SEARCH

SEARCH BY CITATION

Keywords:

  • Carnegie Airborne Observatory;
  • Heterogeneity;
  • LiDAR;
  • Object-based image classification;
  • Vegetation structure;
  • Woody vegetation

Abstract

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results and discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Question

The co-existence of woody plants and grasses characterize savannas, with the horizontal and vertical spatial arrangement of trees creating a heterogeneous biotic environment. To understand the influence of biogeophysical drivers on the spatial patterns of 3-D structure of woody vegetation, these patterns need to be explained over large areas to capture the context. Is there a spatially explicit, ecologically meaningful way to capture the patterns and context of 3-D woody vegetation structure?

Location

Classification development and testing sites: landscapes in Bushbuckridge Municipality, Sabi Sand Wildtuin and Kruger National Park, Mpumalanga province, north-east South Africa.

Methods

The aforementioned structural classification approach requires appropriate 3-D and spatially explicit remote sensing data. A LiDAR-based canopy height model (CHM) and volumetric pixel (voxel) data from the Carnegie Airborne Observatory Alpha system were used to create the structural classification. First, we segmented the CHM images using multi-threshold and multi-resolution image segmentation techniques, and classified the image segments into four height classes, namely shrub (1–3 m), low tree (3–6 m), high tree (6–10 m) or tall tree (>10 m). A hierarchical a priori approach was used to develop classification criteria. The following metrics were calculated for 0.25-ha grid cells based on the cover and spatial arrangement of the four height classes: canopy cover, sub-canopy cover, canopy layers, Simpson's diversity index and cohesion. Top of canopy vegetation was classified using each metric at the 0.25-ha scale, with canopy cover being the primary classification metric. Subsequently, individual layers identified within the canopy were classified using the voxel data. We use a code system for describing classes to ensure standardization between different regions; a more traditional naming system may be used in addition for interpretation.

Conclusion

This system provides a more comprehensive classification of the horizontal and vertical structural diversity of savannas compared to the traditional vegetation classification systems. The description of multi-layers within the canopy should allow for a sensitive change detection method. The classification can be used in many current focus areas, including habitat suitability mapping for biodiversity conservation, strategic adaptive management and monitoring land-cover change.


Nomenclature
Mucina &

Rutherford (2006)

Abbreviations
CC

canopy cover

CL

canopy layers

GLCC

global land cover classification

IGBP

International Geosphere–Biosphere Project

LCCS

land-cover classification system

LiDAR

light detection and ranging

NLC

National land cover

SCC

sub-canopy cover

SDI

Simpson's diversity index

Introduction

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results and discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Woody vegetation classification maps are inherently two-dimensional based on the remotely sensed data used. The position, extent and connectivity of the woody vegetation layer are captured; however, the vertical arrangement of woody plant components is not visible from standard two-dimensional (2-D) passive remotely sensed data. Recording the three dimensional (3-D) structure of vegetation in the field is time consuming and often not feasible, and is not possible at all with standard multispectral images. However, with the development of LiDAR (light detection and ranging), it is now possible to objectively and repeatedly collect measurements of vertical structure over large areas (Goatley & Bellwood 2011). Savannas, defined by a continuous herbaceous layer with a discontinuous woody layer, possess a complex woody architecture best described in three dimensions. This complex vertical and horizontal structure provides habitat for a broad range of vertebrates and invertebrates; and has implications for conservation and natural resource provision, as the importance of the vertical dimension in habitat heterogeneity across large extents is not well understood (Tews et al. 2004; Hall et al. 2011).

Light detection and ranging has previously not been used extensively in structurally heterogeneous African savannas; however, it has recently been used to successfully map savanna biomass (Colgan et al. 2012), investigate the effects of fire and herbivory on vegetation structure (Asner et al. 2009; Levick et al. 2009; Smit et al. 2010; Asner & Levick 2012), explore patterns of structure around communal rural villages (Wessels et al. 2011; Fisher et al. 2012) and map riparian condition indicators (Johansen et al. 2010). LiDAR produces large amounts of data, so it is often necessary to derive summary statistics of canopy height and inferred estimates of diameter at breast height and above-ground biomass for ecological applications (Lefsky et al. 2002a,b; Naesset 2002; Blaschke et al. 2004; Anderson et al. 2006). While this may provide acceptable measurements of structural attributes in areas with homogeneous vegetation, these statistics do not properly describe the variation especially at large extents in heterogeneous landscapes like African savannas, which have more complex vertical structure with high spatial variability. A method to remedy this over-simplification of LiDAR data, while still reducing data volume and complexity and providing relevant ecological information, would be to use it in a spatially explicit 3-D classification. Small footprint LiDAR can address the deficiencies of conventional 2-D savanna classifications (Appendix S1) by providing a 3-D component (plant height as well as layers within the canopy), without the need for extensive fieldwork at a scale relevant to capturing the heterogeneity of savannas.

Such an approach remains challenging, since the spatio-temporal heterogeneity in the horizontal and vertical structure of woody plants in savannas adds complexity when studying pattern and process (Levick & Rogers 2011). Complex patterns in vegetation 3-D structure cannot be effectively characterized by a single measure, as they are driven by climate, rainfall, geology, topography, fire and herbivory (Scholes & Archer 1997; Sankaran et al. 2008), which vary across space and time (Levick & Rogers 2011). The resultant patterns in savannas are not only determined by an individual component, but more importantly, by the interactions between components (Pickett et al. 2003). The spatial context of woody vegetation in the landscape is therefore necessary for appropriate application of the knowledge to management and conservation (Levick & Rogers 2011).

