Integration of LiDAR and QuickBird imagery for mapping riparian biophysical parameters and land cover types in Australian tropical savannas

Forest Ecology and Management 259 (2010) 598–606

Contents lists available at ScienceDirect

Forest Ecology and Management journal homepage: www.elsevier.com/locate/foreco

Integration of LiDAR and QuickBird imagery for mapping riparian biophysical parameters and land cover types in Australian tropical savannas Lara A. Arroyo a,b,*, Kasper Johansen a,b, John Armston a,c, Stuart Phinn a,b a

Joint Remote Sensing Research Program, Australia The University of Queensland, Centre for Remote Sensing and Spatial Information Science, School of Geography, Planning and Environmental Management, Brisbane, QLD 4072, Australia c Remote Sensing Centre, Queensland Department of Environment and Resource Management, Climate Building, 80 Meiers Road, Indooroopilly, QLD 4068, Australia b

A R T I C L E I N F O

A B S T R A C T

Article history: Received 26 June 2009 Received in revised form 11 November 2009 Accepted 12 November 2009

Riparian zones are exposed to increasing pressures because of disturbance from agricultural and urban expansion and overgrazing. Accurate and cost-effective mapping of riparian environments is important for baseline inventories and monitoring and managing their functions associated with water quality, biodiversity, and wildlife habitats. In this study, we integrate remotely sensed light detection and ranging (LiDAR) data and high spatial resolution satellite imagery (QuickBird-2) to estimate riparian biophysical parameters and land cover types in the Fitzroy catchment in Queensland, Australia. An object based image analysis (OBIA) was adopted for the study. A digital terrain model (DTM), a tree canopy model (TCM) and a plant projective cover (PPC) map were first derived from the LiDAR data. A map of the streambed was then produced using the DTM information. Finally, all the LiDAR-derived biophysical map products and the QuickBird image bands were combined in an OBIA to (1) map the following land cover types: riparian vegetation, streambed, bare ground, woodlands and rangelands; (2) determine the distribution of overhang vegetation within the streambed; and (3) measure the width of both the riparian zone and the streambed. The combined use of both datasets allowed accurate land cover mapping, with an overall accuracy of 85.6%. The estimated widths of the riparian zone and the streambed showed strong correlation with the actual field measurements (r = 0.82 and 0.98 respectively). Our results show that the combined use of LiDAR and high spatial resolution imagery can potentially be used for the assessment of the riparian condition in a tropical savanna woodland riparian environment. This work also shows the capacity of OBIA to assist in the assessment of the composition of the riparian environment from multiple image datasets. ß 2009 Elsevier B.V. All rights reserved.

Keywords: Riparian zone Riparian zone and streambed widths Riparian land cover Object based image analysis LiDAR QuickBird Tropical savanna Australia

1. Introduction Riparian zones are defined as the interface of terrestrial and aquatic ecosystems. They constitute a rich ecosystem both in terms of biomass and biodiversity and serve a wide variety of productive, protective and aesthetic functions (Apan et al., 2002). They hold a large number of plant and animal species, act as refuge in times of environmental stress, reduce and control flood events, and play a significant role in erosion processes, river water quality and stream bank stability (Fry et al., 1994; Jones et al., 2008; Langer et al., 2008; Lovett and Price, 2002; Simon and Collison, 2002). Moreover,

* Corresponding author at: The University of Queensland, Centre for Remote Sensing and Spatial Information Science, School of Geography, Planning and Environmental Management, Brisbane, QLD 4072, Australia. Tel.: +61 73365 4355; fax: +61 73365 6899. E-mail address: [email protected] (L.A. Arroyo). 0378-1127/$ – see front matter ß 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.foreco.2009.11.018

they have special cultural and recreational significance, particularly when they are in proximity to urban areas (Innis et al., 2000). These ecosystems are highly vulnerable to disturbances such as changes in water regimes, weed invasion, fire, overgrazing and erosion (Dixon et al., 2005; Innis et al., 2000). Riparian ecosystems have declined in area worldwide since the early 1900s, primarily as a result of construction of dams for flood control and water storage, and pumping of surface and ground water from floodplains for agriculture, human consumption and livestock grazing (Jones et al., 2008; Lytle and Merritt, 2004). Suitable management plans to control and restore riparian areas have become essential in order to preserve these important ecosystems. Hence, accurate and up to date knowledge on the extent, composition, structure and condition of riparian zones is critical. A systematic mapping and monitoring approach is necessary to measure the changes in the composition, structure and condition of riparian habitats and the effectiveness of riparian management and restoration programs (Jones et al., 2008). Several approaches have been proposed and used worldwide for mapping and monitoring the biophysical properties of riparian

