Elsevier

Remote Sensing of Environment

Volume 125, October 2012, Pages 214-226
Remote Sensing of Environment

Mapping tree species composition in South African savannas using an integrated airborne spectral and LiDAR system

https://doi.org/10.1016/j.rse.2012.07.010Get rights and content

Abstract

Mapping savanna tree species is of broad interest for savanna ecology and rural resource inventory. We investigated the utility of (i) the Carnegie Airborne Observatory (CAO) hyperspectral data, and WorldView-2 and Quickbird multispectral spectral data and (ii) a combined spectral + tree height dataset (derived from the CAO LiDAR system) for mapping seven common savanna tree species or genera in the Sabi Sands Reserve and communal lands adjacent to Kruger National Park, South Africa. We convolved the 72 spectral bands of the CAO imagery to eight and four multispectral channels available in the WorldView-2 and Quickbird satellite sensors, respectively. A combination of the simulated WorldView-2 data and LiDAR tree height imagery was also assessed for species classification. First, the simulated WorldView-2 imagery provided a higher classification accuracy (77% ± 3.1 (mean ± standard deviation)) when compared to the simulated Quickbird (65% ± 1.9) and CAO (65% ± 1.2) data. Secondly, the combined spectral + height dataset provided a slightly higher overall classification accuracy (79% ± 1.8) when compared to the WorldView-2 spectral only dataset. The difference was however, statistically significant (p < 0.001; one-way analysis of variance for 30 bootstrapped replicates (n = 100) of the independent validation dataset). Higher classification accuracies were observed for trees with large crowns such as S. birrea, S. africana and A. nigrescens as compared to trees with small crowns. Species composition and diversity maps of trees with large crowns were consistent with established knowledge in the area. For example, the results showed higher tree diversity (number of different species per ha) in the Sabi Sands game reserve than in the communal areas. This study highlights the feasibility of remote sensing of tree species at the landscape scale in African savannas and the potential applicability of WorldView-2 sensor in mapping savanna tree species with a large crown.

Highlights

► Seven savanna tree species were mapped with an overall accuracy of 79% ► WorldView-2 is adequate for mapping savanna tree species with big crowns (  6 m) ► WorldView-2 provided higher accuracies than CAO hyperspectral and Quickbird data ► LiDAR tree height data could be useful for discriminating savanna tree species ► Tree maps revealed patterns in species diversity known for the Kruger Park region

Introduction

Ecologists have long been interested in explaining the distribution of species in ecosystems (He and Legendre, 2002, MacArthur, 1972). This question has gained increased relevance in recent years as rates of species loss have continued to rise. Turner et al. (2003) argue that scientifically sound environmental management of ecosystems requires frequent and spatially detailed assessments of species numbers and distributions. Direct field inventory methods are time consuming, expensive and often inadequate for large geographic areas (Cho, Skidmore and Sobhan, 2009b, Mairs, 1976). The limitations of field approaches at the landscape scale has driven research on the use of remote sensing as a fast and cost-effective means of mapping vegetation functional types and even species (Fuller et al., 1998, Mairs, 1976, Roberts et al., 1998).

Plant species mapping with remote sensing is linked to an understanding that species have unique spectral signatures associated with characteristic biochemical and biophysical properties (Asner and Martin, 2009, Cho et al., 2010, Clark et al., 2005). However, widespread mapping of species at the regional scale has been hampered by the low spectral resolution of most existing spaceborne sensors (e.g. Landsat, Systeme Probatoire d'Observation de la Terre (SPOT) and Quickbird) and the scarcity of appropriate high spectral resolution (hyperspectral) sensors (Huang & Asner, 2009). Hyperspectral imagery is acquired in hundreds of narrow, contiguous bands that can cover the visible, near-infrared, and shortwave-infrared portions of the electromagnetic spectrum (0.4–2.5 μm) (Curran, 1994). There is strong evidence that narrowband spectra are needed to resolve subtle spectral features of leaf and canopy biochemicals including photosynthetic pigments and foliar nutrients (Blackburn, 1998, Cho and Skidmore, 2006, Curran, 1989, Mutanga et al., 2004) and hence, the subtle spectral differences between canopies or crowns of different species in a landscape.

