Object-based land cover classification using airborne LiDAR Remote ...

Remote Sensing of Environment 112 (2008) 2988–2998

Contents lists available at ScienceDirect

Remote Sensing of Environment j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / r s e

Object-based land cover classification using airborne LiDAR A.S. Antonarakis a,⁎, K.S. Richards a, J. Brasington b a b

Department of Geography, University of Cambridge, Downing Place, Cambridge CB2 3EN, UK Institute of Geography and Earth Sciences, The University of Wales, Aberystwyth, Llandinam Building, Penglais Campus, Aberystwyth, Ceredigion SY23 3DB, Wales, UK

a r t i c l e

i n f o

Article history: Received 31 October 2007 Received in revised form 22 January 2008 Accepted 16 February 2008 Keywords: Airborne LiDAR Image classification Riparian forest Vegetation roughness Point cloud Intensity Hydraulic modelling

a b s t r a c t Light Detection and Ranging (LiDAR) provides high resolution horizontal and vertical spatial point cloud data, and is increasingly being used in a number of applications and disciplines, which have concentrated on the exploit and manipulation of the data using mainly its three dimensional nature. LiDAR information potential is made even greater though, with its consideration of intensity. Elevation and intensity airborne LiDAR data are used in this study in order to classify forest and ground types quickly and efficiently without the need for manipulating multispectral image files, using a supervised objectorientated approach. LiDAR has the advantage of being able to create elevation surfaces that are in 3D, while also having information on LiDAR intensity values, thus it is a spatial and spectral segmentation tool. This classification method also uses point distribution frequency criteria to differentiate between land cover types. Classifications were performed using two methods, one that included the influence of the ground in heavily vegetated areas, and the other which eliminated the ground points before classification. The classification of three meanders of the Garonne and Allier rivers in France has demonstrated overall classification accuracies of 95% and 94% for the methods including and excluding the ground influence respectively. Five types of riparian forest were classified with accuracies between 66 and 98%. These forest types included planted and natural forest stands of different ages. Classifications of short vegetation and bare earth also produced high accuracies averaging above 90%. © 2008 Elsevier Inc. All rights reserved.

1. Introduction Better information on roughness of various types of vegetation is needed for use in resistance equations and eventually in flood modelling. These types include woody riparian species with different structural characteristics. Remote Sensing information such as 3D point cloud data from LiDAR can be used as a tool for extracting simple roughness information relevant for the condition of below canopy flow, as well as roughness relevant for more complex tree morphology that affects the flow when it enters the canopy levels. One strategy for extracting roughness parameters from remote sensing techniques is to use a data fusion object classification model. This means that multiple datasets such as LiDAR, digital aerial photography, ground data and satellite data can be combined to produce roughness parameters estimated for different vegetative patches, which can subsequently be mapped spatially using a classification methodology. Airborne LiDAR is used in this study in order to classify forest and ground types quickly and efficiently without the need for manipulating multispectral image files. Classifications have until recently been attempted with multispectral imagery (Duda et al., 1999; Sun et al.,

⁎ Corresponding author. E-mail addresses: [email protected] (A.S. Antonarakis), [email protected] (K.S. Richards), [email protected] (J. Brasington). 0034-4257/$ – see front matter © 2008 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2008.02.004

2003). LiDAR has the advantage of being able to create elevation surfaces that are in 3D, while also having information on LiDAR intensity values, thus it is a spatial and spectral segmentation tool. LiDAR intensity information has not been greatly used either in the commercial sector or in academia, yet it could be an important factor for feature extraction or land cover classification (Flood, 2001). LiDAR has traditionally been used spatially to separate ground points from aboveground points (Cobby et al., 2001; Vosselman, 2000; Lohmann et al., 2000), or classify LiDAR into features such as buildings (Axelsson, 1999) and vegetation (Mason et al., 2003; Cobby et al., 2003). Until very recently, not much had been achieved by using LiDAR point cloud data with elevations and intensity for land cover classifications. Classifications have been attempted by Brennan and Webster (2006) who used derived LiDAR surfaces to differentiate between different layers. Their study used four layers (Mean Intensity, Normalised Height, Digital Surface Model, Multiple Waveform LiDAR Returns) to differentiate between ten land features. These classifications included water, vegetation, roads, saturated and unsaturated soils, coniferous and deciduous trees, and building structures. Other LiDAR based classifications have also been attempted, for example by Charaniya et al. (2004) using LiDAR point cloud elevation and intensity data to classify roofs, grass, trees and roads. Bartels and Wei (2006) performed LiDAR based maximum likelihood classifications fused with co-registered spectral bands, extracting land types including building, vegetation, and ground (i.e. all features at low elevation ranges) from a small 800 m2 urban area.

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

For this study, it was desired to classify known land cover types on river reaches of the Rivers Garonne and Allier, with special consideration given to three meanders. The land cover types have been defined in the field, and relevant parameters have been extracted from them to be able to estimate their roughness at different scales of complexity. Thus these desired land types exist from ground truthing data, aerial photography and from previous research on the study sites including Muller et al. (2002) and De Jong (2005). These desired land types are gravel bars, bare earth, short vegetation, three different ages of planted forests, and two ages of natural riparian forest. Including the river water, this makes nine classes. Each land type was classified using exclusive criteria, in order to classify only known land types. These criteria were chosen based on different rasterised surface interpolations of LiDAR point cloud data. The characteristics of the land types thus need to be known, in order to develop the appropriate LiDAR derived surface models. The objectives of this study were to: - Explore the possibility of identifying and classifying land types in two floodplains, using airborne LiDAR intensity and elevation data. - Accurately identify and classify the different forest types and forest ages.