Woody vegetation structure refers to the position, extent, quantity, type and connectivity of the above-ground components of woody vegetation (Lefsky et al. 2002a) in three dimensions. Therefore, each of these characteristics needs to be measured in order to adequately represent savanna vegetation. Although theoretical methods such as volumetric neutral models capture 3-D spatial structure of vegetation (Kirkpatrick & Weishampel 2005), no ecologically-based classification currently exists that captures this type of heterogeneity in savanna vegetation structure. Five land-cover classifications include a savanna component: structural classification of Edwards (1983), the land-cover classification system (LCCS; Di Gregorio & Jansen 2000), the National Land Cover (NLC) for South Africa (Thompson 1996), International Geosphere–Biosphere Project (IGBP; Loveland & Belward 1997) and the global land cover classification (GLCC; Hansen et al. 2000; Appendix S1). Except for Edwards and LCCS, which are acknowledged for including structural measures, the other classifications (NLC, GLCC and IGBP) do not account for a shrub layer interspersed within the tree layer. The inclusion of the shrub layer is essential as increases in shrubs may indicate bush encroachment, with implications for ecosystem function and biodiversity (Eldridge et al. 2011; Buitenwerf et al. 2012). In addition, the finer scale spatial arrangement of the woody layer as a whole, and of each cover type (tree/shrub), is not taken into consideration in any of these classifications.

It can be argued that a spatial metric such as ‘cohesion’ would give information on the extent, subdivision and contagion of cover classes, indicating habitat suitability (Ishii et al. 2004). At a landscape scale, a low cohesion value could indicate fragmentation, which affects ecological flows within the landscape (McGarigal et al. 2002). Furthermore, habitat suitability is not only determined from the cohesion or fragmentation of the plant canopies in the area, but also the diversity of vegetation structure. Diversity indices such as Shannon–Weiner and Simpson's are commonly used to characterize species diversity (Magurran 2004) and can be applied to structural data (MacArthur & MacArthur 1961). However, diversity indices have not been used in existing land-cover classifications, possibly because different height classes are not identified using conventional multispectral imagery.

Remotely sensed vegetation structural classifications have evolved over the years as image resolution improved, making it possible to now include data to capture the third dimension. There is much benefit from such a classification in savannas, ranging from mapping natural resource availability for ecosystem services, to improved biomass and thus carbon estimates, and enhanced habitat modelling for biodiversity conservation (Ishii et al. 2004; Hall et al. 2011). The aim of the study is therefore to develop an ecologically meaningful savanna classification that captures both the vertical and horizontal heterogeneity of the woody plant canopy, using novel 3-D remote sensing approaches.

Methods

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results and discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

Light detection and ranging (LiDAR)

Woody vegetation was mapped across ca. 35 000 ha of semi-arid savanna in South Africa in April 2008 with the Carnegie Airborne Observatory Alpha System (CAO-Alpha; http://cao.ciw.edu). The CAO-Alpha combined both imaging spectroscopy (hyperspectral imaging) and LiDAR technologies to study ecosystems at the regional scale (Asner et al. 2007). The spectrometer was co-mounted with the LiDAR sensor that acquires both waveform- and discrete-return data; however, only discrete-return data were used in this study.

The integrated global positioning system–inertial motion unit sub-system in the CAO provides the position and orientation of the sensors in 3-D, while the CAO algorithms ensure that data inputs from both the spectrometer and the LiDAR system are co-located and precisely projected to ensure geographically aligned output (Asner et al. 2007). The LiDAR data were collected at 2000 m above ground level with a laser pulse repetition frequency of 50 kHz, laser spot spacing of 1.12 m, and up to four returns per pulse. These specifications are considered to be a minimum requirement for the classification to remain consistent across data sets.

Light detection and ranging produces a 3-D xyz point cloud. A digital elevation model (DEM – interpolated from the LiDAR ground returns) was subtracted from the digital surface model (DSM – LiDAR first return interpolation) to produce the canopy height model (CHM, 1.12 m horizontal pixel resolution). The point cloud frequency values were binned into volumetric pixels (voxels) of 5 × 5 × 1 m (X, Y, Z) for 3-D vegetation analysis. The value in the voxel represents the frequency of LiDAR returns/25 m3 relative to the sum of returns for the entire 5 × 5 m vertical column and is used to assess sub-canopy vegetation. Ground validation of vegetation heights was conducted concurrent to the aerial campaign in 2008. It should be noted that trees less than 2-m tall may be underestimated (Wessels et al. 2011) due to the laser pulse not hitting their small and often sparse canopies.

Test data – site description

The classification was created and tested on sites in communal rangelands in Bushbuckridge Municipality (BBR), and in two adjacent protected areas, Sabi Sand Wildtuin (SSW; a private game reserve) and Kruger National Park (KNP), in Mpumalanga Province, north-east South Africa (Fig. 1). Due to the mosaic of land management techniques and land-use intensities, spatial heterogeneity is high in these areas. This property makes them appropriate sites on which to develop the classification, as they are representative of a wide variety of situations present in global savannas. The sites form a west-to-east gradient in climate and topography. Mean annual precipitation over the study area ranges from >1200 mm in the west, and gradually reduces to an average of 550 mm in the east, with mean annual temperature of 22 °C. The geology of the region is dominated by granite, with Timbavati gabbro intrusions. All sites fall within three vegetation units of the savanna biome: Granite lowveld (dominant), Gabbro grassy bushveld and Legogote sour bushveld (Mucina & Rutherford 2006). Typical woody plant species in the granite lowveld include: Terminalia sericea, Combretum zeyheri and Combretum apiculatum on the deep sandy toplands, and Acacia nigrescens, Dichrostachys cinerea and Grewia bicolor on the more clayey soils of the bottomlands. In the two other vegetation units, additional common species include Sclerocarya birrea, Lannea schweinfurthii, Ziziphus mucronata, Dalbergia melanoxylon, Peltoforum africanum and Pterocarpus rotundifolius.

image

Figure 1. Study location – Bushbuckridge (BBR), Sabie Sands Wildtuin (SSW) and Kruger National Park (KNP) in Mpumalanga Province, South Africa.

