Structural biodiversity monitoring in savanna ecosystems: Integrating ...

Chapter 5.2 Structural biodiversity monitoring in savanna ecosystems: Integrating LiDAR and high resolution imagery through object-based image analysis

S.R. Levick, K.H. Rogers Centre for Water in the Environment, University of the Witwatersrand, Private Bag 3, WITS, 2050, South Africa, [email protected]

KEYWORDS: aerial photography, heterogeneity, normalized canopymodel, remote sensing, woody layer extraction ABSTRACT: Savannas are heterogeneous systems characterized by the coexistence of grasses and woody trees. Growing recognition of the importance of the structural component of biodiversity has highlighted the need to understand the spatial distribution and temporal dynamics of woody plant structural diversity. Advances in LiDAR technology have enabled three dimensional information of vegetation to be obtained remotely over large areas. Whilst the use of LiDAR has gained considerable momentum in forested areas there has been limited application to savanna systems. We explore the applicability of LiDAR and object-based image analysis to the monitoring of woody structural diversity in a savanna system. We demonstrate how an object-based approach to image analysis significantly improves the accuracy of woody layer classification form in a heterogeneous landscape. Furthermore we illustrate how standard approaches to LiDAR derived canopy models suffer from interpolation artifacts in savannas, due to the heterogeneity of the woody layer. By integrating LiDAR with high resolution aerial photography, through object-based analysis, these artifacts can be removed to produce a robust canopy model. The object-based integration of LiDAR with aerial imagery holds immense potential for structural diversity monitoring in savannas.

478

S.R. Levick, K.H. Rogers

1 Monitoring structural biodiversity in savanna ecosystems Savannas are heterogeneous environments driven by a wide range of factors at multiple scales. A key characteristic of savanna landscapes is the co-dominance of two life forms – grasses and woody trees (Scholes and Walker 1993). The spatial structure and composition of savannas is controlled primarily at the broad scale by climate and geology, whilst rainfall, topography, soil type, fire and herbivory influence structure at a range of finer scales (Pickett et al. 2003, Gillson 2004a, Sankaran et al. 2005). In addition to being spatially heterogeneous, savannas are highly dynamic over time (Gillson 2004b). The variability of these systems presents challenges to their management and conservation. Management of savanna systems has historically taken place under a balance of nature/homogeneity paradigm (Rogers 2003). The growing recognition of savanna heterogeneity has led to changes in the management of certain savanna systems. In the Kruger Park (South Africa), for example, management has adopted a heterogeneity paradigm that ‘aims to maintain biodiversity in all its facets and fluxes’ (Braack 1997). This paradigm shift reflects a holistic view of biodiversity which incorporates the composition, structure and function of ecological systems at multiple scales (Noss 1990). Given that heterogeneity is considered to be the ultimate source of biodiversity (Pickett et al. 2003), monitoring system heterogeneity should be of high management priority within savanna systems. 1.1 Monitoring savanna heterogeneity remotely Monitoring of savanna vegetation has traditionally taken place through aerial photographic analyses and field surveys. Ground based field monitoring can provide detailed information of changes in vegetation structure over time, but is very time intensive and can only feasibly be conducted over small spatial scales. Fixed point photography can reveal changes in the three-dimensional structure of vegetation, but it suffers the same constraints as field measurements. Extrapolating results obtained at small spatial scales to larger scales is difficult in heterogeneous systems like savannas. Managers need to be able to monitor large spatial areas to in order to encompass system variability. Remote sensing techniques at broader scales therefore need to be employed. Savannas have historically presented numerous challenges to the field of remote sensing. Given their proximity to the tropics, and the regular occurrence of thunderstorms in summer months, cloud free days are rarely