2989

4854000N), and the second near the village of Monbequi (UTM31; 356000E 4861500N). Both the Verdun and Monbequi sites consist of a large proportion of commercial planted poplar clones of all ages, which are heavily pruned. The ground in the summer consists of ploughed earth and dry, prone grass distributed sparsely around the site, as well as gravel near the edges of the low flow river edge. Vegetation on these meanders also includes natural black poplar (Populus nigra), which can be very dense and are situated on the immediate bank of the river. A secondary natural species was Salix alba, but its distribution was limited (Muller et al., 2002). For the Garonne River, three planted poplar ages as well as mature dense natural riparian forests were considered and measured. The youngest planted poplar was 1–3 years old, with intermediate aged planted poplar at 3–8 years, and mature planted poplar from 8–12 years. One meander section was examined on the Allier near the village of Châtel-de-Neuvre (UTM31; 525250E 5140350N). In this meander most of the surface was bare and consisting of bar forms with variously sized gravel but also included sparsely vegetated areas. The main species was again P. nigra with a limited number of S. alba. For this river, the younger natural riparian forests were between one to five years, and mature natural riparian forests were older than five years. 3. Land types

2. Data description The airborne LiDAR data and digital aerial photographs used in this study were obtained from flights organised by the Natural Environment Research Council Airborne Research & Survey Facility. The flights collected information on the reaches of the two rivers on the 6th of June 2006 for the Garonne, and the 8th of June 2006 for the Allier. A SPOT image of the Garonne River from Toulouse to Montauban was also collected on the 6th of June 2006, with four spectral bands including near infrared (NIR) with a resolution of 10 m. The airborne LiDAR had an average flying height of 1300 m collecting first and last pulse data with an average point density of 1.9 m, and an average spatial resolution of 1.0–2.5 m. The data were provided by the Cambridge University based Unit for Landscape Modelling, where the point clouds had an x–y position accuracy of less than 1 m and an elevation accuracy of less than 15 cm, with the highest point accuracies of 0.020 m, 0.009 m, 0.052 m in x, y, z, resulting from a high fixed point GPS accuracy of 0.0002 m. Apart from xyz coordinates, first and last pulse information on intensity were also collected. LiDAR intensity can be defined as the ratio of the strength of the light reflected from an object related to the light emitted (Song et al., 2002). Song et al. (2002) further stated that different objects can have a different reflectance, and can be related to the on-site light conditions and the spectral band being used by the LiDAR emitter. Three meanders are considered in this study, the first two being from the heavily managed Garonne River, and the second from the almost unmanaged Allier River. The first two meanders considered were near the village of Verdun-sur-Garonne (UTM31; 359500E

There are eight main land cover types that were previously defined with a total of nine layers including water surfaces. Water surfaces are usually recognised as areas that absorb much of the incoming radiation. Of course this depends on the velocity, turbulence and depth of the water as these can cause the surface to be brighter, thus absorbing less radiation (i.e. from ‘white water’ or emerging gravel bars in rivers). Therefore water in remote sensing has traditionally been considered as having the lowest intensity pixels (Harris, 1987; Nedeljkovic, 2006). Song et al. (2002) defined the reflectivity of airborne LiDAR pulses in relation to various surfaces. In the near infrared, clear water was defined as having the lowest reflectivity (0–10% reflectance), pebbles having a very low reflectance similar to water (17%), grass and short vegetation having the highest (50% reflectance), and high reflectances for sand (41–57% for wet to dry soil). Using airborne LiDAR, all these surfaces can be differentiated from trees by simply considering a low height range. Thus water can be defined as having low intensity values as well as having a low range in height, or a small deviation from an average terrain model. Three more low height range surfaces are described in this study, and are related to those stated here in the research done by Song et al. (2002). Short vegetation is described as any land type that is dominated by tall grass and non-woody agricultural crops. Due to the chlorophyll content of the grass and short vegetation, it should have a high intensity value when reflecting pulses in the near infrared. Bare earth is un-vegetated land that is on the whole disturbed agricultural fallow land as well as soils. Gravel bars are described as the riparian sediments that are continuously prone to

Fig. 1. Distribution of LiDAR raw points in natural and planted forests at Verdun.

2990

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

erosion and migration, and are located near the edges of the water surface for flows that do not exceed bankfull discharge. As suggested by Song et al. (2002), gravel has a similar intensity signal to water. This may be explained by the high moisture content present on the gravel related to other un-vegetated surfaces such as soils. Sometimes, though, the intensity values from the un-vegetated surfaces could be confounded, therefore a further criterion for gravel bars has to be used. This is the difference in intensity values from the first and last pulse. Here, the difference in the intensity signal should not be as large as that of bare earth. Gravel is deemed a more homogeneous surface than bare earth as gravel consists of weathered rocks with average diameters estimated to 10–20 cm, while bare earth is rougher ploughed and disturbed earth with regular occurrences of pits and soil mounds, with differences in moisture. Specific ranges in intensity values for these four land cover types are presented in the results section. Indeed water and gravel had the lowest NIR values in this dataset averaging below 70–75; bare earth and soils had intensity values averaged at around 70 to 90; and short vegetation and grasses had values of around 90–150. A more complex surface to classify is that of riparian forest types. Natural and planted forests cannot be classified just from aggregated elevation values or from intensity values, especially when dealing with the same species in different spatial formations. The vertical distribution of points in these different forest types could be a key in classifying them. Fig. 1 shows the distribution of points in both natural and planted forests from a section of the meander at Verdun. The natural forests have a more uniform distribution of points with height than the planted forests. Also in the natural forested section, the concentrations of points have a near uniform distribution with elevation, while there seem to be more concentrated point distributions with height for the planted forest. Therefore, the skewness and kurtosis