Download figure to PowerPoint

Conceptual approach of the classification

Our classification approach was based on a compilation of criteria used in the National Vegetation Classification System (FGDC 1997) and the LCCS (Di Gregorio & Jansen 2000). The classification must furthermore adhere to the following:

  1. have ecologically meaningful metrics
  2. be based on a sound scientific approach that is a logical progression from historical methods and can be repeated
  3. meet the needs of a variety of users
  4. provide a common reference system, and facilitate comparisons between classes used in different classifications
  5. be a flexible system, which can be used at different scales and at different levels of detail, allowing cross-referencing of local and regional features with continental maps without loss of information
  6. be hierarchically organized such that it can be applied at multiple scales
  7. identify spatial units that are appropriately scaled to meet objectives for biodiversity conservation, as well as resource and ecosystem management needs
  8. be a flexible system which is open ended such that it will allow for additions, modifications and continuous refinement
  9. be a well documented system that is easy to execute

.

We used a hierarchical a priori approach to develop the classification criteria. When conducting a global classification, it is often easier to use a data-driven (a posteriori) classification in order to reduce the amount of user interaction, as no prior or local knowledge of the area is a pre-requisite (Achard et al. 2001). However, such methods rely on spectral separability being equated to ecologically meaningful classes, which is not necessarily the case, especially when performing a structural classification. The wealth of existing information about savannas ensures that we can define a priori classes that adequately represent the 3-D nature of savannas (Scholes & Walker 1993; Scholes & Archer 1997; Sankaran et al. 2008). The nature of an a priori classification system is such that category definitions are independent of (1) the area mapped; (2) the data properties; and (3) the mapping techniques, thus making the classification more robust and universally applicable. The hierarchy can be described as a compositional containment hierarchy, where no one metric is more important than the others; however, each by itself is meaningless without the context provided by other metrics and size classes (Parsons 2002).

The classification was based on two levels (Fig. 2). The first level in the hierarchy classifies the top of canopy vegetation based on canopy cover, percentage canopy layers present (derived from the voxel product), cohesion and diversity (Fig. 2). Top of canopy vegetation includes all vegetation captured by the CHM and does not include understorey vegetation. The second level of the classification categorizes each height class present, including both vegetation that appears on the CHM and understorey (sub-canopy) vegetation (Fig. 3). There are four possible canopy layers: shrub (1–3 m), low tree (3–6 m), high tree (6–10 m) and tall tree (>10 m). The cover and cohesion of each layer is described, starting with shrubs and ending with tall trees. If a layer is not present it is excluded from the description (Figs. 2, 4). Level I of the classification is a top-down classification, while Level II is a bottom-up classification (Fig. 2).

image

Figure 2. Classification metrics and how they are combined for a savanna woody structural classification.

Download figure to PowerPoint

image

Figure 3. Schematic representation of savanna plants within a 0.25-ha area shown in both 2-D (top of canopy vegetation – view from above) and 3-D (lateral view). Vacant canopy layers are shown (four layers) and indicate, along with filled canopy spaces (six layers), the number of canopy layers present (6/10; 60%) in the 0.25-ha area.

Download figure to PowerPoint

image

Figure 4. Aerial view and transects through the 3-D point cloud of four 0.25-ha areas of semi-arid savanna with corresponding classifications. The height key refers to vegetation in the aerial view. Height classes are not differentiated in the point clouds. The corresponding classifications of the area using Edwards (1983), land cover classification system (LCCS; Di Gregorio & Jansen 2000), National Land Cover Classification of South Africa (NLC; Thompson 1996), global land cover classification (GLCC; Hansen et al. 2000) and International Geosphere–Biosphere Programme (IGBP; Loveland & Belward 1997).

Download figure to PowerPoint

Traditional land-cover classifications place emphasis on the name of the class; however this may lead to confusion as one land-cover type may be called a different name under two classifications systems (Appendix S1, Fig. 4). We therefore adopted the technique used in the LCCS whereby a code is used to define a class. This makes the classification comparable between countries, which might use alternative names for a vegetation type. In addition, the code system is more robust when investigating change, as the specific metric of the class that is changing, e.g. the level of aggregation, is identified.

Classification development

The building blocks of the classification are individual trees and shrubs. In accordance with Edwards (1983), four growth forms were classified in agreement with the canopy layers: shrub (1–3 m), low tree (3–6 m), high tree (6–10 m) and tall tree (10+ m). Class intervals at the lower end were inclusive and exclusive at the upper end (i.e. 1–3 m height class: 1 m ≤ trees < 3 m). The selected height categories are ecologically meaningful and relate to fire, herbivory and human use. Vegetation <3 m in height is in the fire trap (Govender et al. 2006; Smit et al. 2010) and heavily browsed by small- to medium-size herbivores; vegetation in the 3–6 m height class is targeted by mega-herbivores [elephant (Loxodonta africana) and giraffe (Giraffa camelopardalis)]; finally, vegetation >6 m is less influenced by fire and herbivory (Owen-Smith 1988; Scholes & Walker 1993; Birkett & Stevens-Wood 2005; Neke et al. 2006). People are known to harvest wood for fuel and poles, predominantly from <3 m height class (Neke et al. 2006), although in miombo woodlands where wood is used for charcoal production the entire tree is often harvested (Luoga et al. 2000). Trees >10 m are important in the savanna landscape, providing shade, reducing evapotranspiration of the below-canopy herbaceous layer, and increasing local nutrients accumulated close to the root systems (Belsky 1994; Manning et al. 2006; Treydte et al. 2009), thereby creating high-quality grazing, which may attract higher abundances of ungulates.

Individual vegetation units were identified on the CHM and voxel layers using an object-based image analysis (OBIA) in eCognition Developer v. 8.7 (Trimble Geospatial Imaging, Munich, Germany, 2011). The CHMs were treated with a 3 × 3 low pass filter prior to segmentation to remove noise. A multi-threshold segmentation was performed using the following height thresholds: 0.25; 0.5; 1.0; 1.5; 2.0; 3.0; 4.0; 5.0; 6.0; 7.0; 8.0; 9.0 and 10.0 m. These thresholds allowed for hierarchical segmentation aggregation from a very fine sub-canopy scale (Appendix S2a) to individual tree canopies (Appendix S2e). Image objects with a mean and/or maximum vegetation height <1 m were classified as ‘background’ and removed from further classifications. After initial segmentation, each object was classified into one of the four height classes (1–3 m, 3–6 m, 6–10 m, >10 m) based on maximum height in each image object. As a result of height uniformity in large clumps of trees, and inter-canopy variation in large trees, image objects in areas with high woody cover were not adequately segmented and large trees were often too finely segmented, with canopies consisting of multiple segments (Appendix S2a); however, coarser height thresholds did not identify smaller, often isolated tree canopies. These fine image objects were therefore merged according to their height classification; and a subsequent multi-resolution segmentation was performed on these merged image objects [Segmentation parameters used in eCognition v. 8.7 (2011): scale parameter = 12 (determines size of segmentation in relation to the landscape), shape weighting = 0.5 (0 = irregular shape; 1 = regular shape) and compactness weighting = 0.9 (0 = high perimeter:area ratio; 1 = low perimeter:area ratio)] creating a second segmentation layer. The resulting image objects, which contained finer detail in areas of dense woody cover, were then reclassified into the four vegetation height classes based on maximum height in an image object.

