (e.g. population density, soil properties, particular land use etc.). As a result ecosystem services which rely (partly) on the same input variables, are likely to correlate which each other, as is the case in our set of ecosystem services with nitrogen and Phosphorus in the soil (cf. Table 5 and Figs. 3 and 4). The PCA methodology allows for the quantification and visualisation of how ecosystem services correlate with each other (Jolliffe, 2002). Positive correlated ecosystem services tend to co-occur and negatively correlated services tend to exclude each other. But the co-occurrence of ES does not necessarily imply that management measures are strengthening such a co-occurrence. The ES-bundles are in any case a point of discussion on the how and why these particular services cooccur. Further analysis can reveal whether this is due to manageable (e.g. production forests tend to be closed for recreation) or non-manageable variables (e.g. infiltration occurs on dry sandy soils with low vegetation, conditions that lead to low carbon stocks).
integrated indicators are the result of combining, weighing and scoring lower level indicators (e.g. (Alam et al., 2016; Koschke et al., 2012; Mononen et al., 2016)). A big advantage of our approach is that there is no need for weight matrices or other subjective steps (classifications, scorings) in the analysis. The methodology does not require any assumptions of the importance of each individual ecosystem service. With this method subjectivity can be eliminated as much of as possible and the method can be used without the need for an expert panel. The RGB visualisation has the advantage that it gives unique colour combinations to the location on the different axes and that clear visualisation helps interpretation (Fisher et al., 1993). This allows highlighting of map ecosystem services that are explained by more than one axis as up to three axes can be combined in such an image. The fact that the variance is explained by multiple principal components, may indicate that these services are generated by complex interactions (Limburg et al., 2002; Fisher et al., 2009), which makes them correlate with multiple other services. In this case, the ECOPLAN results for soil organic carbon indeed depends on input variables such as land cover, land use, soil profile, texture and groundwater levels (Ottoy et al., 2015, 2016). A downside is that correct interpretation of the map is highly dependent on an elaborated colour legend. This makes the initial interpretation of the RGB image a bit incomprehensive. Because ecosystem services can also be negatively correlated, the colour scheme is not always as intuitive as expected, which stresses the need for an advanced colour legend. Maps can also be inverted before adding it to the RGB to make it more intuitive. This is especially recommended if the ecosystem services are all grouped on the negative side of principal component axis. The areas with high delivery are indicated in the map but it is not possible to deduce the actual delivery. This requires an extra step outside of the analysis and could be facilitated in an interactive GISenvironment. But even with this disadvantage it can still be used as a powerful communication tool. By first bundling the ecosystem services and then combining the bundles of interest into a RGB the maximum of information is combined in a single map. Yet it is clear that the original ecosystem service maps need to be consulted if one needs to know quantitative values of an ES. The RGB visualisation is constrained to only a limited number of PC axes. But the first three PC axes already explain most of the variation (62.8%). If needed the RGB visualisation can be applied adjustable, replacing one of PC’s in the RGB with other, lower PCs when relevant.