of the LiDAR elevation points could provide a way of differentiating between these two layers, as they illustrate deviations from a normal distribution of points. The skewness of a probability frequency distribution is a measure of its asymmetry compared to a normal distribution. In other words, if the mass of the distribution is concentrated away from the centre of tendency, then it is considered skewed. Skewness is characterised by the ratio between the third moment about the mean (κ3) divided by the second moment about the mean (κ2), and is defined by the equation: Skewness ¼

j3 3=2

j2

Pn P 3 i¼1 ðxi  x Þ ¼  : Pn P 2 3=2 i¼1 ðxi  x Þ

ð1Þ

Each moment about the mean is defined by the sum of a power of the deviation of the individual point (xi) from the mean ( P x ). The kurtosis in a probability frequency distribution is the measure of its peakedness. Mathematically it is described by the ratio between the fourth moment about the mean (κ4) and the second moment about the mean (κ2), and is defined as: Kurtosis ¼

Pn P 4 j4 i¼1 ðxi  x Þ  3 ¼ P   3: 2 P 2 2 j2 n i¼1 ðxi  x Þ

ð2Þ

Four frequency histogram plots of the forest types are shown below in Fig. 2, three being planted polar of different ages, and the first describing the distribution of points in a natural forested section. These elevations are from the original LiDAR point cloud data with an average spatial resolution of 1.5–2.5 m. From Fig. 2, it can be seen that the frequency distributions of elevations for the planted and natural forests are very different. The

Fig. 2. Frequency distribution of points in planted poplars of different ages (mature [graph B], intermediate-aged [graph C], young [graph D]), and natural poplars of all ages (graph A).

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

2991

Fig. 3. Frequency distribution of points in planted poplars of different ages (mature [graph B], intermediate-aged [graph C], young [graph D]), and natural poplars of all ages (graph A) without the influence of ground hits.

intermediate and young poplar plantations have a largely positive skew, due to the very irregular distribution of points with a high proportion of them belonging to the ground. The distribution of points for the natural and mature planted forest shows less of a deviation in terms of its skewness, and a large number of points exist as hits in the planted poplar's canopy. The kurtosis of these, though, is quite different. Due to the strong influence of the canopy compared to the ground hits in the mature planted forests, the kurtosis will be largely platykurtic, while the distribution for the natural forest will have a kurtosis closer to zero. A combination of kurtosis and skewness could aid in differentiating between planted and natural forests. Yet the distributions illustrated in Fig. 2 are bimodal if the ground and undergrowth is included. The problem is that skewness and kurtosis are based on calculations of the moments about the mean that are designed for unimodal distributions. The deviations from the mean are measured in relation to a mean, which in the case of Fig. 2 are the gaps between the two modes. Fig. 3 describes the same areas and vegetation types as Fig. 2 without the influence of the ground or nearground hits. Here, the distributions seem to be more unimodal except for the younger planted forests. The more mature planted forests seem to be more negatively skewed, while the natural forests and the younger planted forests are more positively skewed. Due to the varied point distribution with height of the younger planted forests, a further factor could be used to differentiate between this forest type and natural forest. The percentage of canopy hits could be used to identify the porosity of a spacious younger planted forest and a dense natural forest. This measure may be somewhat discriminatory in itself, as canopy hits could consider points belonging to ground flora, and vice versa. The percentage of canopy hits could also be useful in differentiating between planted and natural forests as a whole if the cell size

defining each distribution is large enough to account for habitual gaps in the planted forests. 4. Surface derivations From the information provided above, six LiDAR derived surfaces were used in order to classify the desired land types for the three meanders described in Section 2. These surfaces were a vegetation height model, a percentage canopy model, an average intensity model, an intensity difference model and probability distribution skewness and kurtosis models. The development of these layers was aided by using C++ code. This code inputs raw point cloud data, and outputs the desired information on points and their position such as maxima and average heights, as well as moments about the mean and skewness and kurtosis values, within a defined cell resolution. This code has the advantage of keeping the original positioning of the output points, depending on the elevation or intensity criteria assigned. The resolution of cells in the code was chosen to be 5 m in order to aggregate areas of similar characteristics. 4.1. Vegetation height model The VHM is derived from the difference between a terrain model (TM; Fig. 4B) and the original point cloud canopy surface model (CSM; Fig. 4A). The terrain model was determined as the point in each 5 m cell with the lowest elevation. The canopy surface model was determined as the point in each 5 m cell with the highest elevation. Cells of 5 m were chosen, as this was the minimum width of tree canopies measured in the field, and thus limits the effects of considering cells with forested and non-forested land. For both of these models, the first and last pulse raw airborne LiDAR information was used. First pulse returns were used in conjunction with last pulse returns because some last return values of a

2992

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

Fig. 4. Airborne LiDAR surfaces for the Verdun meander on the Garonne: A) canopy surface model (CSM); B) terrain model (TM); C) vegetation height model (VHM); D) intensity model (IM); E) intensity difference model (IDM); F) skewness model (SkM); G) kurtosis model (KrM); H) percentage canopy model (PCM).