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

zones. They usually consist of a set of measures and indices of physical properties and biotic attributes, also called riparian health indicators, that are assessed to derive a total score indicative of the riparian condition (Bain et al., 2000). In the case of Australian tropical savannas, Dixon et al. (2005) identified 21 riparian health indicators developed for field assessment, including canopy continuity, canopy cover, riparian zone width and streambed width. Assessment of riparian condition has traditionally been performed by a combination of field measurements, aerial photo-interpretation, and ecological modelling. However, field based assessments are labour-intensive, time consuming and expensive, and thus only suitable for the assessment of medium to small catchments (<200 km of stream length) (Johansen et al., 2007b). Considering that the drainage basins of the world’s rivers range in scale from less than 1 km2 to the Amazon catchment at about 7 million km2 (Mertes, 2002) and Australia’s Murray Darling Catchment of 1.06 million km2, the need for methods to map and monitor changes to riparian ecosystems across large geographic areas becomes clear. Remote sensing techniques may represent a viable alternative for monitoring riparian condition due to their synoptic and repetitive nature and their ability to provide information from areas which are not easily accessible (Johansen et al., 2007a). They have been successfully applied to evaluate several parameters and functions of the riparian zones in temperate, sub-tropical and tropical savanna environments, including plant projective cover (PPC), leaf area index (LAI), canopy continuity, vegetation communities, riparian zone width and bank stability (Congalton et al., 2002; Gilmore et al., 2008; Johansen and Phinn, 2006; Johansen et al., 2007a,b). However, remote sensing techniques have faced some limitations when applied for riparian ecosystem assessment. The information available from the early remote sensors was typically of a relatively coarse spatial resolution (Congalton et al., 2002). Discrepancies between remote sensing classes and field information were found due to the mismatch between the scales of the riparian features and the spatial resolution of the data (Mertes, 2002). Later studies performed with newer sensors confirmed the need of high spatial resolution imagery to identify the small elements of the riparian systems (Akasheh et al., 2008; Gilmore et al., 2008; Goetz et al., 2003; Johansen et al., 2007b), but these datasets were difficult to analyze due to their high degree of spatial detail and detail within individual features (e.g. tree canopy), which could not be processed using standard image processing techniques. The high level of reflectance variability within individual features made their identification difficult, and traditional pixel-based analyses were hampered by the increased complexity of the imagery (Goetz et al., 2003; Johansen et al., 2007a). As an alternative, OBIA approach segments the image into homogeneous groups of pixels (i.e., objects) prior to the classification of such objects. Image segmentation emerged as early as the 1970s and has been extensively used for a range of applications. Within the field of remote sensing, OBIA has shown a rapid development in the last 10 years, mainly as a consequence of the increased spatial resolution of the newer remote sensors (Blaschke, 2009). This approach has been successfully applied with high spatial resolution imagery for a variety of purposes (Arroyo et al., 2006; Cleve et al., 2008; Conchedda et al., 2008; Johansen et al., 2007a; Lamonaca et al., 2008; Mallinis et al., 2008), encouraging its adoption for the analysis of riparian land cover types. Another constraint for evaluating the riparian condition by remotely sensed data is the complex vertical and multi-layered structure of riparian areas and the limitation of passive sensors to an integrated measurement of this complex three-dimensional structure, prohibiting extraction of sub-canopy elements (Akasheh

599

et al., 2008; Johansen et al., 2007b). Optical sensors provide an integrated measurement of structural information on forest height and vertical distribution of foliage. They cannot detect features underneath areas of dense canopy cover and they do not directly provide vertical information on the vegetation attributes. Consequently, they have difficulties distinguishing between canopy and ground covers (e.g. grasses vs trees). Moreover, some measurements need to be based on empirical relationships (Goetz et al., 2003; Johansen et al., 2007b). Airborne altimetric light detection and ranging (LiDAR) introduces the possibility of discrete threedimensional analysis of vegetation and terrain features. This active sensor emits laser pulses over a swath of terrain and records the intensity and round trip travel times of the returning pulses, which are then converted into range measurements. Thus, by accurately knowing the aircraft location, the sensor provides information on both the terrain elevation and the vertical distribution of vegetation elements above the ground level with a vertical and horizontal accuracy of a few centimetres. The potential of airborne laser scanning to retrieve structural attributes of forest communities has been widely illustrated (for example, Clark et al., 2004; Lefsky et al., 2001; Persson et al., 2002; Popescu and Zhao, 2008; Suarez et al., 2005; Zimble et al., 2003). In this study, we incorporate LiDAR data to the analysis in order to evaluate its capacity to provide the structural information that was not captured from the optical data. Its capacity to assist the classification of riparian land cover classes is also evaluated. The OBIA adopted for the analysis of the high spatial resolution multispectral data also serves as a tool to incorporate LiDAR data into the assessment. The main objective of this study was to integrate LiDAR and high spatial resolution QuickBird data to estimate riparian biophysical parameters and land cover types for a study area located in an Australian tropical savanna environment. We intend to overcome the limitations of each sensor by using them in a complementary manner. Both data sets were incorporated into an OBIA in order to map: PPC, riparian vegetation, woodlands, rangelands, bare ground, streambed, streambed width, riparian zone width and overhanging vegetation. The capacity of the OBIA to (1) assist the classification of high spatial resolution data, and (2) integrate LiDAR and QuickBird imagery, was also evaluated. 2. Data and methodology 2.1. Study area The study area covered a stretch of 5 km along Mimosa Creek and associated riparian vegetation situated upstream of the junction with the Dawson River (248310 S; 1498460 E), located within the Fitzroy catchment in Queensland, Australia (Fig. 1). This is the largest river basin that drains into the east coast of Australia (142,000 km2) and flows into the southern end of the Great Barrier Reef. The catchment experiences a tropical to sub-tropical humid to semi arid climate. Annual median rainfall throughout the region is highly variable, ranging from about 500 mm to more than 800 mm annually along the coast where coastal ranges trap moisture from shore airflow. Most rain falls in the summer, with many winters experiencing no rain at all. The riparian vegetation is mainly surrounded by rangelands used for cattle and some agriculture; some remnant patches of woodland vegetation also occur. The major land use in the Fitzroy basin is grazing. 2.2. Field survey Field sampling was conducted between 28 May and 5 June 2007. Field measurements of vegetation cover and structural properties for image calibration and validation were derived along

600

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

(Johansen and Phinn, 2006). The width of the streambed was measured from the toe of the bank to the toe of the opposite bank. 2.3. Image data acquisition and processing

Fig. 1. Location of the riparian zone study area in the Fitzroy catchment, central Queensland, Australia.