The 3-D structural classification is such that once image objects of individual vegetation units have been created, the classification can be carried out at a variety of user specified scales according to user need. The minimum grid size for the classification was determined using semivariograms calculated in ENVI 4.7 (ITT Vis; ITT Visual Information Solutions, Boulder, CO, US, 2009) on the CHM and it was established that the variogram sill occurred at a range of 50 m, translating to a grid size of 0.25 ha (Wessels et al. 2011). Metrics were calculated for each grid cell using the four vegetation height classes exported from eCognition. The following metrics were calculated in ArcMap 10.0 (ESRI 2010, Redlands, CA, USA, www.esri.com): canopy cover (CC), sub-canopy cover (SCC), canopy layers (CL), cohesion and Simpson's diversity index (SDI; Table 1). CC is classified into nine cover classes, which were chosen because of overlap with existing classifications (Appendix S1, Fig. 2).

Table 1. Description of metrics used to construct a three-dimensional structural classification of savanna woody vegetation
MetricDescription and schematic representation

Canopy cover (CC)

Units:%

Range: 0–100

CC refers to the vertical projection of the tree/shrub crown onto the ground, given as a percentage of the area. Cover is measured for the overall woody cover (all height classes). The dominant cover class is measured from the CC metric as the class that constitutes ≥ 50% of the total woody canopy cover. Canopy cover is measured from the top of canopy objects produced in eCognition v. 8.7 (2011) based on the canopy height model (CHM) LiDAR product.image_n/avsc12048-gra-0001.png

Sub-canopy cover (SCC)

Units:%

Range: 0–100

SCC is a measurement of the percentage cover of each height class (1–3 m, 3–6 m, 6–10 m and >10 m) as it occurs below the dominant cover classes. That is, a tree of >10 m may obscure vegetation below. It is measured as a percentage of the grid cell for each height class. Each individual SCC measurement for each height class will fall within the range of 0–100%. SCC is only used to calculate cohesion for each individual height class, and is used as the cover metric in the description of Level II – Plant layers of the classification. SCC is measured from the volumetric pixel (voxel) data.image_n/avsc12048-gra-0002.png

Canopy layers (CL)

Units:%

Range: 0–100

CL is a measure of the percentage of canopy layers present within the canopy. This metric quantifies the thickness of the woody layer. An increase in CL over time might be an indication of bush encroachment. CL is calculated for the entire grid cell of interest using the SCC product (a presence/absence measure – indicated by the solid cylinders and dashed cylinders, respectively, in the figure). It is a measurement of the number of vertical canopy layers present relative to the total possible number of canopy layers (for a tree >10 m four layers are possible) available in each grid cell including the top of canopy object.image_n/avsc12048-gra-0003.png

Cohesion

Units:%

Range: 0–100

Cohesion is a measure of how aggregated the vegetation components (trees and shrubs) are within the designated area in the horizontal plane. Values range between 0 and 100, with 100 representing increased aggregation or clumping. Due to the mix of grass and woody components defining savannas, spatial arrangement is an important consideration with implications for habitat suitability and utilization. At a fine scale, cohesion has implications for organisms' movement and use of the landscape, while at a landscape scale cohesion gives an indication of edge effects (Fischer & Lindenmayer 2007). An increase in cohesion of one or more vertical height classes may indicate increased bushiness. Often, cohesion is inversely proportional to interspersion, with a high cohesion value indicating a low level of interspersion of cover types. Cohesion was measured for both the entire woody layer within the grassland matrix (using the CC metric), as well as for each height class (using the SCC metric) to measure the cohesion of each height class. The following equation was used to calculate cohesion (McGarigal et al. 2002):

inline image

where:

Pij = perimeter of patch ij (either woody vegetation, or each height layer) in terms of number of pixels

aij = area of patch ij in terms of pixels

= total number of pixels in the landscape

Values were then corrected according to percentage area covered.image

Simpson's diversity index (SDI)

Units:%

Range: 0–100

SDI takes into account both the number of height classes present, and their relative abundance. SDI is a measure of structural diversity; the higher the value, the higher the likelihood that two objects within a grid cell differs (i.e. mixture of shrubs, low trees, high trees and tall trees; i.e. increased diversity). The metric is calculated from the tree and shrub objects layer based on the CHM using the following equation (McGarigal et al. 2002):

inline image

where:

ni = number of individuals of height class i

= total number of individuals (trees identified using object-based image analysis) of all height classes

An example of using the classification is as follows: a grid cell contains one tall tree (>10 m), one high tree (6–10 m), one low tree (3–6 m) and one shrub (1–3 m; Fig. 3). The tall tree has three layers within its canopy, representing vegetation in the >10 m, 6–10 m and 1–3 m height classes (Fig. 3). Top of canopy vegetation (CC) covers 37% of the grid cell, percentage of canopy layers present (CL) is 60% (i.e. 60% of the possible sub-canopy layers within the 0.25-ha grid cell are present, Fig. 3), cohesion equals 57, SDI is 87, and all four height classes are present, therefore we use the code E4. Since no height class covers >50% of the total percentage cover, we use the code E40. The classification of Level I in the hierarchy is therefore A6B60C57D87E40 (moderately covered, evenly dispersed, diverse savanna, with understorey vegetation; Table 2). Sub-canopy layers are described in height order from shrub to tall tree. Subsequent layers are therefore given the following codes: e1a3c15, e2a2c0, e3a3c15 and e4a2c0. Lowercase letters are used to indicate layers within the sub-canopy. The resulting full code for the grid cell is thus: A6B60C57D87E40 e1a3c15 e2a2c0 e3a3c15 e4a2c0.