4.3. Scale effects Hotspots are always relative to their surroundings. This makes hotspot mapping prone to scale dependency. It has been previously stated in literature that scale needs to be considered with a PCA technique (Abson et al., 2012). To have an indication of how profound these scale effects are in this type of analysis, we ran the analysis for the same set of ecosystem services on a small and a larger study area. The scale effects are not profound in this case. This could be due to the fact that both study areas are part of the same ecoregion and has similar landscape and abiotic parameters. In cases where the larger study area has is mostly comprised of different parameters a more profound difference in bundling can be expected. 5. Conclusion We have used a well-known remote sensing technique dimension reduction PCA as a simple solution to identify ES-bundles and map hotspots of multifunctionality. The output of this method can be used for a better informed, objective, decision making on land use planning. Such methods are much needed since consumption of open space and seemingly less profitable land use is problematic in many regions (Lambin and Meyfroidt, 2011). How to deal with land use competition is an increasing problem in this day and age (Raudsepp-Hearne et al., 2010; Setala et al., 2014). Land use policies often contradict each other and strive for conflicting objectives (Bennett et al., 2009; Daily et al., 2009; Setala et al., 2014). With this method it requires little effort to identify ES-bundles and map multifunctionality hotspots. If the methods for quantitifcation and mapping are sufficiently independent and biophysically sound, the PCA methods can reveal multifunctionality between services and lead to (new) insights that can be used for better management and planning. When taking ES into account for conservation planning (in terms of 217
Noise attenuation Wood production Food production Water provisioning Carbon in biomass Carbon in soil Air quality regulation Nitrogen in the soil Phosphorus in the soil Infiltration Water retention Denitrification Added value to houses due to a green environment Health effects of nature Cooling effects from green infrastructure
3.63 100 0 0 98.72 13.9 82.55
0.03 0.03
0 1.07 0.31 5.5
3.27
0
5.6 5.6
5.8 4.84 4.65 18.87
40.58
3.15
Wood production
100 3.63 4.69 8.69 3.65 27.26 5.66
Noise attenuation
3.19
10.22
26.95 4.17 3.04 13.56
35.83 35.83
4.69 0 100 19.86 0 0 0
Food production
0
2.31
63.66 0 0 5.74
22.6 22.6
8.69 0 19.86 100 0 1.46 0
Water provisioning
0
2.83
0 0.75 0.25 5.26
0.01 0.01
3.65 98.72 0 0 100 13.37 83.74
Carbon in biomass
37.48
19.73
0 12.97 10.27 12.56
26.18 26.18
27.26 13.9 0 1.46 13.37 100 18.44
Carbon in soil
0
3.3
0 3.97 1.97 6.63
1.71 1.71
5.66 82.55 0 0 83.74 18.44 100
Air quality regulation
4.21
13.37
29.77 11.19 8.61 14.43
100 100
5.6 0.03 35.83 22.6 0.01 26.18 1.71
Nitrogen in the soil
Table 5 A crosstable of the overlap (in%) of the top 10% between each of the individual ecosystem services in the small study area.
4.21
13.37
29.77 11.19 8.61 14.43
100 100
5.6 0.03 35.83 22.6 0.01 26.18 1.71
Phosphorus in the soil
4.21
8.15
100 0 0 13.63
29.77 29.77
5.8 0 26.95 63.66 0 0 0
Infiltration
0
9.01
0 100 34.78 8.36
11.19 11.19
4.84 1.07 4.17 0 0.75 12.97 3.97
Water retention
0
11.64
0 34.78 100 12.14
8.61 8.61
4.65 0.31 3.04 0 0.25 10.27 1.97
Denitrification
7.71
56.88
13.63 8.36 12.14 100
14.43 14.43
18.87 5.5 13.56 5.74 5.26 12.56 6.63
Added value to houses due to a green environment
80.57
100
8.15 9.01 11.64 56.88
13.37 13.37
40.58 3.27 10.22 2.31 2.83 19.73 3.3
Health effects of nature
100
80.57
4.21 0 0 7.71
4.21 4.21
3.15 0 3.19 0 0 37.48 0
Cooling effects from green infrastructure
C. Marsboom et al.
Ecological Indicators 87 (2018) 209–260
218
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
scale or transfer findings between study sites as local variations in landscapes and land use (history) can have an impact on ES delivery (Verhagen et al., 2016). Visualisation and communication remains a challenge, but by use of the more complex, but intergrated RGB-map and/or the separate bundle maps and the PCA graphs, this method does provide objective output that can be used for communication and discussion with stakeholders. It gives insight in co-occurence of services and challenges to look for answers to the why things are the way they are. Finally we advocate ES-research to look beyond the obvious ES-research and literature. Although remote sensing is frequently used in ecosystem services research, there still are a lot of unexplored potentials (Feng et al., 2010).
Fig. 6. Histogram of overlapping top 10% ecosystem services maps.