25 m wide and 70–100 m long transects located perpendicular to the stream at each field site. Each site had six transect lines, each separated by 5 m starting at the edge of the streambed, going through the riparian zone and finishing 10–20 m beyond the external perimeter of the riparian zone. The geographic location at the start and end points of transects lines 1 and 6 (first and last ones) were registered with a Global Positioning System (GPS) receiver averaging the position for >1000 s or until the estimated positional error was less than 2.0 m. The orientation of the transect lines and identification of features visible both in the field and in the imagery were also recorded for more precise location of the field sites. Quantitative field measurements of PPC were derived along each of the six transect lines from upward looking photos taken at 5 m intervals. The photos were subsequently classified into canopy elements and sky (Fig. 2) to calculate the PPC within the field of view using the approach suggested by Van Gardingen et al. (1999). Ground cover and vegetation structure (number of trees and woody vegetation, DBH, height of understory elements) were measured for 5 m  5 m quadrants within each transect. The transect lines were located 5 m apart and served as a grid reference for these measurements. Riparian zone width and streambed width were measured to the nearest metre using a measuring tape. The riparian zone width was defined as the perpendicular distance from the toe of the stream bank to the external perimeter of the riparian zone, where an abrupt change in vegetation height, structure and density occurred and the bank slope decreased

A QuickBird image was captured of the study area on 11 August 2007 with an off-nadir angle of 14.68. The image was first radiometrically corrected to at sensor spectral radiance using the pre-launch calibration coefficients provided by DigitalGlobe Inc. The FLAASH module in ENVI 4.3 was then used to atmospherically correct the image to at-surface spectral reflectance. A total of 18 ground control points derived in the field were used to geometrically correct the image (root mean square error (RMSE) = 0.59 pixels for the multi-spectral bands). LiDAR data, acquired by the Leica ALS50-II LiDAR sensor on 15 July 2007 with a mean platform altitude of approximately 850 m above the ground level and an average point spacing of 1 m, were provided in American Society for Photogrammetry and Remote Sensing (ASPRS) Lidar Exchange Format (LAS), specification 1.1. LiDAR returns were classified as ground or non-ground by the data provider using proprietary software. According to the data provider, vertical accuracy of the LiDAR elevations was 0.15 m. Four products were derived from this dataset according to the methods described below: digital terrain model (DTM), slope map, tree canopy model (TCM) and PPC. A 0.5 m DTM was produced from the LiDAR data by inverse distance weighted interpolation of returns classified as ground with an exponent of two. Elevation of the ground at the position of non-ground returns was also estimated using the same interpolation technique. A slope map was produced using ArcMap software as the maximum rate of change between each cell of the DTM and its neighbours. The height of all first returns above the ground was calculated by subtracting the ground elevation from the first return elevation. These estimates of first returns were then aggregated into 2.4 m  2.4 m data bins to match the QuickBird multi-spectral spatial resolution and employed for the derivation of the TCM and the PPC. The TCM is a representation of the top of the canopy (Suarez et al., 2005) and it was calculated as the maximum height of first returns in each bin. PPC was estimated from the LiDAR cover fraction, defined as one minus the gap fraction probability, Pgap, at a zenith of zero. This was calculated from the proportion of counts in each data bin by 1  Pgap ðzÞ ¼

C V ðzÞ ; C V ð0Þ þ C G

(1)

where CV(z) is the number of first return counts above z metres, CV(0) is the number of first returns above the ground and CG is the number of first return counts from the ground (Lovell et al., 2003). z was set to 2 m to match the field measurements. All pulses were

Fig. 2. Example of the quantitative field measurements of PPC from upward looking photos. Photos (left) were classified (right) into canopy elements (black) and sky (white). PPC corresponds to the relative area of canopy elements in the classified photo (0.49 in this example).

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

assumed to be vertical as no account is made for scan angle. The fraction of LiDAR pulses intercepted by the canopy above a height of z is determined by the PPC, but calibration is required to account for the sensitivity of LiDAR cover fraction to the sensor configuration (Goodwin et al., 2006). A regional calibration of LiDAR cover fraction to PPC for woody vegetation communities, including tropical savannas, in Queensland was developed using independent LiDAR survey data from an Optech ALTM3025 with a mean platform altitude of approximately 1000 m and an average point spacing of 0.5 m. Since these data had a similar footprint diameter (0.2 m) and the calculation of PPC only used first returns, it was decided to test the regional calibration. It is important to note that other differences between the sensor and survey configurations were not accounted for in the estimation of PPC. A total of 47 field measurements of PPC were acquired coincident with these LiDAR data. These LiDAR and field surveys were used to develop a calibration curve from LiDAR fractional cover to PPC and are described in detail by Armston et al. (2009). Using the same procedure as Armston et al. (2009), a simple power function was found to fit the scatter well (RMSE 3.33% PPC) and had the property of being bounded from 0 to 100%: 0:6447 PPC ¼ 1  Pgap :

(2)

Since there was excellent agreement between the field estimates of PPC and LiDAR-derived fractional cover and the residuals were consistent with a binomial sampling distribution, the LiDAR cover fraction estimates were calibrated to estimates of PPC using Eq. (2). All the acquired information (field, QuickBird and LiDAR datasets) was collected within three months. The total precipitation recorded in that period of time for the area was 16.9 cm, which was considered to produce no changes between the field data collection and the image captures.

chosen for the second OBIA, due to the very different information content of the different data sets. An initial segmentation was carried out on the basis of the LiDAR-derived information using the PPC and TCM products. Those objects that showed low and similar TCM values (areas with no or low vegetation) were merged into larger segments. A second segmentation was then performed including the optical information. The location of the streambed was also incorporated into this segmentation, to make sure the outline of the classified streambed was preserved in the segmentation. After the segmentation, both multi-spectral and LiDAR-derived information were used to define the following land cover types: riparian vegetation, woodlands, rangelands, bare ground and streambed. Objects were classified on the basis of spectral values, spectral variability, size, shape and in relation to neighbouring objects. Four types of features were used for the classification: mean, standard deviation, context information and the normalised difference vegetation index (NDVI). Mean refers to the mean value of all pixels within an object (e.g. mean RED is the mean spectral value of the red band of all pixels within an object). The standard deviation features were also derived from the analysis of all pixels included in each object. These features were employed as an estimation of the level of variability within each object. For instance, rangeland areas, which characteristically showed smooth surfaces because of its homogenous grass cover, displayed low values of standard deviation in the near infrared (NIR) band. Context features refer to those features expressed in relation to one specific class, or set of classes (e.g., the relative border to objects classified as ‘‘class/set of classes’’). For example, the context feature ‘‘distance to riparian vegetation’’ was used to discard isolated forested areas misclassified as riparian vegetation. Finally, the NDVI values were calculated for each object by NDVI ¼