single pulse returned a higher value than the first pulse return, due to noise in the LiDAR receiver. The terrain model and surface model points were then used to construct triangulated irregular networks (TINs) based on the elevations and this was then interpolated to a rasterised image with a 5 m resolution. The difference of these two rasterised datasets formed the vegetation height model (Fig. 4C). It should be noted though that the problem of accurately defining the ground and canopy tops using LiDAR point cloud data is unresolved, with suggestions of moving into calibrating multiple signal airborne LiDAR (Harding et al., 2001; Lefsky et al., 2002). Nevertheless, Antonarakis (in press) found that LiDAR derived tree heights from crown apices using these surfaces were around 81% of the field measured values. 4.2. Percentage canopy hits model The Percentage Canopy hits Model (PCM) was defined as the percentage of LiDAR hits that were reflected from the canopy compared to the total LiDAR hits in a 10 m cell resolution. This resolution is different than the other LiDAR derived surfaces and was chosen for two reasons. First, this resolution was chosen in order to be able to identify more than just one point when considering spacious young planted forests. Second, a 10 m resolution would account for gaps in the planted forests, causing sharper contrasts between mature planted and mature natural forests. The ground and the ground flora were considered as being in the lower third of the forest's elevation range. This was chosen from the limiting factor being the youngest planted poplar raw LiDAR point frequency distribution with elevation (Fig. 2D). A TIN was subsequently created relating each resulting percentage per 10 m cell, and a raster image was finally interpolation. 4.3. Intensity models The average intensity model (IM; Fig. 4D) was constructed from the LiDAR point intensity values using an amalgamation of both raw first

and last pulse points. The average intensity was used only for land cover types with a low elevation range. Keeping that in mind, the total difference in intensity values for short vegetation and bare earth was only 15–20, with each cell having an average standard deviation of only around 4. Thus it was considered that averaging intensities did not affect the eventual classification of a surface, and did not significantly confuse land cover types. The same method of surface development was deployed for the mean intensity layer as the terrain and surface models. This was by constructing a TIN based on the average intensity returns, and subsequently interpolating the triangular network into a rasterised image with a 5 m resolution. The intensity difference model (IDM; Fig. 3E) was constructed using the differences of the maxima values obtained from the first pulses and the minima values from the last pulse intensities. Again the maximum first pulse points and minimum last pulse points were interpolated separately into rasterised images, and subsequently the difference between the two was extracted. 4.4. Skewness and kurtosis models The skewness model (SkM; Fig. 4F) was constructed using Eq. (1), and the kurtosis model (KrM; Fig. 4G) was calculated using Eq. (2) in this study. These layers were calculated for each 5 m cell of the C++ code using the combination of first and last pulse point clouds, considering the elevation values to construct the moments about the mean. Raw LiDAR data points in the specified cell were grouped, and each individual point elevation was subtracted from the mean elevation value of the cell. The exponent was applied according to the moment degree required, and the resulting values were summed. TIN and raster interpolations were subsequently performed at the desired resolution to create the two surfaces. The skewness and kurtosis models were created for both ground-on and ground-off conditions, and used in two different methodologies described in the subsequent section. All surfaces for the meander at the Verdun site

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

2993

Fig. 5. Flow chart of method 1 classification logic.

on the Garonne River are shown below (Fig. 4). From these surfaces, the classification criteria can now be described. 5. Classification methods The classifications of each land type were performed using three software packages, C++ programming, ArcGIS, and MATLAB. The first two are used to develop the surfaces and the class separation is performed through MATLAB. Two classification methods were performed, and differ in the consideration of the forested classifications. The first classification method describes criteria to differentiate between forest types that do include the ground hits in their point frequency distribution as shown in Fig. 2. This first method focuses on the use of bimodal distribution skewness and kurtosis models to differentiate between the natural and planted poplars. Fig. 5 is a flow chart describing the classification hierarchy for the first method. The second classification method describes a technique to differentiate forest types without the inclusion of ground hits as shown in the frequency distributions in Fig. 3. This method considers the differentiation between natural and planted poplars using unimodal distribution skewness and kurtosis models, and includes a further layer of the percentage of canopy hits compared to ground hits. Fig. 6 is a flow chart describing the forest classification hierarchy for the second method. Descriptions of definition criteria for water, gravel, bare earth, and short vegetation are the same for both methods. Both of these methods are used and compared. The use of the second method is better justified because the skewness and kurtosis models created have unimodal frequency point distributions with height. The first method is worth considering as well, as the resulting forest types could

have very high classification accuracies insinuating the strength of including the ground influence. Also, the comparison of both these methods should demonstrate the need for omitting ground hits in the present or any future airborne LiDAR classification techniques. All classifications are defined as belonging to one or multiple criteria thresholds. The spatial location and extent of each class are known before the classification algorithm is applied. Accordingly, multiple pixels of a known class were examined for their value ranges from a certain surface model type, and threshold values were subsequently chosen. For example skewness and kurtosis threshold values for planted poplar forests were defined as a representative range of values from a large sample of their pixels values. 5.1. Water The first classification was of open water surfaces, in this case rivers. River water was classified as having a height range of less than 0.5 m (VHM b 0.5), and average intensity values of less than 55 (IM b 55). Potentially in some cases, the river may not have returned any points, due to water absorption and no backscatter. In these cases, the SPOT image obtained on the 6th June 2006 could be used in the NIR to determined water bodies as having very dark pixel values. 5.2. Land feature elevations Land features such as agricultural land or forests are first considered from their elevation deviations from the local minimum (as defined by the terrain model). To classify low features and tall features, two elevations need to be determined and separated. Hence, near-ground

Fig. 6. Method 2 classification logic excluding the influence of the ground. The skewness and kurtosis models are now defined from their canopy as SkM(c) and KrM(c).