ES) we advise against the top 10% method since it rather captures mono-functionality than multifunctionality. By applying the top 10% on the PCA maps, we were able to detect hotspots of multifunctionality that potentially have a high societal value. Although scale effects did not play an important role in the results of this study, we advise to use this method on relatively small scales and repeat analysis rather than generalizing large scale results to the local
Acknowledgements This research was conducted under the ECOPLAN project (Planning for Ecosystem Services; IWT-SBO 100420; https://www.uantwerpen. be/en/rg/ecoplan/). We would like to thank three anonymous reviewers for their useful feedback on the manuscript.
Appendix A The PCA graphs of the small study area, the interpretation of these graphs can be found in Table 3.
Figure
Page number
PC1-PC2 PC1-PC3 PC1-PC4 PC1-PC5 PC1-PC6 PC2-PC3 PC2-PC4 PC2-PC5 PC2-PC6 PC3-PC4 PC3-PC5 PC3-PC6 PC4-PC5 PC4-PC6 PC5-PC6
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55
219
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
220
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
221
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
222
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
223
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
224
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
225
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
226
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
227
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
228
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
229
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
230
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
231
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
232
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
233
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Appendix B The PCA graphs of the large study area, the interpretation of these graphs can be found in Table 3.
Figure
Page number
PC1-PC2 PC1-PC3 PC1-PC4 PC1-PC5 PC1-PC6 PC1-PC7 PC2-PC3 PC2-PC4 PC2-PC5 PC2-PC6 PC2-PC7 PC3-PC4 PC3-PC5 PC3-PC6 PC3-PC7 PC4-PC5 PC4-PC6
58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 234
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
PC4-PC7 PC5-PC6 PC5-PC7 PC6-PC7
75 76 77 78
235
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
236
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
237
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
238
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
239
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
240
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
241
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
242
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
243
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
244
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
245
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
246
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
247
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
248
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
249
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
250
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
251
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
252
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
253
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
254
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Appendix C The first 3 principal components presented respectively as red, green and blue.
255
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
256
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
257
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Appendix D A map representing the total number of ES in the overlay analysis of the top 10% maps.
258
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
Appendix E. Supplementary data Supplementary data associated with this article can be found, in the online version, at https://doi.org/10.1016/j.ecolind.2017.10.049.
multispectral bands with PAN band using HIS and wavelet transform. In: In: Ehlers, M. (Ed.), Remote Sensing for Environmental Monitoring, Gis Applications, and Geology 4545. Spie-Int Soc Optical Engineering, Bellingham, pp. 130–136. Crossman, N.D., Bryan, B.A., 2009. Identifying cost-effective hotspots for restoring natural capital and enhancing landscape multifunctionality. Ecol. Econ. 68 (3), 654–668. http://dx.doi.org/10.1016/j.ecolecon.2008.05.003. Crouzat, E., Mouchet, M., Turkelboom, F., Byczek, C., Meersmans, J., Berger, F., Lavorel, S., 2015. Assessing bundles of ecosystem services from regional to landscape scale: insights from the French Alps. J. Appl. Ecol. 52 (5), 1145–1155. http://dx.doi.org/ 10.1111/1365-2664.12502. Daily, G.C., Polasky, S., Goldstein, J., Kareiva, P.M., Mooney, H.A., Pejchar, L., Shallenberger, R., 2009. Ecosystem services in decision making: time to deliver. Front. Ecol. Environ. 7 (1), 21–28. http://dx.doi.org/10.1890/080025. Duncan, C., Thompson, J.R., Pettorelli, N., 2015. he quest for a mechanistic understanding of biodiversity–ecosystem services relationships. Proc. R. Soc. B (Vol. 282, No. 1817, p. 20151348). The Royal Society (2015, October). Ecoplan, 2016. Planning for Ecosystem Services. (UAntwerpen: IWT-SBO 100420). Egoh, B., Reyers, B., Rouget, M., Richardson, D.M., Le Maitre, D.C., van Jaarsveld, A.S., 2008. Mapping ecosystem services for planning and management. Agric. Ecosyst. Environ. 127 (1), 135–140. Egoh, B., Reyers, B., Rouget, M., Bode, M., Richardson, D.M., 2009. Spatial congruence between biodiversity and ecosystem services in South Africa. Biol. Conserv. 142 (3), 553–562. Eigenbrod, F., Armsworth, P.R., Anderson, B.J., Heinemeyer, A., Gillings, S., Roy, D.B., Gaston, K.J., 2010. The impact of proxy-based methods on mapping the distribution of ecosystem services. J. Appl. Ecol. 47 (2), 377–385. http://dx.doi.org/10.1111/j. 1365-2664.2010.01777.x. Feng, X.M., Fu, B.J., Yang, X.J., Lu, Y.H., 2010. Remote sensing of ecosystem services: an opportunity for spatially explicit assessment. Chinese Geogr. Sci. 20 (6), 522–535. http://dx.doi.org/10.1007/s11769-010-0428-y.