Several studies have demonstrated the ability of discriminating tree species with hyperspectral data (Andrew and Ustin, 2008, Asner, Knapp, et al., 2008, Cho et al., 2010, Dennison and Roberts, 2003, Hestir et al., 2008, van Aardt and Wynne, 2001). Both parametric and non-parametric classification techniques have been used. Parametric methods such as maximum likelihood and discriminant analysis consider both first order variations (e.g. mean values) and second order variations (e.g., covariance matrices) in the data, thus accounting for within-species variability in the classification (Clark et al., 2005). However, the high data dimensionality of hyperspectral data limits the application of parametric classifiers, that is, the number of training spectra per species must be equal to or greater than the number of spectral bands (Landgrebe, 1997). This requirement is often hard to meet, e.g. having at least 220 training spectra per species in the case of the Hyperion (EO-1) sensor with 220 spectral bands (Balzter et al., 2007). Hyperspectral data may also suffer from band redundancy for specific applications, i.e. neighbouring bands or even bands from different parts of the spectrum may be strongly correlated and therefore contain highly similar information. Some studies have employed data reduction techniques such as principal component analysis and wavelet energy feature vectors to mitigate the high dimensionality problem (e.g. Kalacska et al., 2007). On the other hand, non-parametric techniques including spectral similarity measures e.g. spectral angle mapper (Cho et al., 2010), sub-pixel classification techniques e.g. spectral mixture analysis (Dennison and Roberts, 2003, Somers et al., 2011), machine learning methods e.g. artificial neural networks and support vector machine (Barnard et al., 2010) and decision tree classification techniques e.g. Random forests (Naidoo et al., 2012) make no assumption of the data distribution. However, they are computationally intensive, particularly when applied on high spatial and spectral resolution data over broad areas.

New spaceborne multispectral sensors such as DigitalGlobe WorldView-2 (Ozdemir & Karnieli, 2011)) and RapidEye (Ramoelo et al., 2012) designed with strategically located wavebands in the absorption features of leaf biochemicals offer promise for widespread remote sensing of plant species. The WorldView-2 spectral bands are centred at 425 nm (absorbed by chlorophyll), 480 nm (absorbed by chlorophyll), 545 nm (sensitive to plant health), 605 nm (absorbed by carotenoids — detects ‘yellowness’ of vegetation), 660 nm (absorbed by chlorophyll), 725 nm (sensitive to vegetation health), 835 nm (sensitive to leaf mass and moisture content) and 950 nm (sensitive to leaf mass and moisture content) (see review by Ustin et al., 2009). It is therefore suggested that by preserving some of the inherent features of hyperspectral data in WorldView-2 multispectral data, such as the carotenoids and chlorophyll sensitive bands, some of the dimensionality problems of hyperspectral data with respect to species classification could be mitigated. It however, remains to be established, (i) whether the WorldView-2 provides additional information for species mapping when compared to conventional multispectral sensors such as DigitalGlobe Quickbird (bands: 430–545 nm, 466–620 nm, 590–710 nm, 715–918 nm (Longmont, CO, USA) and (ii) whether WorldView-2 could produce comparable results for species classification when compared to hyperspectral data, at least for the visible to near-infrared spectral range. Using Quickbird imagery, Ardila et al. (2011) mapped tree species in a city with an accuracy of 66%. The potential of WorldView-2 for mapping tree species (e.g. Latif et al. (2010)) for various ecosystems including savannas needs to be established.

Remote sensing of savanna tree species using spectral information from passive sensors is a difficult challenge given the inherent complexity of the savanna ecosystem structure. Savanna ecosystems are generally characterised by a continuous grass layer scattered with variable densities of shrubs and isolated trees (Belsky and Canham, 1994, House et al., 2003). Furthermore, trees of varying heights and crown dimensions co-exist (Wessels et al., 2011), and the grass background is an additional source of spectral confusion between trees, particularly for trees at various phenological stages. An important question is whether the combination of structural (tree height) and spectral information can be used to map savanna tree species, and whether the structural/spectral data combination provides an improvement over spectral data alone. Light Detection and Ranging (LiDAR) remote sensing provides information on vegetation structure, notably vegetation height (Lefsky et al., 2005, Levick et al., 2009, Wessels et al., 2011). LiDAR is an active sensing technology using a laser (light amplification by stimulated emission of radiation) to transmit a light pulse towards a target and a receiver to measure the backscattered or reflected light from that target (Lefsky et al., 2005). Distance to the object is determined by recording the time interval between the transmitted and backscattered pulses. The combination of spectral and LiDAR data was successfully used to improve the mapping of reedbed habitats in Cumbria, UK (Onojeghuo & Blackburn, 2011), invasive species in Hawaiian rainforest (Asner et al., 2008a) and for species (11 tree species) distribution mapping in the coastal Pacific Northwest, Canada (Jones et al., 2010).