2994

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

features were determined as having a deviation from the terrain model of less than 3.5 m (VHMb 3.5), while elevated features (namely tall vegetation) is considered as having a deviation from the terrain model of more than 3.5 m (VHMN 3.5). Low features are differentiated from the water layer by the difference in intensity values attributed (intensity values described in Section 5.3). This elevation was chosen to signify the witnessed cut-off point between the tallest nettles and shortest trees. 5.3. Short vegetation and bare earth Short vegetation such as nettles, grasses, and agricultural crops is categorised as those pixels that have a deviation from the terrain of less than 3.5 m, and an average intensity of greater than 90 (IMN 90). Bare earth is subsequently characterised as pixels with the same deviation from the terrain model, but with an average intensity of between 75 and 90 (75b IMb 90). These classification criteria are supervised and were selected from knowledge of the land types and where they belonged spatially. 5.4. Gravel Gravel is considered as having a very low height range (VHM b 0.5), with a witnessed intensity between 55 and 75 (55 b IM b 75). Gravel also should produce a difference in the maxima and minima intensity values that is less than bare earth and short vegetation, and this could sharpen the classification of this land type. For this a difference of less than 30 was chosen (IDM b 30). Also gravel was considered as being close to the river, so any gravel-defined pixels that were not directly connected to the low flow river's edge were omitted. 5.5. Forests (inclusion of ground hits) Both skewness and kurtosis surface models were used to define the natural and planted poplar riparian forests. From Fig. 2 and from supervisions of the two layers, planted poplars were defined as those with skewness values of more than 1.3 m (SkMN 1.3) and kurtosis values of less than −1.7 (KrM b −1.7). Here, the skewness criterion helps define pixels that potentially belong to younger planted forests, and the kurtosis criterion helps define the more mature planted poplar forest. The values were chosen to be as extreme as possible in order to best differentiate

between planted and natural forests. Natural forests were defined from skewness and kurtosis values that were less negative and less platykurtic respectively. The frequency distribution of points is more normal than those of planted forest sections, so values were chosen for skewness of between 0.5 and −0.5 (−0.5 b SkMb 0.5), and for kurtosis of between 0 and −0.5 (−0.5 b KrMb 0). Some post-processing needed to be performed on these resulting classifications. Plantations and natural forested sections were unified through a few steps. First, in some instances, small areas of a few pixels in size were classified even if they did not belong to a forest. Thus areas that were defined by being smaller than 3–4 pixels were deleted. A filtering technique was subsequently applied to even out the edges of defined areas. This was achieved by first eroding the binary image using a 3×3 kernel, and subsequently dilating it using the same kernel size. The eroding process using the same kernel size acted to delete jagged or thin peninsular edges, while not eroding pixels that belonged to a consolidated area. Using non-consistent kernel sizes would ultimately change the fabric and resolution of the image. Finally, gaps in a classified forest were filled in to further consolidate the main areas of the forested sections. The gaps were defined as zero values in the binary image that were enclosed by forest values defined with values of 1. 5.6. Forests (omission of ground hits) The second classification method (Fig. 6) was performed with skewness and kurtosis layers that did not include the influence of the ground or ground flora. To create a more accurate classification, the percentage of canopy hits model (PCM) was also included. As above, the values chosen for each of the two forested land types were as extreme as possible for the most accurate differentiation between them. Planted forests were defined with large negative skewness (SkM b −0.8) and large leptokurtic values (KrM N 2). Natural forests were defined as having skewness and kurtosis values being near zero (−0.15 b SkM b 0.15; −0.1 b KrM b 0.1). The percentage of canopy points were used to identify the planted forests that would have a small proportion of the points being intercepted by their canopies (0 b PCM b 0.2), and identify the denser natural forests as areas where most of the LiDAR points are intercepted by the canopy (0.4 b PCM b 0.9). Values between 0.2 b PCM b 0.4 were not included as this would create spatial overlapping of planted and natural poplars,

Fig. 7. Classification of the Verdun meander.

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

Fig. 8. Classification of the Monbequi meander.

Fig. 9. Classification of the Chatel meander.

2995

2996

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

as this range includes both dense mature planted poplar and natural poplar. Again, plantations and natural forested sections were unified as noted in Section 5.5.

Table 1 Method 1 accuracy indices (AI) of forested areas in the two Garonne sites Type

Verdun classification

5.7. Forest ages Three ages of planted forests and two ages of natural forests were defined. Natural Forests were thus segregated into Young Natural Forest (VHM b 12) and Mature Natural Forests (VHM N 12). Planted forests were further divided into Young Planted Forests (VHM b 10), Intermediate Planted Forests (10 b VHM b 18), and Mature Planted Forests (VHM N 18). These heights were chosen from previous field measurements of individual tree characteristics in June 2006 and February 2007. More classes of forest age could be defined if desired using further segmentations of the canopy heights. It is important to note that trees with the same height in planted and natural poplar may not necessarily be of the same age due to the antecedent conditions in each vegetation stand.

Water Gravel Short vegetation Bare earth Young planted poplar Inter planted poplar Mature planted poplar Young natural poplar Mature natural poplar Total

Monbequi classification

Chatel classification

Class pixels (km2)

AI (%)

Class pixels (km2)

AI (%)

Class pixels (km2)

AI (%)