References Abson, D.J., Dougill, A.J., Stringer, L.C., 2012. Using principal component analysis for information-rich socio-ecological vulnerability mapping in southern africa. Appl. Geogr. 35 (1–2), 515–524. http://dx.doi.org/10.1016/j.apgeog.2012.08.004. Alam, M., Dupras, J., Messier, C., 2016. A framework towards a composite indicator for urban ecosystem services. Ecol. Indic. 60, 38–44. Assessment, M.E., 2005. Ecosystems and Human Well-Being 200 Island Press, Washington, DC. Bai, Y., Zhuang, C.W., Ouyang, Z.Y., Zheng, H., Jiang, B., 2011. Spatial characteristics between biodiversity and ecosystem services in a human-dominated watershed. Ecol. Complexity 8 (2), 177–183. http://dx.doi.org/10.1016/j.ecocom.2011.01.007. Bennett, E.M., Peterson, G.D., Gordon, L.J., 2009. Understanding relationships among multiple ecosystem services. Ecol. Lett. 12 (12), 1394–1404. Beverly, J.L., Uto, K., Wilkes, J., Bothwell, P., 2008. Assessing spatial attributes of forest landscape values: an internet-based participatory mapping approach. Can. J. For. Res.-Revue Canadienne De Recherche Forestiere 38 (2), 289–303. http://dx.doi.org/ 10.1139/x07-149. Birkhofer, K., Diehl, E., Andersson, J., Ekroos, J., Früh-Müller, A., Machnikowski, F., Mader, V.L., Nilsson, L., Sasaki, K., Rundlöf, M., Wolters, V., Smith, H.G., 2015. Ecosystem services – current challenges and opportunities for ecological research. Front. Ecol. Evol. 2, 1–12. Boerema, A., Rebelo, A.J., Bodi, M.B., Esler, K.J., Meire, P., 2016. Are ecosystem services adequately quantified? J. Appl. Ecol. Carpenter, S.R., Mooney, H.A., Agard, J., Capistrano, D., DeFries, R.S., Diaz, S., Whyte, A., 2009. Science for managing ecosystem services: beyond the millennium ecosystem assessment. Proc. Natl. Acad. Sci. U. S. A. 106 (5), 1305–1312. http://dx.doi.org/10. 1073/pnas.0808772106. Chitwong, S., Cheevasuvit, F., Homthong, J., 2002. Fusion of principal component of