The aims of this study were: (i) to compare the utility of WorldView-2 spectral data for discriminating savanna tree species to the potential of hyperspectral data and a conventional multispectral image such as Quickbird, (ii) to investigate the utility of an integrated airborne spectral and LiDAR system for mapping common savanna tree species in the Sabi Sands Reserve and neighbouring communal lands in South Africa (Fig. 1), and (iii) lastly, to determine whether the species diversity maps generated from the classified species maps corroborate with conventional knowledge on species diversity in the region. For example, we assumed that the maps produced would show that granite soils are richer in tree species than gabbro or whether Acacia nigrescens is more abundant on gabbro than on granite (Schmidt et al., 2002).

Section snippets

Materials and methods

The methodology used in the study is summarised in Fig. 2. The basic steps of the research protocol consist of image acquisition and pre-processing, field data to calibrate or train the remote sensing image and to validate the species classification results, and the mapping and assessment of species numbers and diversity within the study area.

Species classification results

The classification using the simulated WorldView-2 spectral only dataset provided a higher overall accuracy (76.8 ± 3.1 (mean ± standard deviation)) when compared to the blue, green, red and near-infrared bands of Quickbird (65.1% ± 1.9) and to all 72 VNIR bands of the CAO data (65% ± 1.2) (Table 3). It should be noted that the classification involving the 72 CAO bands did not include T. sericea because there were fewer pixels (n = 40) than the number of CAO bands. The producer's accuracy for the

Discussion and conclusions

Successful discrimination and mapping of tree species with remote sensing has been reported in several studies (Andrew and Ustin, 2008, Asner, Knapp, et al., 2008, Dennison and Roberts, 2003, Hestir et al., 2008, van Aardt and Wynne, 2001) but rarely for African savannas (e.g. Cho et al., 2010, Sarrazin et al., 2010). This study shows that remote sensing of African savanna tree species is feasible using WorldView-2 spectral data. High classification accuracies were achieved, particularly for

Acknowledgments

We thank the Council for Scientific and Industrial Research (CSIR), South Africa for proving the funding for the study. Hyperspectral and LiDAR imagery were supplied by the Carnegie Airborne Observatory, which is funded by the Andrew Mellon Foundation. The CAO system is further supported by the W.M. Keck Foundation, the Gordon and Betty Moore Foundation and William Hearst III. Data were pre-processed by T. Kennedy-Bowdoin, D. Knapp, J. Jacobson and R. Emerson at the Carnegie Institution for

References (70)

  • P.J. Curran

    Remote sensing of foliar chemistry

    Remote Sensing of Environment

    (1989)
  • P.E. Dennison et al.

    The effects of vegetation phenology on endmember selection and species mapping in southern California chaparral

    Remote Sensing of Environment

    (2003)
  • P.L. Emanuel et al.

    Modelling the sustainable harvest of Sclerocarya birrea subsp. caffra fruits in the South African lowveld

    Forest Ecology and Management

    (2005)
  • J.-B. Féret et al.

    Spectroscopic classification of tropical forest species using radiative transfer modeling

    Remote Sensing of Environment

    (2011)
  • R.M. Fuller et al.

    The integration of field survey and remote sensing for biodiversity assessment: A case study in the tropical forests and wetlands of Sango Bay, Uganda

    Biological Conservation

    (1998)
  • E.L. Hestir et al.

    Identification of invasive vegetation using hyperspectral remote sensing in the California Delta ecosystem

    Remote Sensing of Environment

    (2008)
  • T.G. Jones et al.