2.4. Land cover type classification and distribution of overhang vegetation within the streambed Land cover types were mapped using the software Definiens Developer 7 (Definiens AG, 2007), which was specifically designed for OBIA. The OBIA can be divided into two main processing steps. The first one is the segmentation of the data into homogenous segments (image objects); and the second is the assignment of these objects to discrete classes. Segmentation is controlled by the digital values of the input bands and the shape of the segmented objects. Two classifications were carried out using OBIA approaches. In the first, the DTM and the slope map were first employed for the location of the streambed within the study area. In this case, only LiDAR-derived information (DTM and slope map) was needed. Areas with steep slope were then identified as the edges of potential streambeds, which were flat and less elevated than the surrounding areas in the case of the streambed. All the available information (four QuickBird multi-spectral bands, DTM, PPC, TCM and the streambed map) was then incorporated into a new OBIA for further analysis. A stepwise segmentation approach was

NIR  RED ; NIR þ RED

(3)

where NIR is the mean of the NIR band values of all pixels within each object, and RED is the mean of the red band values of all pixels within each object. Each class was described by one or more of these features. The classification was performed in a hierarchical manner, with objects of one level informing the classification of other-level objects. Table 1 shows an overview of the features used for each class. The context feature ‘‘existence of streambed’’ was used for assessing the distribution of overhang vegetation within the streambed. All objects assigned to the class ‘‘streambed’’ were analyzed and divided into ‘‘clear streambed’’ or ‘‘streambed with overhang vegetation’’. 2.5. Riparian zone and streambed widths estimation The riparian zone width was estimated as the Euclidean distance from the toe of the streambed to the external perimeter of the riparian vegetation in the land cover type classification. All objects classified as riparian vegetation were first subdivided into smaller objects consisting of one pixel, and only those ones

Table 1 Object- and class-related features used for the object based classification. Class

Features used

Bare ground Riparian vegetation

Mean RED; mean TCM NDVI; number of neighbour ‘‘riparian vegetation’’ objects; enclosed by class ‘‘riparian vegetation’’; distance to ‘‘streambed’’ NDVI; standard deviation NIR; mean TCM NDVI; relative border to ‘‘riparian vegetation’’ Mean RED; presence of ‘‘streambed’’ NDVI; presence of ‘‘streambed’’

Rangelands Woodlands Streambed without vegetation Vegetation overhang

601

602

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

Fig. 3. LiDAR-derived products: (a) DTM; (b) streambed map (in blue); (c) PPC and (d) TCM. Bright areas correspond to high values for the terrain elevation (a and b), PPC (c) and tree heights (d). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of the article.)

corresponding to the external edge of the riparian vegetation were considered for the analysis of the riparian zone width. The riparian zone width was then extracted from the context feature ‘‘Distance to class’’, which measures the distance from the centre of each object to the closest object of the specified class. The distance from the pixels located at the edge of the riparian vegetation to the closest object classified as streambed was extracted. The same approach was employed for the estimation of the streambed width.

width was measured from both edges of the streambed (right and left hand side of the river) to the external perimeter of the riparian zone. Visual assessment from the optical information was unreliable for the validation of the streambed width estimations, since the streambed was frequently hidden underneath the canopy cover of the riparian vegetation. Hence, only field measurements were used for this validation.

2.6. Validation

3. Results

All the obtained products were validated using field measurements and visual interpretation of the QuickBird imagery. This includes the PPC estimations; the land cover type classification; and the riparian zone and streambed width measures. Validation of the LiDAR-derived PPC estimations was important for two main reasons: firstly, PPC represents itself a useful biophysical parameter in order to assess the riparian condition and it has been described as a riparian health indicator (Dixon et al., 2005). Secondly, the PPC map was an input for the OBIA, and therefore its accuracy affects the accuracy of the land cover type classification. PPC estimations were validated using field measurements of PPC derived from upward looking photos. To allow this validation, each field site was subdivided into smaller plots of 225 m2. This plot size (15 m  15 m) is approximately equivalent to the area covered by nine of the upward looking photos taken in the field (3  3 photos) and it represents a feasible compromise to allow geographic correspondence between both data sets. The average of the PPC estimates for each plot was then compared to the average of the corresponding nine field PPC measurements. A total number of 48 plots were used. An error matrix was constructed to estimate the accuracy of the land cover type classification. A stratified random selection of ninety validation sites, including fifteen sites per land cover type, was used for the accuracy assessment. Each validation site had an area of 23 m2 (four QuickBird pixels). After visual classification using both the multi-spectral and the panchromatic bands from the QuickBird image, the user’s, producer’s and overall accuracies of the classification were calculated, as well as the Kappa statistic. Measurements of the riparian zone width were visually derived from the QuickBird imagery and compared to those automatically obtained from the OBIA. The information collected in the field was also used to assist the visual assessment of the riparian zone width. The QuickBird image provided an adequate perspective for visual identification of the abrupt change in vegetation associated with the external edge of the riparian zone. Thus, identification of the riparian zone from the QuickBird imagery was in some areas easier and more precise than its identification in the field. A set of 34 measurements of riparian zone width was produced using the multi-spectral and panchromatic QuickBird bands. They corresponded to 17 sites located along the river, where the riparian zone