259
Ecological Indicators 87 (2018) 209–260
C. Marsboom et al.
classification of hyperspectral image data using sequences of extended morphological transformations. IEEE Trans. Geosci. Remote Sens. 43 (3), 466–479. http://dx.doi. org/10.1109/tgrs.2004.841417. Plieninger, T., Dijks, S., Oteros-Rozas, E., Bieling, C., 2013. Assessing, mapping, and quantifying cultural ecosystem services at community level. Land Use Policy 33, 118–129. http://dx.doi.org/10.1016/j.landusepol.2012.12.013. Plieninger, T., van der Horst, D., Schleyer, C., Bieling, C., 2014. Sustaining ecosystem services in cultural landscapes. Ecol. Soc. 19 (2), 59. Qiu, J.X., Turner, M.G., 2013. Spatial interactions among ecosystem services in an urbanizing agricultural watershed. Proc. Natl. Acad. Sci. U. S. A. 110 (29), 12149–12154. http://dx.doi.org/10.1073/pnas.1310539110. R Development Core Team, 2016. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. Raudsepp-Hearne, C., Peterson, G.D., Bennett, E.M., 2010. Ecosystem service bundles for analyzing tradeoffs in diverse landscapes. Proc. Natl. Acad. Sci. U. S. A. 107 (11), 5242–5247. http://dx.doi.org/10.1073/pnas.0907284107. Rodriguez, N., Armenteras, D., Retana, J., 2015. National ecosystems services priorities for planning carbon and water resource management in Colombia. Land Use Policy 42, 609–618. http://dx.doi.org/10.1016/j.landusepol.2014.09.013. Rustum, Rabee, Adeloye, Adebayo J., 2007. Replacing outliers and missing values from activated sludge data using Kohonen self-organizing map. J. Environ. Eng. 133 (9), 909–916. Schröter, M., Remme, R.P., 2015. Spatial prioritisation for conserving ecosystem services: comparing hotspots with heuristic optimisation. Landscape Ecol. 31 (2), 431–450. Schulp, C.J.E., Lautenbach, S., Verburg, P.H., 2014. Quantifying and mapping ecosystem services: demand and supply of pollination in the European Union. Ecol. Indic. 36, 131–141. http://dx.doi.org/10.1016/j.ecolind.2013.07.014. Setala, H., Bardgett, R.D., Birkhofer, K., Brady, M., Byrne, L., de Ruiter, P.C., van der Putten, W.H., 2014. Urban and agricultural soils: conflicts and trade-offs in the optimization of ecosystem services. Urban Ecosyst. 17 (1), 239–253. http://dx.doi.org/ 10.1007/s11252-013-0311-6. Termorshuizen, J.W., Opdam, P., 2009. Landscape services as a bridge between landscape ecology and sustainable development. Landscape Ecol. 24 (8), 1037–1052. Timilsina, N., Escobedo, F.J., Cropper, W.P., Abd-Elrahman, A., Brandeis, T.J., Delphin, S., Lambert, S., 2013. A framework for identifying carbon hotspots and forest management drivers. J. Environ. Manage. 114, 293–302. http://dx.doi.org/10.1016/j. jenvman.2012.10.020. Townshend, J.R.G., 1984. Agricultural land-cover discrimination using thematic mapper spectral bands. Int. J. Remote Sens. 5 (4), 681–698. Turner, K.G., Odgaard, M.V., Bocher, P.K., Dalgaard, T., Svenning, J.C., 2014. Bundling ecosystem services in Denmark: trade-offs and synergies in a cultural landscape. Landscape Urban Plann. 125, 89–104. http://dx.doi.org/10.1016/j.landurbplan. 2014.02.007. Vallés-Planells, M., Galiana, F., Van Eetvelde, V., 2014. A classification of landscape services to support local landscape planning. Ecol. Soc. 19 (1). Van der Biest, K., D'Hondt, R., Jacobs, S., Landuyt, D., Staes, J., Goethals, P., Meire, P., 2014. EBI: An index for delivery of ecosystem service bundles. Ecol. Indic. 