0.23 0.05 0.88 1.15 0.23

99.19% 70.04% 99.96% 99.12% 84.23%

0.18 0.05 0.31 0.73 0.18

98.72% 84.08% 99.72% 99.07% 97.67%

0.14 0.17 1.13 0.26 0.004

95.20% 80.14% 99.89% 84.08% –

0.23

95.33%

0.21

98.22%

0.001

–

0.36

94.57%

0.28

97.38%

0.0001

–

0.19

82.93%

0.15

97.87%

0.38

94.60%

0.18

80.99%

0.16

83.84%

0.18

81.46%

3.50

95.44%

2.25

97.21%

2.25

93.76%

6. Results and classification accuracy Classifications were produced for the three meander areas on the Garonne and Allier Rivers. The resulting classified images are presented in Figs. 7, 8 and 9. At first glance, it can be noticed that very few pixels were classified as planted forests in the Chatel meander site, while the majority of forested land cover in the two Garonne meanders show planted forest dominance. Prior knowledge of the three sites is enough to confirm that the Allier River has no or very few planted forests in the immediate riparian zone, while the Garonne river is characterised by at least 81% planted forests in the whole floodplain with natural forests occupying the below bankfull discharge river edges (Muller et al., 2002). In fact for both of the Garonne meanders, around 70% of the classified forest cover was found to be occupied by poplar plantations, just in these small reaches. Estimates on the Allier River from Clermont-Ferrand to Moulins have indicated that live forest cover can account for around one third of the total floodplain land type, while pioneer vegetation and undergrowth can account for 32% of the total area (Peters et al., 2000). The forest cover estimated in the two classification methods amounted to around 33% and 35% of the total meander areas investigated, while the shorter vegetation accounted for 29% and 28% of the total areas. Even if these meanders do not take into account the full floodplain in the sections investigated, the percentages can give a good indication of land cover ratios. These accuracies that have been stated here are all generalised spatially for the three meander sites, resulting in broad accuracies of the forest cover types discussed. Both methods developed in this study show good to very high levels of accuracy when concerning both the non-vegetation and vegetation land cover types, with the accuracy values presented below. There are some principle differences that can be distinguished when first comparing the classified images for the two methods. The largest differences are on the Verdun meander, especially in the northeast corner. The first method classifies a patch of young poplar forest, while the second classifies it as young natural poplar. There are also some other sections inside the Verdun meander in the southeast of the classified image that are classified as mature planted poplar and natural forest for the first and second method respectively. Hence, when removing the influence of the ground for these areas, the point cloud distributions of the area become more platykurtic, and more evenly skewed, resulting in pixels in these regions to be classified as natural poplar forest rather than planted poplar forest. There are also some regions in the Monbequi meander that exhibit the same pattern. Further speculations on this land cover type divergences are discussed in the concluding section.

The accuracy of the classification results was assessed primarily by using aerial digital photographs acquired on the same date as the airborne LiDAR data (6th June 2006). The three sites were visited as well, so memory of the land types was also useful in defining the accuracy of each meander. The classified image was visited in relation to the digital aerial photographs, and each area that seemed to be falsely classified was noted with the number of pixels. This was performed for the entirety of each of the three meanders. Omissions and commissions of all pixels were investigated for all land types for the three meanders and for both classification methods. The accuracy index (AI) was used to take commission and omission pixels into account for each land type. This accuracy method is defined by Pouliot et al. (2002). Overall classification accuracy can be defined as: AIðkÞ ¼

  ðn  ðO þ C ÞÞ  100: n

ð3Þ

AI is an accuracy index in percent, O and C represent the number of omission and commission errors, and n is the total number of trees in the image to be detected. The purpose of the index is to count all error against the correct number of trees to be detected. Commission errors are where pixels are falsely assigned to another class, and omission errors are where pixels were not assigned to their correct land type. The classification accuracy results for the first method are show below in Table 1. The accuracy results in Table 1 show that the first classification method had overall accuracies of around 95%, with the best overall classification accuracy being for the Monbequi site with 97.2%, and the lowest for the Chatel site with 93.8% accuracy. The total accuracy for all three sites combined was 95.5%. Just considering the total woody vegetation, this method accurately classified 91.5% of all trees. The tree category that was least accurately classified was mature natural riparian forest with accuracies for the three sites ranging from around 81–84%, while mature and intermediate planted poplar were accurately classified with percentages between 95 and 98%. The young planted poplar at the Verdun site seemed to be under-classified compared to the other planted poplar forest ages (84.2% accuracy). Comparing the planted and natural poplar forests, the natural forest had 88% classification accuracy, and the planted forests had 94% classification accuracy. The weakest land cover classification was for the gravel bars with values as low as 70%. The weaker classification of the natural forest compared to the planted forest for this first method may be attributed to the unclassified pixels. Around half of the pixels omitted from the mature and young natural riparian forest were left unclassified. A small number of unclassified pixels also belonged to the gravel bar surface. Only small areas were classified as gravel for the Garonne meanders, while the

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

7. Conclusions

Table 2 Method 2 accuracy indices (AI) of forested areas in the two Garonne sites Type

Water Gravel Short vegetation Bare earth Young planted poplar Inter planted poplar MaturePlantedPoplar Young natural poplar Mature natural poplar Total

2997

Verdun classification

Monbequi classification

Chatel classification

Class pixels (km2)

AI (%)

Class pixels (km2)

AI (%)

Class pixels (km2)

AI (%)