The 0.5 m DTM extracted from the LiDAR data revealed a fairly flat area, with a total height difference of only 25 m (Fig. 3a). This information was employed for mapping the streambed of the river according to its geomorphology (Fig. 3b). The PPC product showed the percentage of land covered by green foliage and non– photosynthetic vegetation (branches, trunks, dead leaves) (Fig. 3c). Finally, a TCM estimating the heights of the top of the canopy was derived from the LiDAR data (Fig. 3d). The canopy height ranged from 0 to 41.35 m. The PPC product was validated against the measurements obtained from the upward looking pictures, averaged each 255 m2. A comparison of the LiDAR-derived PPC with the field estimates of PPC showed a high level of agreement between both measurements (Fig. 4; r = 0.81; RMSE = 0.168), however some bias is evident. LiDAR-derived estimates of PPC are higher than the ones derived in the field. The scatter exhibits a bias similar to that shown in Armston et al. (2009). This suggests the value for the exponent in Eq. (2) was too high for the riparian environment and sensor configuration parameters in this study. Recalibration of Eq. (2) would be ideal but was not conducted in this study for two reasons: (i) there were insufficient direct measurements of PPC (i.e. point-intercept transects; Armston et al., 2009) for calibration and

Fig. 4. Scatter plot of the PPC estimations from the LiDAR data vs the PPC measurements extracted from upward looking photos.

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

603

Fig. 5. Segmentation levels: (a) LiDAR-derived segmentation and (b) incorporation of optical information.

validation; and (ii) the OBIA did not require unbiased estimates of PPC. Image segmentation was first carried out using only the LiDARderived information (Fig. 5a). This information on its own was useful for the tree identification, but it was insufficient for segmenting cover types with similar heights, such as bare ground and grasslands. The second segmentation level was created by incorporating the multi-spectral information (Fig. 5b), and the result was a more suitable separation of tree crowns and also other features. The multi-spectral information allowed identification of features with a similar structural behaviour. Finally, the incorporation of the boundaries of the streambed assisted the land cover classification and allowed the estimation of the riparian zone width. Classification was performed on the final segmentation level using the parameters defined in Table 1. The combined use of LiDAR, spectral and context information allowed accurate identification of the five land cover types, as well as the location of areas with and without overhang vegetation within the streambed (Fig. 6). Seventy-seven out of the ninety validation sites were correctly identified, which provided an overall classification accuracy of 85.6% (Table 2). The Kappa statistic for the land cover classification was 83.7%. Riparian vegetation and woodland classes were predicted with the lowest accuracy (62 and 68% respectively) due to the high level of spectral and positional similarity between them, especially in the transitional area between riparian and woodland vegetation. A total area of 4.1 ha of streambed (83.5% of the total streambed mapped for this study area) were located underneath vegetation overhang and would have been impossible to map by means of optical sensors alone. At the same time, the original streambed map, derived in the first OBIA from the LiDAR data, was missing 4.5% of the final streambed area. The rest of the streambed was identified and mapped when the QuickBird multi-spectral bands were incorporated in the second OBIA. Measurements of the riparian zone width and the streambed width were derived from the land cover classification map (Fig. 7). Reducing the size of the objects to one single pixel ensured a reliable measurement of both the riparian zone width and the

Fig. 6. Land cover classification: (a) subset of the study area (green, red and NIR bands) and (b) classification result for the same subset. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of the article.)

streambed width. The distance in Definiens is estimated from the centre of each object, being influenced by its shape and orientation. This bias was eliminated by working with individual pixels. The average riparian zone width and streambed width for the study area were 57.11 and 13.23 m respectively. Thirty-four measurements of the riparian zone width were employed for the validation of the riparian zone width assessment. Comparison between the reference and estimated riparian zone widths showed a strong correlation (r = 0.82; RMSE = 13.9 m), with a slight overestimation of the automatic assessment of riparian zone width in some areas (Fig. 8). This overestimation was linked to the presence of woodland areas close to the riparian zone, which were in some cases misclassified as riparian vegetation, and therefore included in the riparian zone width estimation. Although

Table 2 Error matrix of the land cover classification for bare ground (BG), riparian vegetation (RV), woodlands (WL), rangelands (RL), streambed without vegetation overhang (SC) and streambed with vegetation overhang (VO) expressed in m2. Reference data

Classified data BG RV WL RL SC VO Producer’s accuracy

User’s accuracy

BG

RV

WL

RL

SC

VO

230.4 0 0 23.04 0 0

0 345.6 23.04 0 0 23.04

0 138.24 299.52 0 0 0

23.04 0 0 253.44 0 0

0 46.08 0 0 276.48 0

0 23.04 0 0 0 368.64

90.91%

Overall classification accuracy = 85.6%.

88.24%

68.42%

91.67%

85.71%

94.12%

90.91% 62.50% 92.86% 91.67% 100.00% 94.12%

604

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

Fig. 7. Riparian zone width estimation. Pixels representing the edge of the riparian zone are presented in different colours according to their distance to the streambed. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of the article.)