37, 252–265. http://dx.doi.org/10.1016/j.ecolind.2013.04.006. Verhagen, W., Van Teeffelen, A.J.A., Compagnucci, A.B., Poggio, L., Gimona, A., Verburg, P.H., 2016. Effects of landscape configuration on mapping ecosystem service capacity: a review of evidence and a case study in Scotland. Landscape Ecol. 31 (7), 1457–1479. http://dx.doi.org/10.1007/s10980-016-0345-2. Willaarts, B.A., Volk, M., Aguilera, P.A., 2012. Assessing the ecosystem services supplied by freshwater flows in Mediterranean agroecosystems. Agric. Water Manage. 105, 21–31. http://dx.doi.org/10.1016/j.agwat.2011.12.019. Willemen, L., Hein, L., van Mensvoort, M.E.F., Verburg, P.H., 2010. Space for people, plants, and livestock? Quantifying interactions among multiple landscape functions in a Dutch rural region. Ecol. Indic. 10 (1), 62–73. http://dx.doi.org/10.1016/j. ecolind.2009.02.015. Wu, J.S., Feng, Z., Gao, Y., Peng, J., 2013. Hotspot and relationship identification in multiple landscape services: a case study on an area with intensive human activities. Ecol. Indic. 29, 529–537. http://dx.doi.org/10.1016/j.ecolind.2013.01.037. Zhang, W., Ricketts, T.H., Kremen, C., Carney, K., Swinton, S.M., 2007. Ecosystem services and dis-services to agriculture. Ecol. Econ. 64 (2), 253–260. http://dx.doi.org/ 10.1016/j.ecolecon.2007.02.024. Zuur, A.F., Ieno, E.N., Elphick, C.S., 2010. A protocol for data exploration to avoid common statistical problems. Methods Ecol. Evol. 1 (1), 3–14. http://dx.doi.org/10. 1111/j.2041-210X.2009.00001.x. van der Zanden, E.H., Levers, C., et al., 2016. Representing composition, spatial structure and management intensity of European agricultural landscapes: a new typology. Landscape Urban Plann. 150, 36–49.
Fisher, P., Dykes, J., Wood, J., 1993. Map design and visualization. Cartographic J. 30 (2), 136–142. Fisher, B., Turner, R.K., Morling, P., 2009. Defining and classifying ecosystem services for decision making. Ecol. Econ. 68 (3), 643–653. Fodor, I.K., 2002. A Survey of Dimension Reduction Techniques: Technical Report UCRLID-148494. Lawrence Livermore National Laboratory. Forouzangohar, M., Crossman, N.D., MacEwan, R.J., Wallace, D.D., Bennett, L.T., 2014. Ecosystem services in agricultural landscapes: a spatially explicit approach to support sustainable soil management. Sci. World J. 13. http://dx.doi.org/10.1155/2014/ 483298. Gorban, A.N., Kégl, B., Wunsch, D.C., Zinovyev, A.Y. (Eds.), 2008. Principal Manifolds for Data Visualization and Dimension Reduction 58. Springer, Berlin, pp. 96–130. Gos, P., Lavorel, S., 2012. Stakeholders' expectations on ecosystem services affect the assessment of ecosystem services hotspots and their congruence with biodiversity. Int. J. Biodivers. Sci. Ecosyst. Serv. Manage. 8 (1–2), 93–106. Greenland, S., O'rourke, K., 2001. On the bias produced by quality scores in meta-analysis, and a hierarchical view of proposed solutions. Biostatistics 2 (4), 463–471. Hekstra, G.P., 1983. Indicators in complex systems. Ecological Indicators for the Assessment of the Quality of Air, Water, Soil, and Ecosystems. Springer, Netherlands, pp. 369–373. Hotelling, H., 1933. Analysis of a complex of statistical variables into principal components. J. Educ. Psychol. 