Structural biodiversity monitoring in savanna ecosystems

479

found during the growing season. Synthetic Aperture Radar (SAR), however, holds a lot of potential for monitoring vegetation communities in savanna landscapes as it has the ability to penetrate cloud cover (Menges et al. 1999). SAR can therefore provide valuable insight into temporal changes in ecosystems by enabling land cover monitoring at all times of the year. SAR has also shown potential for broad scale biomass estimation, although accuracy has been shown to fall with increasing biomass and leaf area index (Waring et al. 1995). Fine scale three-dimensional representation of vegetation requires the use of Light Detection and Ranging (LiDAR). In recent years LiDAR has become commercially available and provides a robust means of measuring the three-dimensional structure of terrain and vegetation surfaces remotely. LiDAR has been utilized extensively in forestry applications and has been shown to reliably return ground elevation and tree height data in forested systems (Lefsky et al. 2002). LiDAR has experienced limited use in savanna landscapes, although Dowling and Accad (2003) and Lovell et al. (2003) have explored its potential in Australian savannas. Given that savannas are heterogeneous systems at multiple scales, interpretation of remotely sensed data should be conducted in hierarchical manner which accounts for spatial variation across the landscape. Objectbased image analysis provides a means for achieving this objective. 1.2 The object-based approach to image analysis Object-based image analysis arose through the realization that image objects hold more real world value than pixels alone (Blaschke and Strobel 2001). The software eCognition 4.0, developed by Definiens Imaging, adopts an object-based image analysis approach and provides a platform for incorporating contextual and ancillary data in image classification. The first step in the analysis is the mutiresolutional segmentation of an image into areas of homogeneity. Homogeneity criteria are based on both spectral and shape properties. A bottom-up region merging technique is employed where smaller objects are merged into larger ones based on the criteria set. The approach allows for segmentation at different scales, which is used to construct a hierarchical network of image objects representing the image information in different spatial resolutions simultaneously (Laliberte et al. 2004). The image objects have relationships to both adjacent objects on the same level and objects on different hierarchical levels. Classification is then performed on the image objects, not the pixels, at the desired scale.

480


In this chapter we demonstrate how object-based image analysis of both aerial imagery and LiDAR can aid the structural diversity monitoring process in savanna landscapes. Our first example explores the classification of black and white aerial photography, which is an important record of historical woody cover and is a critical step in temporal change analyses. The second example examines the integration of both LiDAR and color aerial photography for representing the three-dimensional structure of savanna vegetation. The research was conducted in the Shingwedzi Catchment of the northern Kruger Park, South Africa (Figure 1).

Fig. 1. Study location – Shingwedzi Catchment, Kruger Park, South Africa

2 Woody canopy delineation from black and white aerial photographs Black and white aerial photographic records provide a means of exploring vegetation changes over fairly large spatial areas. These records provide valuable evidence of changes in woody vegetation cover over time, but accurately extracting the woody layer has proved difficult. Much of the difficulty in extracting woody cover from aerial photographs stems from the heterogeneity inherent in savanna landscapes.


481

Fig. 2. a) Aerial photograph of a heterogeneous savanna landscape. b) Pixel-based classification of woody vegetation resulting in an overestimate of woody cover and 68% accuracy

Figure 2a, for example, depicts an aerial view of the riparian fringe adjacent to the Phugwane river of the Shingwedzi Catchment. The image was taken in May 2001. A split between dark basaltic soils (top left-hand corner) and white alluvial soils (bottom right-hand corner) runs diagonally through the image. This variation in soil color presents challenges when trying to extract woody cover from the image. The dark basaltic soils are of similar brightness to some of the woody vegetation types. A traditional pixel based classification of woody cover (Figure 2b) fails to extract only the woody plants and overestimates woody coverage immensely. This is primarily the result of some dark soil areas being classified as woody canopy. Four 1km X 1km sites were selected along each of the four rivers of the Shingwedzi catchment for classification validation purposes. Field based validation was not possible due to the historical nature of the photographs. Visual validation was therefore performed whereby 50 woody cover and 50 bare ground points where digitized onscreen for each of the 16 sites. An error matrix was constructed to access woody cover classification accuracy at the 800 validation point locations. A maximum likelihood pixel based approach only achieved 68% accuracy when compared against the validation data. This is clearly not acceptable for monitoring purposes. To im-