establishing the boundary between riparian vegetation and woodlands is challenging (even when it is performed in the field), the overall accuracy of the automatic estimation was high, with an average error of 3.9 m, equivalent to less than two pixels in the image. Because the riparian zone width estimation was based on the land cover classification map, the results relied heavily on the image classification accuracy. Finally, the streambed width measurements obtained from the streambed map and the ones measured in the field showed a very high correlation, with a correlation coefficient (r) of 0.98 (RMSE = 1.53 m). 4. Discussion 4.1. Integration of active and passive remote sensors for the analysis of riparian biophysical parameters and land cover types Most of the products derived in this study were obtained through the combination of both optical and LiDAR-derived information. The spectral information was a key element in identifying land cover types; and the inclusion of LiDAR data allowed the incorporation of the vertical distribution of the vegetation strata to the analysis. Integration of LiDAR and optical data also facilitated the image segmentation. The incorporation of LiDAR data to the analysis allowed accurate mapping of PPC and TCM, which were essential inputs

for the OBIA. Moreover, these structural parameters can be used for the analysis of the riparian condition, as tree heights and PPC act as riparian health indicators (Dixon et al., 2005). Previous research has demonstrated that the potential of optical data for estimating such structural parameters of the riparian zone is limited, since the accuracy of such estimates can be significantly affected by the type of vegetation present in the area (Johansen and Phinn, 2006). The results found in the present study show that such riparian health indicators can be accurately derived with the inclusion of LiDAR data to the analysis. The TCM and PPC layers also assisted in the identification of forested areas (woodlands and riparian vegetation). The LiDAR data were also essential for mapping elements from underneath the canopy cover (e.g. the streambed), from where optical sensors cannot provide information. The streambed map can be used as an indicator of water flow volume in the river (Akasheh et al., 2008). Accurate location of the streambed was also necessary in order to map vegetation overhang and to estimate the riparian zone width and the streambed width. The distribution of vegetation overhang is an important riparian health condition indicator, as shading by vegetation overhang affects incident solar radiation and water temperature (Ghermandi et al., 2009). According to our results, 83.5% of the streambed was located underneath vegetation and was not detectable from the QuickBird data. At the same time, the incorporation of the QuickBird information allowed the improvement of the LiDAR-derived streambed map, which underestimated the streambed extent in some areas. The optical information was decisive for the land cover type classification. Although the intensity of the LiDAR pulse can be related to the different surfaces that the pulses have hit (Antonarakis et al., 2008), this information appeared less sensitive to land cover type than optical data, and discrimination of certain surfaces from LiDAR information was difficult. For example, the intensity values appeared similar for bare ground and grass at the single wavelength (1047 nm) of the LiDAR pulse. In contrast, these land cover types were easily separated using the multi-spectral optical information. The integration of LiDAR information to the OBIA allowed discrimination of those land cover classes that share similar spectral behaviour but different vertical distribution (e.g. green grasses and dense forested areas). The combined use of both datasets provided the best possible land cover type classification. This confirmed the feasibility of combining both sensors for the assessment of the riparian condition, rather than selecting one over the other. 4.2. Adoption of an object based approach for the analysis of riparian zones

Fig. 8. Scatter plot of the riparian zone width estimation vs the reference values. The ellipse shows examples of overestimation of the riparian zone width. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of the article.)

The adoption of an object based approach for the analysis of the riparian biophysical parameters and land cover types was beneficial in the study area environment for several reasons. According to our results, the integration of both datasets was beneficial for the analysis of the riparian zone. In that sense, the use of an OBIA was decisive in order to combine spectral and structural data in a single analysis. The OBIA acted as a valuable platform for combining two datasets with very different patterns and resolution. According to Goetz et al. (2003), land cover type classification was problematic when applying traditional supervised and unsupervised classifications on IKONOS imagery due to limited spectral discrimination of covers. The OBIA aided this limitation by enabling the incorporation of context features to the description of the land cover types. This was very useful for discriminating between riparian vegetation and woodlands, which, in addition to other features, were described in relation to context features (e.g.

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

distance to the streambed). Even though the exact extent and boundary of the riparian zone is often difficult to delineate and remains an issue of debate (Apan et al., 2002), the riparian zone was accurately mapped by the OBIA (user’s accuracy of 63.2% and producer’s accuracy of 100%). The incorporation of context features also assisted to identify streambed with and without vegetation overhang. Similar results were obtained by Conchedda et al. (2008) after adopting an OBIA for mapping mangrove ecosystems. Finally, the use of an OBIA was found suitable for the analysis of the high spatial resolution datasets employed in this study. Although high spatial resolution imagery is essential in order to map the high spatial variability of the riparian zone (Akasheh et al., 2008; Goetz et al., 2003; Johansen et al., 2007a; Muller, 1997), this imagery can be difficult to analyze. When the size of the pixel becomes smaller than the size of the elements of interest, the within-class variance increases and spectral signatures of single pixels only correspond to components of the elements of interest, making them less representative (Arroyo et al., 2006). After segmentation, the average size of the objects to classify matched the average size of the elements to classify, assisting the identification of riparian land cover types. Also, in the OBIA the within-class variance did not represent ‘‘noise’’, but useful information to incorporate into the classification. Similar to ecological studies, where local structure and heterogeneity is an intrinsic characteristic of riparian vegetation patches (Muller, 1997), the OBIA used the within-class variance as representative characteristics of the riparian land covers. 4.3. The adoption of remote sensing techniques for the assessment of riparian condition Our results show that remote sensing techniques can potentially be used for the assessment of riparian condition. This method may be useful for assessing riparian condition in other savanna areas, although the robustness of the OBIA approach should be tested further. This especially significant in the case of the Australian savannas, where the riparian zones are spread out across a vast area (approximately 25% Australia), much of which is remote from major population centres and is characterised by a low population density (Dixon et al., 2005). Muller (1997) established two specific requirements for adopting remote sensing techniques for the riparian assessment: high spatial resolution and spatially based classification algorithms. Later studies revealed that information of the vertical distribution of vegetation was also critical in order to accurately assess structural parameters of the riparian zone (Akasheh et al., 2008; Johansen et al., 2007b). The results presented in this paper show that current sensors provide the information required at an appropriate spatial resolution, both for optical and LiDAR data. In this sense, the commercial availability of LiDAR sensors represents a very important advance. OBIA completes the requirements established by Muller (1997). This approach has developed rapidly in the 21st century and it is providing encouraging results in several areas of research, such change detection (Conchedda et al., 2008; Johansen et al., 2008); forest fuels mapping (Arroyo et al., 2006), classification of wildland–urban interface (Cleve et al., 2008), and forest structure characterization (Pascual et al., 2008). It is important to note here that the riparian health indicators estimated for this study were originally developed for assessment in the field. A better result would be expected if riparian health indicators were developed based on the significant possibilities that remote sensing provides. Remote sensing techniques have the potential of mapping several biophysical parameters (e.g. canopy cover, water content, biomass) with high levels of precision. In some cases, such biophysical parameters have not been considered when evaluating the riparian condition because their assessment