24 (6), 417. Jolliffe, I., 1986. Principal Component Analysis. Wiley Online Library. Jolliffe, I., 2002. Principal Component Analysis. John Wiley & Sons, Ltd. Jones, K.B., Zurlini, G., Kienast, F., Petrosillo, I., Edwards, T., Wade, T.G., et al., 2013. Informing landscape planning and design for sustaining ecosystem services from existing spatial patterns and knowledge. Landscape Ecol. 28 (6), 1175–1192. Kohonen, T., 2001. Self-Organizing Maps. Springer-Verlag, Berlin Heidelberg. Koschke, L., Fuerst, C., Frank, S., Makeschin, F., 2012. A multi-criteria approach for an integrated land-cover-based assessment of ecosystem services provision to support landscape planning. Ecol. Indic. 21, 54–66. Lambin, E.F., Meyfroidt, P., 2011. Global land use change, economic globalization, and the looming land scarcity. Proc. Natl. Acad. Sci. 108 (9), 3465–3472. Li, W., Prasad, S., Fowler, J.E., Bruce, L.M., 2012. Locality-Preserving dimensionality reduction and classification for hyperspectral image analysis. IEEE Trans. Geosci. Remote Sens. 50 (4), 1185–1198. http://dx.doi.org/10.1109/tgrs.2011.2165957. Limburg, K.E., O'Neill, R.V., et al., 2002. Complex systems and valuation. Ecol. Econ. 41 (3), 409–420. Locatelli, B., Imbach, P., Wunder, S., 2014. Synergies and trade-offs between ecosystem services in Costa Rica. Environ. Conserv. 41 (01), 27–36. Mononen, L., Auvinen, A.P., Ahokumpu, A.L., Rönkä, M., Aarras, N., Tolvanen, H., et al., 2016. National ecosystem service indicators: measures of social?ecological sustainability. Ecol. Indic. 61, 27–37. Mouchet, M.A., Paracchini, M.L., et al., 2017. Bundles of ecosystem (dis)services and multifunctionality across European landscapes. Ecol. Indic. 73, 23–28. O'Farrell, P.J., Reyers, B., Le Maitre, D.C., Milton, S.J., Egoh, B., Maherry, A., et al., 2010. Multi-functional landscapes in semi arid environments: implications for biodiversity and ecosystem services. Landscape Ecol. 25 (8), 1231–1246. http://dx.doi.org/10. 1007/s10980-010-9495-9. Onaindia, M., de Manuel, B.F., Madariaga, I., Rodriguez-Loinaz, G., 2013. Co-benefits and trade-offs between biodiversity, carbon storage and water flow regulation. For. Ecol. Manage. 289, 1–9. http://dx.doi.org/10.1016/j.foreco.2012.10.010. Ottoy, S., Beckers, V., Jacxsens, P., Hermy, M., Van Orshoven, J., 2015. Multi-level statistical soil profiles for assessing regional soil organic carbon stocks. Geoderma 253, 12–20. http://dx.doi.org/10.1016/j.geoderma.2015.04.001. Ottoy, S., Elsen, A., Van De Vreken, P., Gobin, A., Merckx, R., Hermy, M., Van Orshoven, J., 2016. An exponential change decline function to estimate soil organic carbon stocks and their changes from topsoil measurements. Eur. J. Soil Sci. 67 (6), 816–826. http://dx.doi.org/10.1111/ejss.12394. Pan, Y., Xu, Z., Wu, J., 2013. Spatial differences of the supply of multiple ecosystem services and the environmental and land use factors affecting them. Ecosyst. Serv. 5, 4–10. Pandey, P.C., Tate, N.J., Balzter, H., 2014. Mapping tree species in coastal Portugal using statistically segmented principal component analysis and other methods. IEEE Sens. J. 14 (12), 4434–4441. http://dx.doi.org/10.1109/jsen.2014.2335612. Pataki, D.E., Carreiro, M.M., Cherrier, J., Grulke, N.E., Jennings, V., Pincetl, S., et al., 2011. Coupling biogeochemical cycles in urban environments: ecosystem services, green solutions, and misconceptions. Front. Ecol. Environ. 9 (1), 27–36. http://dx. doi.org/10.1890/090220. Pearson, K., 1901. LIII. On lines and planes of closest fit to systems of points in space. London Edinburgh Dublin Philos. Mag. J. Sci. 2 (11), 559–572. Plaza, A., Martinez, P., Plaza, J., Perez, R., 2005. Dimensionality reduction and
260