Table 2. Examples of two areas classified using the structural classification and suggested names for each code. The breaks in the code shown below (canopy cover and volume, cohesion and diversity, etc.) are suggested break points along the classification hierarchy where users may end their classification
Classifier usedCodeSuggested name
Example 1 – No dominant layer
Canopy cover & % layersA6B60Moderately covered savanna with understorey vegetation
Cohesion and diversityA6B60C57D87Moderately covered evenly dispersed diverse savanna with understorey vegetation
Life forms present and dominanceA6B60C57D87E40Moderately covered evenly dispersed diverse savanna with understorey vegetation
Example 2
Canopy cover & % layersA4B4Open savanna with understorey vegetation
Cohesion and diversityA4B4C82D23Aggregated, even open savanna with understorey vegetation
Life Forms present and dominanceA4B4C82D23E32Low tree aggregated even open savanna with understorey vegetation
Understorey layersA4B4C82D23E32 e1a2c12 e2a3c61Multi-layered low tree aggregated open savanna with shrubs

Results and discussion

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results and discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

The 3-D structural classification put forward in this paper creates a standard for comparison with existing vegetation classifications (Appendix S1), while at the same time incorporating novel 3-D technology, creating an ecologically meaningful and useful classification. Using an adaptation of the well-known LCCS code system, the structural classification can be used on LiDAR data in different countries with different naming conventions, but remain comparable. We propose a set of suggested names for classes (Fig. 2, Table 2); however, the order of the names for each metric within the full name may be modified as long as the code remains consistent. While the code may become cumbersome, it can be shortened according to user needs (i.e. only report Level I) or according to available information (Table 2). An intermediate option is to present the classification for the top of canopy layer and for just the dominant height class layer if one is present (Table 2).

While other classifications can identify changes in land cover (e.g. Edwards 1983; Di Gregorio & Jansen 2000), the change has to be considerable before being detected. Conversion from one land-cover category to another through land-use change is easily identified, as cover is drastically altered (e.g. clear-cutting or planting trees). However, modifications within one land-cover category through land-use intensification, especially when the changes are occurring below the top canopy (e.g. fuelwood harvesting or coppice regeneration), are difficult to detect with traditional land-cover classifications (Jansen & Di Gregorio 2002). The 3-D classification provides an advantage over Edwards' (1983) classification, which classifies the amount of cover of four life forms (trees, shrubs, grasses and herbs) and describes vegetation based on aerial cover of dominant life form (e.g. high closed woodland). Spatial configuration and number of layers within the canopy are not included. LCCS does provide more detail than Edwards, such as leaf type and phenology (e.g. broad-leaved, deciduous), as well as information on the stratification of the canopy; however, stratification only refers to life forms that can be identified from an aerial view and does not include sub-canopy layers (Di Gregorio & Jansen 2000).

The 3-D structural classification will be able to identify subtle changes in sub-canopy vegetation density and spatial arrangement before a state shift occurs, by identifying changes in height class dominance, as well as changes in cover, cohesion and diversity. The metrics can be used as part of a monitoring system contributing to better management by early detection of areas of concern, such as areas with woody encroachment, loss of big trees or excessive fuelwood removal; which might not be detected with traditional 2-D classifications that only use percentage canopy cover as a classifier. For change detection, we would recommend using the largest amount of detail (i.e. use the code for all layers within the canopy) to ensure increased sensitivity to identify change. Changes may also be investigated separately for each metric for ease of interpretation (Fig. 5, Appendix S3).

image

Figure 5. True colour image of an area in Bushbuckridge Municipality (a) Mpumalanga Province, north-west South Africa and the corresponding Level I classifications (b. Canopy cover, c. Cohesion, d. Canopy layers, e. Number of height classes present and f. Simpson's diversity index) using the 3-D woody structural classification for savannas.

Download figure to PowerPoint

Areas with the same cover may have different structural compositions (Fig. 4), which will result in dissimilar functional habitats. The two areas might also contain varying assemblages of height classes, different arrangement of these height classes within the area and different canopy layers. Vegetation structural complexity has been shown to increase species richness and diversity of both small mammals and reptiles (Price et al. 2010). Birds (MacArthur & MacArthur 1961; Bergen et al. 2007; Seymour & Dean 2009), arthropods (Halaj et al. 2000), mammals (Williams et al. 2002; Lumsden & Bennett 2005) and reptiles (Smart et al. 2005) rely on fine-scale spatial niches created by complex vertical architecture present in savannas for their habitat. In addition, ungulates, both browsers and grazers, and predators show definite preferences for areas with different amounts of woody cover, ranging from grasslands to densely wooded areas (De Knegt et al. 2007; Winnie et al. 2008). The structural classification provides the level of detail needed to map areas of suitable habitat, which is essential for effective management and conservation of biodiversity. Furthermore, the classification makes it possible to monitor heterogeneity throughout the landscape. Maintenance of this heterogeneity is an explicit management goal pursued by some conservation agencies (e.g. SANParks Thresholds of Potential Concern for heterogeneity; http://www.sanparks.org/) to facilitate biodiversity conservation (Rogers 2003; Ishii et al. 2004).

A further advantage of a 3-D classification, over one that is 2-D, is that it is a combination of plant structure and percentage cover, which influences biomass and subsequently estimates of carbon. While numerous methods are available to monitor carbon stocks using satellite remote sensing (see Goetz et al. 2009), it is unclear how accurate and precise these estimates of biomass and biomass change are (Maniatis & Mollicone 2010). The 3-D structural classification can be used to improve understanding on the relationship between biomass and habitat structure. Current biomass estimates for savanna vegetation are derived from adult, often single-stemmed, trees (Colgan et al. 2012), yet they are applied to multi-stemmed coppicing vegetation and may contain up to 40% error. In communal rangelands in southern Africa, where large areas of vegetation are coppiced, biomass might be overestimated using the standard allometry. The necessity to estimate biomass more accurately highlights the need to map the 3-D structure of vegetation (Hall et al. 2011). The 3-D maps of savannas would provide increased monitoring potential to identify subtle changes and increased thickening of these woody components (Jansen & Di Gregorio 2002; Hall et al. 2011).