482


prove this classification, multi-scaled object-based image analysis is needed. 2.1 Object-based processing of historical aerial photographs for woody canopy extraction eCognition 4.0 was used to conduct multi-resolutional segmentation and classification on Figure 3a. Prior to segmentation the image was filtered with a 3 X 3 low-pass filter to remove excessive variation. Smoothing of layers prior to segmentation helps produce fewer, and more homogeneous image objects, so that individual trees are represented by fewer polygons (Laliberte et al. 2004). A fine level of segmentation was initially chosen to ensure that image objects were small enough to represent individual trees (Figure 3b). Larger scale segmentation was then conducted to group areas of similar vegetation/soil type units together (Figure 3c). The primary aim of this broader segmentation is to provide some spatial context for the smaller ‘tree’ objects at the lower level. Laliberte et al. (2004) used this technique to successfully extract shrubs from aerial photographs of arid rangelands in southern New Mexico. Although there is little difference in brightness between some of the woody trees and the basalt soils in Figure 3a, the difference between the mean of image objects in Figure 3b and the larger image objects in Figure 3c can be used to differentiate trees from soil. During the classification process, first level objects were considered woody cover if they had a mean brightness value of between 0 and 90, as well as a ratio of between 0 and 0.95 relative to their super object. The resulting classification (Figure 3d) was 97% accurate when tested against the validation data. By adopting the object-based approach, contextual and ancillary data can be included in the classification process to produce more robust results.


483

Fig. 3. a) Aerial photograph of heterogeneous landscape. b) Fine scale segmentation c) Large scale segmentation d) Object-based classification of woody cover resulting in 97% accuracy.

484


3 Extracting woody vegetation structural attributes from LiDAR and high resolution aerial photography Aerial photography records are a valuable resource for monitoring changes in vegetation cover over time. They are, however, limited in their ability to depict three-dimensional changes in woody structure. Managers concerned with changes in the vertical structure of trees need to look for alternate monitoring techniques. LiDAR, in conjunction with high resolution aerial photography, provides a remote sensing solution for monitoring vegetation structural diversity in savanna landscapes. The instrument used in this example was a first/last pulse ALTM 1225 (Optech Inc., Canada) with a pulse frequency of 25kHz. The flight path focused on the four major rivers of the Shingwedzi Catchment (Figure 4).

Fig. 4. LiDAR coverage of the four major rivers of the Shingwedzi Catchment

3.1 Normalized vegetation canopy model (nVCM) construction A key advantage of using LiDAR, from a vegetation monitoring point of view, is the ability to create a normalized vegetation canopy model (nVCM). An nVCM is a spatial raster representation of above ground tree height. Normalized vegetation canopy models are widely used in forestry and can be utilized for monitoring changes in tree height and biomass over time. An nVCM is typically constructed from first creating a digital terrain model (DTM) from the LiDAR ground returned points and a digital surface model (DSM) from the total LiDAR points. The DTM can then be subtracted from the DSM to create a surface of above ground elevation. In natural landscapes this generates an nVCM. This system is widely used in


485

forested systems and has been shown to return reliable estimates of above ground vegetation height (Maier et al. 2006, Tiede et al. 2006).

Fig. 5. a) Subset of a standard nVCM derived from DSM and DTM subtraction. b) Regression of field measured woody canopy height against standard nVCM derived from LiDAR. c) nVCM corrected for discontinuous canopy structure through object-based analysis. d) Regression of field measured woody canopy height against the corrected nVCM.