605

in the field would be too expensive, time consuming and lack representativeness. However, there are available powerful sensors and methods to measure such biophysical parameters from remotely sensed data. Further research is encouraged in order to develop and evaluate new remotely sensed riparian health indicators for the analysis of riparian condition. 5. Conclusions Using remote sensing for mapping riparian vegetation is difficult due to its complexity, high diversity and spatial variability occurring at finer spatial scales. In this study, we suggest to address these limitations by adopting an OBIA and the combined use of LiDAR and high spatial resolution optical image data. Riparian zone and streambed width, vegetation overhang, and riparian land cover classes were accurately mapped from LiDAR and high spatial resolution optical data within a 5 km stretch of riparian zones in an Australian tropical savanna environment. The integration of both sources of information produced an accurate land cover map, despite the high level of heterogeneity within the riparian landscape. This allowed accurate identification of riparian vegetation, vegetation overhang and the streambed, all of which are commonly used indicators of riparian zone condition. Moreover, the analysis developed allowed an accurate assessment of the riparian zone width and areas of streambed with vegetation overhang. The high levels of accuracy obtained confirm that combining LiDAR and high spatial resolution satellite imagery can significantly improve the mapping and assessment of the riparian condition. The OBIA was suitable for this type of data integration. This approach also assisted the classification by allowing the incorporation of context information to the classes’ definition. Our results have implications for riparian management in tropical savannas as a tool for monitoring vegetation structure and composition remotely. Further research in this direction should be focused on the estimation and incorporation of other remotely derived riparian health indicators. Acknowledgements Help with the fieldwork and analysis was provided by Santosh Bhandari and Andrew Clark. L.A. Arroyo is funded by the Fundacion Alonso Martin Escudero (Spain). K. Johansen is supported by an Australian Research Council Linkage Grant to K. Mengersen, S. Phinn, and C. Witte. Part of the results presented in this paper was presented at the Silvilaser conference, held in September 2008 in Edinburgh. References Akasheh, O.Z., Neale, C.M.U., Jayanthi, H., 2008. Detailed mapping of riparian vegetation in the middle Rio Grande River using high resolution multi-spectral airborne remote sensing. Journal of Arid Environments, doi:10.1016/ j.jaridenv.2008.03.014. Antonarakis, A.S., Richards, K.S., Brasington, J., 2008. Object-based land cover classification using airborne LiDAR. Remote Sensing of Environment 112, 2988–2998. Apan, A.A., Raine, S.R., Paterson, M.S., 2002. Mapping and analysis of changes in the riparian landscape structure of the Lockyer Valley catchment, Queensland, Australia. Landscape and Urban Planning 59, 43–57. Armston, J.D., Denham, R.J., Danaher, T.J., Scarth, P.F., Moffiet, T., 2009. Prediction and validation of foliage projective cover from Landsat-5 TM and Landsat-7 ETM+ imagery for Queensland, Australia. Journal of Applied Remote Sensing 3, 033540, doi:10.1117/1.3216031. Arroyo, L.A., Healey, S.P., Cohen, W.B., Cocero, D., 2006. Using object-oriented classification and high-resolution imagery to map fuel types in a Mediterranean region. Journal of Geophysical Research 111, G04S04, doi:10.1029/ 2005JG000120. Bain, M.B., Harig, A.L., Loucks, D.P., Goforth, R.R., Mills, K.E., 2000. Aquatic ecosystem protection and restoration: advances in methods for assessment and evaluation. Environmental Science & Policy 3, S89–S98.