In order to ensure that the proposed structural classification method is comparable with existing classifications (Appendix S1), we used codes as a naming system and percentage canopy cover as a primary classifier. The cover classes chosen here are narrow, but can be combined to be directly comparable to existing classifications (Appendix S1). The classification adds to existing classifications not only by including understorey layers, but also in the description of the spatial arrangement of the woody components in terms of their cohesion and diversity of woody layers within each area. These metrics aid in classifying the landscape in an ecologically meaningful way, especially for habitat suitability mapping (McGarigal et al. 2002). Although we used a static grid for the classification, we do acknowledge that a grid has arbitrary boundaries and vegetation often has no clearly defined boundaries (Fisher 1997). A solution would be to use a moving window analysis, with the kernel size equal to distance at which spatial autocorrelation ceases, in this case 50 m, providing a spatially continuous description of the vegetation. This may, however, impact on change detection analyses.

We provide a classification method to reduce the large volumes of data associated with LiDAR while still capturing the spatially variable structural heterogeneity in savannas. In addition, since the classification can be done over large extents, the context of the structural patterns is captured. This aids in understanding the drivers of savanna woody structure and can be used in regional change predictions. The 3-D maps of woody vegetation structure for conservation and resource planning would be invaluable. In addition, the structural maps can be used to model the potential percolation of fire through the landscape (Archibald et al. 2012) as well as mapping surface roughness parameters, which will affect storm surge (Medeiros et al. 2012). Although future satellite-borne LiDAR campaigns are in the process of being planned, such as ICESat II, airborne LiDAR is currently the only method available to collect high resolution 3-D information to detect individual tree canopies (Hall et al. 2011).

Acknowledgements

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results and discussion
  6. Acknowledgements
  7. References
  8. Supporting Information

The Carnegie Airborne Observatory is supported by the Gordon and Betty Moore Foundation, John D. and Catherine T. MacArthur Foundation, Grantham Foundation for the Protection of the Environment, Avatar Alliance Foundation, W.M. Keck Foundation, Margaret A. Cargill Foundation, Mary Anne Nyburg Baker and G. Leonard Baker Jr., and William R. Hearst III. Application of the CAO data in South Africa is made possible through the Andrew Mellon Foundation and the endowment of the Carnegie Institution for Science. We thank D. Knapp, T. Kennedy-Bowdoin, J. Jacobson and R. Emerson for analysing the LiDAR data. Additional funding was supplied by the Carnegie Foundation of New York through the Global Change and Sustainability Research Institute at the University of the Witwatersrand, Johannesburg, ZA.