0.23 0.05 0.87 1.15 0.19 0.21 0.33 0.29 0.23 3.55

99.08% 74.03% 99.96% 99.12% 74.56% 89.92% 81.32% 79.83% 66.37% 91.77%

0.18 0.04 0.31 0.73 0.19 0.21 0.28 0.16 0.17 2.27

98.72% 80.52% 99.72% 99.07% 97.11% 94.41% 91.21% 93.00% 84.99% 95.69%

0.14 0.16 1.13 0.26 0.01 0.00 0.01 0.40 0.18 2.27

95.20% 80.73% 99.89% 84.08% – – – 98.33% 95.36% 95.17%

Chatel site on the other hand contained large areas of gravel bars spread through the meander. In the Chatel site, many pixels classified as bare earth actually were of gravel bars. The second method had fewer overall unclassified pixels, with more in the Monbequi section. Classification accuracy results for the second method are shown below in Table 2. This second classification method had overall accuracies of around 94% with the highest average accuracy index being for the Monbequi site with 95.7%, and the lowest for the Verdun site with 91.8% accuracy. The total woody vegetation accurately classified from this second method was lower that the first, with an average accuracy index of 86.8%. Most to all of the individual forest categories had lower classification accuracies associated with this second method with the least accurately classified forest type being again the natural riparian forest with accuracies for the three sites from 66–95%. In this method, the planted and natural riparian forests had the same classification accuracy of 86.8%. This is because there were less unclassified pixels than in the first method, and much of the error was caused through overlap of natural forest into planted forest, or vice versa. The gravel land cover type increased in classification accuracy for the second method mainly through the correct classification of pixels as forest. Using intensity and elevation information from airborne LiDAR data, Charaniya et al. (2004) were able to classify four land types with successful classification accuracy ranging from 66–84%. The four classes derived from LiDAR data were trees, grass, roads and roofs, and the terrain model used was not derived using LiDAR points, but obtained from the United States Geological Survey (USGS). Using multiple LiDAR derived surfaces, Brennan and Webster (2006) classified ten land types with 94% overall accuracy. This classification method was based on identifying bright and dark vegetation, coniferous and deciduous forests, defining building structures, and also the saturation of soils. The LiDAR derived surfaces that were used were based on elevation values and intensity, but the study was also able to obtain data of multiple LiDAR returns (greater than 2 returns). The results from this study can also be compared to recent classifications performed using multispectral imagery only. Lu et al. (2004) used Thematic Mapper imagery to classify different stages of rainforest succession and agricultural land with an overall accuracy range of 70–86%. Buddenbaum et al. (2005) used hyperspectral remote sensing data (HyMap) to classify forests, and reported an overall accuracy of 74%. Stow et al. (2007) used Landsat TM/ETM+ data to classify forests and shrubland with overall accuracies of 64% and forest classification accuracies of around 80%. Vohland et al. (2007) used Landsat TM to classify 8 land cover types including forests, and reported an average accuracy of 87.5%. Bork and Su (2007) defined 8 vegetation types, from grasses to shrubs and forest. They defined classification accuracies using 3-band multispectral data of 59.4%. Combined with LiDAR elevation range values, this classification went up to 80.3%.

This classification method first demonstrates that airborne LiDAR is an effective tool in classifying land types, and has been successful in accurately classifying 95% and 94% of the land types in three sites in the Garonne and Allier floodplains for the first and second methods respectively. These methods also demonstrate the possibility of using frequency distribution parameters such as skewness and kurtosis to accurately identify planted and natural poplar riparian forests in twodimensional interpolated LiDAR surfaces. Natural poplar forests were classified with 88% accuracy, and planted poplar forests were classified with 94% accuracy in the first method, and all forests were classified with 86.8% accuracy using the second method. The combination of intensity and elevation data from the LiDAR point clouds can be enough to classify multiple land types. Both intensity and elevation data played a prominent role in defining regions, and the combination of both could aid in the better classification of a land type. Both methods demonstrated their effectiveness in classifying forests with high accuracies, and the effect of removing the ground influence did not hinder this classification process significantly. The second method though, justifiably used skewness and kurtosis with unimodal canopy point frequency distributions, while the first method used kurtosis and skewness of forests dominated by ground hits for sparser forests and canopy hits for denser forests. Therefore, it may be expected that the second method would produce higher classification accuracies. A reason for this not being the case may be that removing the ground decreases the polarity of the kurtosis and skewness ranges in the forests, resulting in more overlapping of one forest type or the other. A second reason may be that for younger poplars, the distributions of points were not truly unimodal as the mature forest distributions were, resulting in skewness and kurtosis range errors. Although the overall accuracy of both methods was very high, there were some issues with the method, as discussed below. One source of error could almost certainly result from the TIN interpolation of raw LiDAR point cloud data. In certain areas on the floodplain, especially on the river surface, points were absent either through pulses not returning, or through the fault of the flight path coverage. This could have caused unnaturally high elevations for the river surface. Secondly, some of the areas have been aggregated in order to follow the desired choices stated in Section 3, arising from the need to represent land types with a previously calculated roughness associated with it. For example, the short vegetation land type most likely is an aggregation of distinctive surfaces such as pioneer vegetation in a riparian zone, grasses and short shrubs. Roads and buildings were also not considered in this classification method, although they have been included in other studies. On the three sites investigated, the land area occupied by buildings was relatively small, but they could potentially be definable. Different species of woody vegetation were not identified in this method either. One area defined as planted forest in the northeast of the Verdun site for example is occupied by orchards rather than planted poplar. Potentially different tree species could be identified from their different point cloud elevation frequency distributions, but this could be difficult to implement on the floodplain scale. It was considered that the effects of calibration errors of the airborne LiDAR point clouds on the final classifications were minimal. First this is because the spatial (0.001–4%) and elevation (0.005–6%) errors of the cell size (5× 5 m) and height range (3.5–30 m) considered were too small to affect the creation of the surfaces used in classification. The calibration errors would only directly affect the choice of height range values to represent the different forest ages, yet these chosen values were not centimetric and broadly represented a stage in the maturity of a tree. Finally this information could be used for many applications including monitoring of vegetation change through time. In the broader sense of this research, forest classification information could be very useful when considering 2D hydraulic modelling. In some computational fluid dynamics (CFD) packages such as grid based flow