606

L.A. Arroyo et al. / Forest Ecology and Management 259 (2010) 598–606

Blaschke, T., 2009. Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, doi:10.1016/j.isprsjprs.2009.06.004. Clark, M.L., Clark, D.B., Roberts, D.A., 2004. Small-footprint lidar estimation of subcanopy elevation and tree height in a tropical rain forest landscape. Remote Sensing of Environment 91, 68–89. Cleve, C., Kelly, M., Kearns, F.R., Moritz, M., 2008. Classification of the wildland– urban interface: a comparison of pixel- and object-based classifications using high-resolution aerial photography. Computers, Environment and Urban Systems 32, 317–326. Conchedda, G., Durieux, L., Mayaux, P., 2008. An object-based method for mapping and change analysis in mangrove ecosystems. ISPRS Journal of Photogrammetry and Remote Sensing 63, 578–589. Congalton, R.G., Birch, K., Jones, R., Schriever, J., 2002. Evaluating remote sensed techniques for mapping riparian vegetation. Computers and Electronics in Agriculture 37, 113–126. Definiens, A.G., 2007. Definiens Developer 7—User Guide. Definiens AG, Germany, 497 pp. Dixon, I., Douglas, M., Dowe, J., Burrows, D.W., Townsend, S.A., 2005. A rapid method for assessing the condition of riparian zones in the wet/dry tropics of northern Australia. In: Proceedings of the 4th Australian Stream Management Conference, Launceston, Australia, 20–22 October 2004. Department of Primary Industries, Water and Environment, pp. 173–178. Fry, J., Steiner, F.R., Greenb, D.M., 1994. Riparian evaluation and site assessment in Arizona. Landscape and Urban Planning 28, 179–199. Ghermandi, A., Vandenberghe, V., Benedetti, L., Bauwens, W., Vanrolleghem, P.A., 2009. Model-based assessment of shading effect by riparian vegetation on river water quality. Ecological Engineering 35 (1), 92–104. Gilmore, M.S., Wilson, E.H., Barrett, N., Civco, D.L., Prisloe, S., Hurd, J.D., Chadwick, C., 2008. Integrating multi-temporal spectral and structural information to map wetland vegetation in a lower Connecticut River tidal marsh. Remote Sensing of Environment 112, 4048–4060. Goetz, S.J., Wright, R.K., Smith, A.J., Zinecker, E., Schaub, E., 2003. IKONOS imagery for resource management: tree cover, impervious surfaces, and riparian buffer analyses in the mid-Atlantic region. Remote Sensing of Environment 88, 195– 208. Goodwin, N.R., Coops, N.C., Culvenor, D.S., 2006. Assessment of forest structure with airborne LiDAR and the effects of platform altitude. Remote Sensing of Environment 103, 140–152. Innis, S.A., Naiman, R.J., Elliott, S.R., 2000. Indicators and assessment methods for measuring the ecological integrity of semi-aquatic terrestrial environments. Hydrobiologia 422/423, 111–131. Johansen, K., Phinn, S., 2006. Mapping structural parameters and species composition of riparian vegetation using IKONOS and Landsat ETM+ data in Australian tropical savannahs. Photogrammetric Engineering and Remote Sensing 72, 71–80. Johansen, K., Coops, N.C., Gergel, S.E., Stange, Y., 2007a. Application of high spatial resolution satellite imagery for riparian and forest ecosystem classification. Remote Sensing of Environment 110, 24–44. Johansen, K., Phinn, S., Dixon, I., Douglas, M., Lowry, J., 2007b. Comparison of image and rapid field assessments of riparian zone condition in Australian tropical savannas. Forest Ecology and Management 240, 42–60. Johansen, K., Arroyo, L.A., Phinn, S., Witte, C., 2008. Object-oriented change detection of riparian environments from high spatial resolution multi-spectral images. In: Proceedings of the GEOgraphic Object Based Image Analysis for

the 21st Century (GEOBIA): Pixels, Objects, Intelligence, Calgary, Canada, pp. 1– 6. Jones, K.B., Edmonds, C.E., Slonecker, E.T., Wickham, J.D., Neale, A.C., Wade, T.G., Riitters, K.H., Kepner, W.G., 2008. Detecting changes in riparian habitat conditions based on patterns of greenness change: a case study from the Upper San Pedro Riber Basin, USA. Ecological Indicators 8, 89–99. Lamonaca, A., Corona, P., Barbati, A., 2008. Exploring forest structural complexity by multi-scale segmentation of VHR imagery. Remote Sensing of Environment 112, 2839–2849. Langer, E.R., Steward, G.A., Kimberley, M.O., 2008. Vegetation structure, composition and effect of pine plantation harvesting on riparian buffers in New Zealand. Forest Ecology and Management 256, 949–957. Lefsky, M.A., Cohen, W.B., Spies, T.A., 2001. An evaluation of alternate remote sensing products for forest inventory, monitoring and mapping of Douglas fir forests in western Oregon. Canadian Journal of Forestry Research 31, 78–87. Lovell, J.L., Jupp, D.L.B., Culvenor, D., Coops, N.C., 2003. Using airborne and groundbased ranging lidar to measure canopy structure in Australian forests. Canadian Journal of Remote Sensing 29, 607–622. Lovett, S., Price, P., 2002. A second phase of the National Riparian Lands R&D program. In: Rip Rap: River and Riparian Lands Management Newsletter, Land and Water Australia, pp. 1–5. Lytle, D.A., Merritt, D.M., 2004. Hydrologic regimes and riparian forests: a structured population model for cottonwood. Ecology 85, 2493–2503. Mallinis, G., Koutsias, N., Tsakiri-Strati, M., Karteris, M., 2008. Object-based classification using Quickbird imagery for delineating forest vegetation polygons in a Mediterranean test site. ISPRS Journal of Photogrammetry and Remote Sensing 63, 237–250. Mertes, L.A.K., 2002. Remote sensing of riverine landscapes. Freshwater Biology 47, 799–816. Muller, E., 1997. Mapping riparian vegetation along rivers: old concepts and new methods. Aquatic Botany 58, 411–437. Pascual, C., Garcia-Abril, A., Garcia-Montero, L.G., Martin-Fernandez, S., Cohen, W.B., 2008. Object-based semi-automatic approach for forest structure characterization using lidar data in heterogeneous Pinus sylvestris stands. Forest Ecology and Management 255, 3677–3685. Persson, A., Holmgren, J., Soderman, U., 2002. Detecting and measuring individual trees using an airborne laser scanner. Photogrammetric Engineering and Remote Sensing 28, 1–8. Popescu, S.C., Zhao, K., 2008. A voxel-based lidar method for estimating crown base height for deciduous and pine trees. Remote Sensing of Environment 112, 767– 781. Simon, A., Collison, A.J.C., 2002. Quantifying the mechanical and hydrologic effects of riparian vegetation on streambank stability. Earth Surfaces Processes and Landforms 27, 527–546. Suarez, J.C., Ontiveros, C., Smith, S., Snape, S., 2005. Use of airborne LiDAR and aerial photography in the estimation of individual tree heights in forestry. Computers & Geosciences 31, 253–262. Van Gardingen, P.R., Jackson, G.E., Hernandez-Daumas, S., Russel, G., Sharp, L., 1999. Leaf area index estimates obtained for clumped canopies using hemispherical photography. Agricultural and Forest Meteorology 94, 243–257. Zimble, D.A., Evans, D.L., Carlson, G.C., Parker, R.C., Grado, S.C., Gerard, P.D., 2003. Characterizing vertical forest structure using small-footprint airborne LiDAR. Remote Sensing of Environment 87, 171–182.