References

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results and discussion
  6. Acknowledgements
  7. References
  8. Supporting Information
  • Achard, F., Eva, H. & Mayaux, P. 2001. Tropical forest mapping from coarse spatial resolution satellite data: production and accuracy assessment issues. International Journal of Remote Sensing 22: 27412762.
  • Anderson, J.E., Martin, M.E., Smith, M.-L., Dubayah, R.O., Hofton, M.A., Hyde, P., Peterson, B.E., Blair, J.B. & Knox, R.G. 2006. The use of waveform lidar to measure northern temperate mixed conifer and deciduous forest structure in New Hampshire. Remote Sensing of Environment 105: 248263.
  • Archibald, S., Staver, A.C. & Levin, S.A. 2012. Evolution of human-driven fire regimes in Africa. Proceedings of the National Academy of Sciences of the United States of America 109: 847852.
  • Asner, G.P. & Levick, S.R. 2012. Landscape-scale effects of herbivores on tree fall in African savannas. Ecology Letters 15: 12111217.
  • Asner, G.P., Knapp, D.E., Kennedy-Bowdoin, T., Jones, M.O., Martin, R.E., Boardman, J. & Field, C.B. 2007. Carnegie Airborne Observatory: in-flight fusion of hyperspectral imaging and waveform light detection and ranging (wLiDAR) for three-dimensional studies of ecosystems. Journal of Applied Remote Sensing 1: 013536.
  • Asner, G.P., Levick, S.R., Kennedy-Bowdoin, T., Knapp, D.E., Emerson, R., Jacobson, J., Colgan, M.S. & Martin, R.E. 2009. Large-scale impacts of herbivores on the structural diversity of African savannas. Proceedings of the National Academy of Sciences of the United States of America 12: 261278.
  • Belsky, A.J. 1994. Influences of trees on savanna productivity: tests of shade, nutrients, and tree–grass competition. Ecology 75: 922932.
  • Bergen, K.M., Gilboy, A.M. & Brown, D.G. 2007. Multi-dimensional vegetation structure in modelling avian habiat. Ecological Informatics 2: 922.
  • Birkett, A. & Stevens-Wood, B. 2005. Effect of low rainfall and browsing by large herbivores on an enclosed savannah habitat in Kenya. African Journal of Ecology 43: 123130.
  • Blaschke, T., Tiede, D. & Heurich, M. 2004. 3D landscape metrics to modelling forest structure and diversity based on laser scanning data. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences XXXVI-8/W2: 129132.
  • Buitenwerf, R., Bond, W.J., Stevens, N. & Trollope, S.W. 2012. Increased tree densities in South African savannas: >50 years of data suggest CO2 as a driver. Global Change Biology 18: 675684.
  • Colgan, M.S., Asner, G.P., Levick, S.R., Martin, R.E. & Chadwick, O.A. 2012. Topo-edaphic controls over woody plant biomass in South African savannas. Biogeosciences Discussions 9: 957987.
  • De Knegt, H.J., Hengeveld, G.M., van Langevelde, F., de Boer, W.F. & Kirkman, K.P. 2007. Patch density determines movement patterns and foraging efficiency of large herbivores. Behavioral Ecology 18: 10651072.
  • Di Gregorio, A. & Jansen, L.J.M. 2000.Land Cover Classification System (LCCS): classification concepts and user manual LCCS version 1.0. GCP/RAF/287/ITA Africover, Food and Agriculture Organization of the United Nations Publishing Service, Rome, IT.
  • Edwards, D. 1983. A broad-scale structural classification of vegetation for practical purposes. Bothalia 14: 705712.
  • Eldridge, D.J., Bowker, M.A., Maestre, F.T., Roger, E., Reynolds, J.F. & Whitford, W.G. 2011. Impacts of shrub encroachment on ecosystem structure and functioning: towards a global synthesis. Ecology Letters 14: 709722.
  • FGDC (Federal Geographic Data Committee, Vegetation Subcommittee). 1997. Vegetation classification standard. FGDC, VA. FGDC-STD-005.
  • Fischer, J. & Lindenmayer, D.B. 2007. Landscape modification and habitat fragmentation: a synthesis. Global Ecology and Biogeography 16: 265280.
  • Fisher, P. 1997. The pixel: a snare and a delusion. International Journal of Remote Sensing 18: 679685.
  • Fisher, J.T., Witkowski, E.T.F., Erasmus, B.F.N., van Aardt, J., Asner, G.P., Wessels, K.J. & Mathieu, R. 2012. Human-modified landscapes: patterns of fine-scale woody vegetation structure in communal savannah rangelands. Environmental Conservation 39: 7282.
  • Goatley, C.H.R. & Bellwood, D.T. 2011. The roles of dimensionality, Canopies and complexity in ecosystem monitoring. PLoS ONE 6: e27307.
  • Goetz, S.J., Baccinin, A., Laporte, N.T., Johns, T., Walker, W., Kellndorfer, J., Houghton, R.A. & Sun, M. 2009. Mapping and monitoring carbon stocks with satellite observations: a comparison of methods. Carbon Balance and Management 4: 2.
  • Govender, N., Trollope, W.S. & van Wilgen, B.W. 2006. The effect of fire season, fire frequency, rainfall and management on fire intensity in savanna vegetation in South Africa. Journal of Applied Ecology 43: 748758.
  • Halaj, J., Ross, D.W. & Moldenke, A.R. 2000. Importance of habitat structure to the arthropod food-web in Douglas-Fir canopies. Oikos 90: 139152.
  • Hall, F.G., Bergen, K., Blair, J.B., Dubayah, R., Houghton, R., Hurtt, G., Kellndorfer, J., Lefsky, M., Ranson, J., Saatchi, S., Shugart, H.H. & Wickland, D. 2011. Characterizing 3D vegetation structure from space: mission requirements. Remote Sensing of Environment 115: 27532775.
  • Hansen, M.C., DeFries, R.S., Townshend, J.R.G. & Sohlberg, R. 2000. Global land cover classification at 1 km spatial resolution using a classification tree approach. International Journal of Remote Sensing 21: 13311364.
  • Ishii, H.T., Tanabe, S. & Hiura, T. 2004. Exploring the relationships among canopy structure, stand productivity, and biodiversity of temperate forest ecosystems. Forest Science 50: 342355.
  • Jansen, L.J.M. & Di Gregorio, A. 2002. Parametric land cover and land-use classifications as tools for environmental change detection. Agriculture, Ecosystems and Environment 91: 89100.
  • Johansen, K., Arroyo, L.A., Armston, J., Phinn, S. & Witte, C. 2010. Mapping riparian condition indicators in a sub-tropical savanna environment from discrete return lidar data using object-based image analysis. Ecological Indicators 10: 796807.
  • Kirkpatrick, L.A. & Weishampel, J.F. 2005. Quantifying spatial structure of volumetric neutral models. Ecological Modelling 186: 312325.
  • Lefsky, M.A., Cohen, W.B., Parker, G.G. & Harding, D.J. 2002a. Lidar remote sensing for ecosystem studies. BioScience 52: 1930.
  • Lefsky, M.A., Cohen, W.B., Harding, D.J., Parker, G.G., Acker, S.A. & Gower, S.T. 2002b. Lidar remote sensing of above-ground biomass in three biomes. Global Ecology and Biogeography 11: 393399.
  • Levick, S.R. & Rogers, K.H. 2011. Context-dependent vegetation dynamics in an African savanna. Landscape Ecology 26: 515528.
  • Levick, S.R., Asner, G.P., Kennedy-Bowdoin, T. & Knapp, D.E. 2009. The relative influence of fire and herbivory on savanna three-dimensional vegetation structure. Biological Conservation 142: 16931700.