    Assessing the utility of airborne hyperspectral and LiDAR data for species distribution mapping in the coastal Pacific Northwest, Canada

    Remote Sensing of Environment

    (2010)
  • M. Kalacska et al.

    Hyperspectral discrimination of tropical dry forest lianas and trees: Comparative data reduction approaches at the leaf and canopy levels

    Remote Sensing of Environment

    (2007)
  • A. Ramoelo et al.

    Regional estimation of savanna grass nitrogen using the red-edge band of the spaceborne RapidEye sensor

    International Journal of Applied Earth Observation and Geoinformation

    (2012)
  • D.G. Leckie et al.

    Stand delineation and composition estimation using semi-automated individual tree crown analysis

    Remote Sensing of Environment

    (2003)
  • M.A. Lefsky et al.

    Geographic variability in LiDAR predictions of forest stand structure in the Pacific Northwest

    Remote Sensing of Environment

    (2005)
  • S.R. Levick et al.

    The relative influence of fire and herbivory on savanna three-dimensional vegetation structure

    Biological Conservation

    (2009)
  • R. Lucas et al.

    Classification of Australian forest communities using aerial photography, CASI and HyMap data

    Remote Sensing of Environment

    (2008)
  • J.W. Mairs

    The use of remote sensing techniques to identify potential natural areas in oregon

    Biological Conservation

    (1976)
  • A.D. Manning et al.

    Scattered trees are keystone structures — Implications for conservation

    Biological Conservation

    (2006)
  • O. Mutanga et al.

    Predicting in situ pasture quality in the Kruger National Park, South Africa, using continuum-removed absorption features

    Remote Sensing of Environment

    (2004)
  • L. Naidoo et al.

    Classification of savanna tree species, in the Greater Kruger National Park region, by integrating hyperspectral and LiDAR data in a Random Forest data mining environment

    ISPRS Journal of Photogrammetry and Remote Sensing

    (2012)
  • A.O. Onojeghuo et al.

    Optimising the use of hyperspectral and LiDAR data for mapping reedbed habitats

    Remote Sensing of Environment

    (2011)
  • I. Ozdemir et al.

    Predicting forest structural parameters using the image texture derived from WorldView-2 multispectral imagery in a dryland forest, Israel

    International Journal of Applied Earth Observation and Geoinformation

    (2011)
  • D.A. Roberts et al.

    Mapping chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models

    Remote Sensing of Environment

    (1998)
  • S. Saatchi et al.

    Impact of spatial variability of tropical forest structure on radar estimation of aboveground biomass

    Remote Sensing of Environment

    (2011)
  • K. Schumann et al.

    Impact of land-use type and harvesting on population structure of a non-timber forest product-providing tree in a semi-arid savanna, West Africa

    Biological Conservation

    (2011)
  • C.M. Shackleton

    Comparison of plant diversity in protected and communal lands in the Bushbuckridge lowveld savanna, South Africa

    Biological Conservation

    (2000)
  • B. Somers et al.

    Endmember variability in Spectral Mixture Analysis: A review

    Remote Sensing of Environment

    (2011)
  • W. Turner et al.

    Remote sensing for biodiversity science and conservation

    Trends in ecology & evolution

    (2003)
  • Cited by (184)

    • Using object-oriented coupled deep learning approach for typical object inspection of transmission channel

      2023, International Journal of Applied Earth Observation and Geoinformation
    • Evaluation of classification techniques in Very-High-Resolution (VHR) imagery: A case study of the identification of deadwood in the Chilean Central-Patagonian Forests

      2022, Ecological Informatics
      Citation Excerpt :

      The detection of changes in both uses and coverage of a particular area permits to make decisions on the management and conservation of ecosystems quickly and efficiently (Ávila-Pérez et al., 2020). In this sense, pattern recognition techniques have already been extensively used (Ávila-Pérez et al., 2020; Yin et al., 2018); in particular, satellite image classification techniques have stood out as an excellent tool for monitoring changes in land cover and, precisely, classifying tree species (Deur et al., 2021). Earlier classifiers (decades from 1970 to 1990, approximately) assigned pixels in the image to a certain number of classes based exclusively on the numerical value representing the radiometric intensity received by a remote sensor (Chuvieco, 2006; Toro and Lizarazo, 2012).

    View all citing articles on Scopus
    View full text