486


We constructed an nVCM by standard DSM-DTM subtraction techniques for a subset of the Shingwedzi dataset. The resulting model produces a visually realistic representation of woody canopy height (Figure 5a). When this subset of the total LiDAR coverage was tested against 200 ground validated data points, however, it is clear that the model overestimates vegetation height in areas of bare ground and grass cover (Figure 5b). Ground validation was performed at 200 stratified random points. Points were located with a differential GPS and canopy height was measured with a Vertex III hypsometer. This artifact is due to the discontinuous nature of savanna canopy layers. In forested systems with continuous canopies, standard nVCM calculations are sufficient, but gaps between savanna trees result in interpolation artifacts between trees which results in an overestimation of tree height in the canopy gaps. In order to address this issue the actual woody canopy coverage needs to be extracted from high resolution aerial imagery. We achieved this by combining the LiDAR data with high-resolution color aerial photography through object-based image analysis. Building on from the black and white photograph workflow, the image was segmented at both a fine and broad scale. The fine scale segmentation (scale parameter = 3, shape factor = 0.2, smoothness = 0.8) delineated individual tree objects and the broad scale segmentation (scale parameter = 250, shape factor = 0.5, smoothness = 0.5) provided context for woody canopy classification. Once the woody canopy layer was clearly defined, it was used as a mask on the standard nVCM to eliminate artifacts created by the gaps in the canopy and to produce an nVCM corrected for discontinuous canopies. The corrected nVCM (Figure 5c) is more robust when tested against the ground validated data (Figure 5d). Interpolation errors between the tree canopies are removed through the object-based image analysis workflow that was followed (Figure 6).


487

Fig. 6. Workflow model for developing a robust normalized vegetation canopy model (nVCM) for landscapes with discontinuous vegetation cover

The cross-sectional profile in Figure 7 runs through both the standard nVCM (grey) and the corrected nVCM (black) and highlights the differences between the two approaches. There is very little difference between the standard and corrected nVCM in areas where trees are present, it is in the gaps between trees that the standard nVCM overestimates above ground canopy height. The cross-section through the standard nVCM returns a mean canopy height of 5.17m with a coefficient of variation equal to 78.975. The corrected nVCM, however, returns a mean of 3.74m and a coefficient of variation equal to 121.16. This has important implications for the monitoring of vegetation structure and diversity. Without applying the object-based approach in savannas, managers may greatly overestimate the above ground canopy height and standing biomass of the system, and underestimate the level of structural diversity.

488


Fig. 7. Cross-section through the standard nVCM (grey) and the nVCM corrected through object-based image analysis (black)

This also has important implications for the ground validation process. The random approach taken to our ground point selection enforced the sampling of both trees and the bare ground between them. If validation points had only been collected at locations where trees were present, as is often the case, the interpolation artifacts and overestimation errors would not have been detected. After refining the workflow on the subset of the LiDAR, the technique was applied across the entire LiDAR coverage. A stratified random sampling technique was used to select 500 canopy and a 500 inter-canopy points across the LiDAR coverage. Points were located with a differential GPS and canopy height was measured with a Vertex III hypsometer. The resulting nVCM (Figure 8) was validated against 1000 ground control points and returned and R-squared value of 0.851 (p < 0.0001).


489

Fig. 8. Corrected normalized vegetation canopy model for the entire study site (R2=0.851, p < 0.0001, 1000 ground control points)

4 Implications for the monitoring of savanna structural diversity The heterogeneity of complex systems at different scales proves problematic for traditional pixel based classification techniques. The object-based approach, however, produces an accurate representation of woody cover from both black and white historical aerial photographs and high resolution color aerial photography. By using a multi-resolutional segmentation approach, and grouping homogeneous objects together at different scales, contextual and hierarchical information can be incorporated into the classification process. This procedure returns reliable woody cover classifications despite the complex heterogeneity of savanna systems. Combining elevation data from LiDAR with high resolution digital color imagery through object-based image analysis greatly enhances the structural description of a landscape by adding the three-dimensional height component. The corrected normalized canopy model provides a more realistic representation of vegetation height distribution than standard DSMDTM subtraction approaches. This is primarily due to the fact it does not assume continuous vegetation cover and accounts for the spatial heterogeneity of savanna woody cover.

490


Comparisons of remotely measured tree height with ground validated points indicate that structural attributes of woody vegetation can be reliably obtained from fusing LiDAR and imagery through object-based image analysis in a savanna system. This holds significant implications for vegetation management in savannas by providing a tool for monitoring vegetation structure remotely. Understanding the patterns of spatial and temporal heterogeneity of a system is fundamental to its successful management. If we consider the reciprocal relationship between pattern and process in ecological systems (Turner 1989), understanding where structural changes occur spatially in the landscape can help elucidate the drivers of vegetation change. The fusion of LiDAR and imagery in an object-based image analysis environment provides the means for generating this spatio-temporal understanding. The multi-scale, contextual approach inherent in object-based image analysis provides managers with a powerful tool for monitoring changes in vegetation structural diversity across heterogeneous landscapes at different scales.