2998

A.S. Antonarakis et al. / Remote Sensing of Environment 112 (2008) 2988–2998

model HYDRO2DE, 2D spatial GIS type information can be used as an input with a lumped roughness value associated with each land type. This would ultimately involve LiDAR point cloud data eventually being used in flood simulations. Acknowledgments This paper is the result of airborne LiDAR data obtained from a NERC project in June 2006. This research was supported by the British Society for Geomorphology (BSG) and from the William Vaughan Lewis and Phillip Lake Funds. References Antonarakis, A. S., Richards, K. S., Brasington, J., Muller, E., Bithell, M. (in press).The retrieval of vegetative fluid resistance terms for rigid stems using Airborne LiDAR. Journal of Geophysical Research-Biogeosciences. doi:10.1029/2007JG000543 Axelsson, P. (1999). Processing of laser scanner data — algorithms and applications. ISPRS Journal of Photogrammetry and Remote Sensing, 54, 138−147. Bartels, M., & Wei, H. (2006). Rule-based improvement of maximum likelihood classified LIDAR data fused with co-registered bands. Annual Conference of the Remote Sensing and Photogrammetry Society, CD Proceedings, 05–08 September 2006 (pp. 1−9). Bork, E. W., & Su, J. G. (2007). Integrating LIDAR data and multispectral imagery for enhanced classification of rangeland vegetation: a meta analysis. Remote Sensing of Environment, 111(1), 11−24. Brennan, R., & Webster, T. L. (2006). Object-oriented land cover classification of lidarderived surfaces. Canadian Journal of Remote Sensing, 32(2), 62−172. Buddenbaum, H., Schlerf, M., & Hill, J. (2005). Classification of coniferous tree species and age classes using hyperspectral data and geostatistical methods. International Journal of Remote Sensing, 26(24), 5453−5465. Charaniya, A. P., Manduchi, R., & Lodha, S. K. (2004). Supervised parametric classification of aerial LiDAR data. CVPRW'04, Proceedings of the IEEE 2004 Conference on Computer Vision and Pattern Recognition Workshop, 27 June – 2 July 2004, Baltimore, Md, Vol. 3 (pp. 1−8). Cobby, D. M., Mason, D. C., & Davenport, I. J. (2001). Image processing of airborne scanning laser altimetry for improved river flood modelling. ISPRS Journal of Photogrammetry and Remote Sensing, 56(2), 121−138. Cobby, D. M., Mason, D. C., Horritt, M. S., & Bates, P. D. (2003). Two-dimensional hydraulic flood modelling using a finite element mesh decomposed according to vegetation and topographic features derived from airborne scanning laser altimetry. Hydrological Processes, 17, 1979−2000. De Jong, J. (2005). Modelling the influence of vegetation on the morphodynamics of the river Allier. MSc. thesis, Delft University of Technology, Delft.

Duda, T., Canty, M., & Klaus, D. (1999). Unsupervised land-use classification of multispectral satellite images. A comparison of conventional and fuzzy-logic based clustering algorithms. IEEE Geoscience and Remote Sensing Symposium IGARSS, Vol. 2 (pp. 1256−1258). Flood, M. (2001). LIDAR activities and research priorities in the commercial sector. International Archives of Photogrammetry and Remote Sensing, 34(3), 3−7. Harding, D. J., Lefsky, M. A., Parker, G. G., & Blair, J. B. (2001). Laser altimetry canopy height profiles: methods and validation for closed canopy, broadleaf forests. Remote Sensing of Environment, 76, 283−297. Harris, R. (1987). Satellite remote sensing; an introduction (pp. 220). London: Routledge and Kegan Paul. Lefsky, M. A., Cohen, W. B., Parker, G. G., & Harding, D. J. (2002). Lidar remote sensing for ecosystem studies. Bioscience, 52(1), 19−30. Lohmann, P., Koch, A., & Schaeffer, M. (2000). Approaches to the filtering of laser scanner data. International Archives on Photogrammetry and Remote Sensing, 33, 540−547. Lu, D. S., Batistella, M., & Moran, E. (2004). Multitemporal spectral mixture analysis for Amazonian land-cover change detection. Canadian Journal of Remote Sensing, 30(1), 87−100. Mason, D. C., Cobby, D. M., Horritt, M. S., & Bates, P. D. (2003). Floodplain friction parameterization in two-dimensional river flood models using vegetation heights derived from airborne scanning laser altimetry. Hydrological Processes, 17, 1711−1732. Muller, E., Guilloy-Froget, H., Barsoum, N., & Brocheton, L. (2002). Populus nigra L. en vallée de Garonne: legs du passé et constraints du présents. Comptes Rendus Biologies, 325, 1129−1141. Nedeljkovic, I. (2006). Image classification based on fuzzy logic. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Commission VI, Vol. 34 (pp. 1−6) Part XXX. Peters, B., van Looy, K., & Kurstjens, G. (2000). Pioniervegetaties langs grindrivieren: de Allier en de Grensmaas. Natuurhistorisch Maandblad, 7(89), 123−136. Pouliot, D. A., King, D. J., Bell, F. W., & Pitt, D. G. (2002). Automated tree crown detection and delineation in high-resolution digital camera imagery of coniferous forest regeneration. Remote Sensing of Environment, 82, 322−334. Song, J. H., Han, S. H., Yu, K. Y., & Kim, Y. I. (2002). Assessing the possibility of land-cover classification using lidar intensity data. Photogrammetric Computer Vision Symposium 2002, ISPRS Commission III, 9 – 13 September 2002, Graz, Austria, Vol. B (pp. 259-263). Stow, D., Lopez, A., Lippitt, C., Hinton, S., & Weeks, J. (2007). Object-based classification of residential land use within Accra, Ghana based on QuickBird satellite data. International Journal of Remote Sensing, 28(22), 5167−5173. Sun, W., Heidt, V., Gong, P., & Xu, G. (2003). Information fusion for rural land-use classification with high-resolution satellite imagery. IEEE Transactions on Geoscience and Remote Sensing, 41(4), 883−890. Vohland, M., Stoffels, J., Hau, C., & Schuler, G. (2007). Remote sensing techniques for forest parameter assessment: multispectral classification and linear spectral mixture analysis. Silva Fennica, 41(3), 441−456. Vosselman, G. (2000). Slope based filtering of laser altimetry data. International Archives of Photogrammetry and Remote Sensing, 33(B3/2), 935−942.