  • Loveland, T.R. & Belward, A.S. 1997. The International Geosphere–Biosphere Programme Data and Information System Global Land Cover Data Set (DISCover). Acta Astronautica 41: 681689.
  • Lumsden, L.F. & Bennett, A.F. 2005. Scattered trees in rural landscapes: foraging habitat for insectivorous bats in south-eastern Australia. Biological Conservation 122: 205222.
  • Luoga, E.J., Witkowski, E.T.F. & Balkwill, K. 2000. Subsistence use of wood products and shifting cultivation within a miombo woodland of eastern Tanzania, with some notes on commercial uses. South African Journal of Botany 66: 7285.
  • MacArthur, R.H. & MacArthur, J.W. 1961. On bird species diversity. Ecology 42: 594598.
  • Magurran, A.E. 2004. Measuring biological diversity. Blackwell, Oxford, UK.
  • Maniatis, D. & Mollicone, D. 2010. Options for sampling and stratification for national forest inventories to implement REDD+ under the UNFCCC. Carbon Balance and Management 5: Available at: http://www.cbmjournal.com/content/5/1/9 (accessed 11 May 2013).
  • Manning, A.D., Fischer, J. & Lindenmayer, D.B. 2006. Scattered trees are keystone structures – Implications for conservation. Biological Conservation 132: 311321.
  • McGarigal, K., Cushman, S.A., Neel, M.C. & Ene, E. 2002. FRAGSTATS: spatial pattern analysis program for categorical maps. Computer software program produced by the authors at the University of Massachusetts, Amherst. Available at: http://www.umass.edu/landeco/research/fragstats/fragstats.html (accessed 11 May 2013).
  • Medeiros, S.C., Hagen, S.C. & Weishampel, J.F. 2012. Comparison of floodplain surface roughness parameters derived from land cover data and field measurements. Journal of Hydrology 452–453: 139149.
  • Mucina, L.& Rutherford, M.C. (eds.) 2006. The vegetation of South Africa, Lesotho and Swaziland: Strelitzia 19, pp. 439539. South African National Biodiversity Institute, Pretoria, ZA.
  • Naesset, E. 2002. Predicting forest stand characteristics with airborne laser scanning using a practical two-stage procedure and field data. Remote Sensing of Environment 80: 8899.
  • Neke, K.S., Owen-Smith, N. & Witkowski, E.T.F. 2006. Comparative resprouting response of Savanna woody plant species following harvesting: the value of persistence. Forest Ecology and Management 232: 114123.
  • Owen-Smith, N. 1988. Megaherbivores: the influence of very large body size on ecology. Cambridge University Press, Cambridge, UK.
  • Parsons, D. 2002. Object Oriented Programming in C++, 2nd ed, pp. 110185. Cengage Learning, Continuum, London, UK.
  • Pickett, S.T.A., Cadenasso, M.L. & Benning, T.L. 2003. Biotic and abiotic variability as key determinants of savanna heterogeneity at multiple spatiotemporal scales. In: Du Toit, J.T., Rogers, K.H. & Biggs, H.C. (eds.) The Kruger experience: ecology and management of savanna heterogeneity, pp. 422446. Island Press, Washington, DC, US.
  • Price, B., Kutt, A.S. & McAlpine, C.A. 2010. The importance of fine-scale savanna heterogeneity for reptiles and small mammals. Biological Conservation 143: 25042513.
  • Rogers, K.H. 2003. Adopting a heterogeneity paradigm: implications for management of protected savannas. In: Du Toit, J.T., Rogers, K.H. & Biggs, H.C. (eds.) The Kruger experience: ecology and management of savanna heterogeneity, pp. 4158. Island Press, Washington, DC, US.
  • Sankaran, M., Ratnam, J. & Hanan, N. 2008. Woody cover in African savannas: the role of resources, fire and herbivory. Global Ecology and Biogeography. 17: 236245.
  • Scholes, R.J. & Archer, S.R. 1997. Tree–grass interactions in savannas. Annual Review of Ecology and Systematics 28: 517544.
  • Scholes, R.J. & Walker, B.H. 1993. An African savanna – Synthesis of the Nylsvley study. Cambridge University Press, Cambridge, UK.
  • Seymour, C.L. & Dean, W.R.J. 2009. The influence of changes in habitat structure on the species composition of bird assemblages in the southern Kalahari. Austral Ecology 35: 581592.
  • Smart, R., Whiting, M.J. & Twine, W. 2005. Lizards and landscapes: integrating field surveys and interviews to assess the impact of human disturbance on lizard assemblages and selected reptiles in a savanna in South Africa. Biological Conservation 122: 2331.
  • Smit, I.P.J., Asner, G.P., Govender, N., Kennedy-Bowdoin, T., Knapp, D.E. & Jacobson, J. 2010. Effects of fire on woody vegetation structure in African savanna. Ecological Applications 20: 18651875.
  • Tews, J., Brose, U., Grimm, V., Tielborger, K., Wichmann, M.C., Schwager, M. & Jeltsch, F. 2004. Animal species diversity driven by habitat heterogeneity/diversity: the importance of keystone structures. Journal of Biogeography 31: 7992.
  • Thompson, M. 1996. A standard land-cover classification scheme for remote sensing applications in South Africa. South African Journal of Science 92: 3442.
  • Treydte, A.C., Grant, C.C. & Jeltsch, F. 2009. Tree size and herbivory determine below-canopy grass quality and species composition in savannahs. Biodiversity and Conservation 18: 39894002.
  • Wessels, K.J., Mathieu, R., Erasmus, B.F.N., Asner, G.P., Smit, I.P.J., van Aardt, J., Main, R., Fisher, J., Marais, W., Kennedy-Bowdoin, T., Knapp, D.E., Emerson, R. & Jacobson, J. 2011. Impact of contrasting land use on woody vegetation structure in the Lowveld savannas of South Africa. Forest Ecology and Management 261: 1929.
  • Williams, S.E., Marsh, H. & Winter, J. 2002. Spatial scale, species diversity, and habitat structure: small mammals in Australian tropical rain forest. Ecology 83: 13171329.
  • Winnie, J.A., Cross, P. & Getz, W. 2008. Habitat quality and heterogeneity influence distribution and behaviours in African buffalo (Syncerus caffer). Ecology 89: 14571468.

Supporting Information

  1. Top of page
  2. Abstract
  3. Introduction
  4. Methods
  5. Results and discussion
  6. Acknowledgements
  7. References
  8. Supporting Information
FilenameFormatSizeDescription
avsc12048-sup-0001-AppenidxS1.pdfapplication/PDF417KAppendix S1. Key features of vegetation structure classifications and land-cover classifications and the subsequent classification of semi-arid savannas using each type.
avsc12048-sup-0002-AppendixS2.pdfapplication/PDF675KAppendix S2. The process of using first a multi-threshold segmentation (a–d) and then a multi-resolution segmentation (e & f) to identify savanna woody vegetation tree canopies using object-based image analysis on a canopy height model derived from light detection and ranging (LiDAR).
avsc12048-sup-0003-AppenidxS3.pdfapplication/PDF1282KAppendix S3. Woody vegetation structural metrics (canopy cover, number of canopy layers present, canopy cohesion, dominant height classes, number of height classes present and Simpson' s diversity index) in 0.25-ha grid cells for eight sites across Kruger National Park, Sabi Sand Wildtuin and Bushbuckridge, South Africa.

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.