6 References Blaschke T and Strobel J (2001) What's wrong with pixels? Some recent developments interfacing remote sensing and GIS. GIS - Zeitschrift fur Geoinformationssysteme 6:12-17 Braack L (1997) An objectives hierarchy for the management of the KNP. A revision of parts of the management plan for the Kruger National Park.Volume VII Dowling R and Accad A (2003)Vegetation classification of the riparian zone along the Brisbane River, Queensland Australia, using light detection and ranging (lidar) data and forward looking digital video. Can. J. Remote Sensing 29:556-563. Gillson L (2004a) Evidence of hierarchical patch dynamics in an East African savanna? Landscape Ecology 19:883-894. Gillson L( 2004b) Testing non-equilibrium theories in savannas: 1400 years of vegetation change in Tsavo National Park, Kenya. Ecological Complexity 1:281-298 Laliberte AS, Rango A, Harvstad KM, Paris JF, Beck RF, McNeely R, Gonzalez AL ( 2004) Object-oriented image analysis for mapping shrub encroachment from 1937 to 2003 in southern New Mexico. Remote Sensing of Environment 93:198-210 Lefsky MA, Cohen BW, Parker GG, Harding DJ (2002) Lidar remote sensing for ecosystem studies. Bioscience 52:19-30


491

Lovell JL, Jupp DLB, Culvenor DS and Coops NC (2003. Using airborne and ground-based ranging lidar to measure canopy structure in Australian forests. Canadian Journal of Remote Sensing 29:607-622. Maier B, Tiede D, Dorren L (2006) Assessing mountain forest structure using airborne laser scanning and landscape metrics. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences XXXVI-4/C42 Menges CH, Van Zyl JJ, Ahmad W and Hill GJE (1999) Classification of vegetation communities in the tropical savannas of northern Australia using AirSAR data. Pacific Rim AIRSAR Significant Results and Planning Workshop, Maui, Hawaii, 24-26 August Noss RF (1990) Indicators for monitoring biodiversity: A hierarchical approach. Conservation Biology 4:355-364 Pickett STA, Cadenasso ML, Benning TL (2003) Biotic and Abiotic Variability as Key Determinants of Savanna Heterogeneity at Multiple Spatiotemporal Scales. In: du Toit JT, Rogers KH, Biggs H (ed) The Kruger Experience: Ecology and Management of Savanna Heterogeneity. Island Press, Washington, DC,pp 23-40 Rogers KH (2003) Adopting a heterogeneity paradigm: Implications for management of protected savannas. In: du Toit JT, Rogers KH, Biggs H (ed) The Kruger Experience: Ecology and Management of Savanna Heterogeneity. Island Press, Washington, DC,pp 41-58 Sankaran M, Hanan NP, Scholes RJ, Ratnam J, Augustine DJ, Cade BS, Gignoux J, Higgins SI, Le Roux X, Ludwig F, Ardo J, Banyikwa F, Bronn A, Bucini G, Caylor KK, Coughenour MB, Diouf A, Ekaya W, Feral CJ, February EC, Frost PGH, Hiernaux P, Hrabar H, Metzger KL, Prins HHT, Ringrose S, Sea W, Tews J, Worden J, Zambatis N (2005) Determinants of woody cover in African savannas. Nature 438:846-849 Scholes RJ and Walker BH (1993) An African savanna: a synthesis of the Nylsvley study. Cambridge University Press Tiede D, Lang S, Hoffmann C (2006) Supervised and forest type-specific multiscale segmentation for a one-level-representation of single trees. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences Vol. No. XXXVI-4/C42 Turner MG (1989) Landscape Ecology: The effect of pattern on process. Ann. Rev. Ecol. Syst. 20:171-197 Waring RH, Way J, Hunt ER Jr, Morrissey L, Ranson KJ, Weishampel JF, Oren R and Franklin SE (1995) Imaging radar for ecosystem studies. BioScience 45: 715-723