OPERATIONAL USE OF ANNUAL LANDSAT-5 TM AND LANDSAT-7 ETM+ IMAGE TIME SERIES FOR MAPPING WOODED EXTENT AND

FOLIAGE PROJECTIVE COVER IN NORTH-EASTERN AUSTRALIA
Joanna Kitchen , John Armston , Andrew Clark , Tim Danaher Peter Scarth 1,3
1 1 1,3 1 2,3

,

Remote Sensing Centre, Queensland Department of Environment and Resource Management, 80 Meiers Road, Indooroopilly, Qld, Australia, 4068 Tel: +61 7 3896 9616, Email: Andrew.J.Clark@derm.qld.gov.au
2

Information Sciences Branch, Department of Environment, Climate Change and Water, PO Box 856, Alstonville, NSW, Australia, 2477 Tel: +61 2 6627 0224, Email: Tim.Danaher@environment.nsw.gov.au

3

Joint Remote Sensing Research Program, Centre for Spatial Environmental Research, School of Geography, Planning and Environmental Management, The University of Queensland

Abstract Large area and long-term spatial data on vegetation cover is critical for environmental legislation and carbon emission calculation and assessment. In Queensland and New South Wales (NSW) there is a requirement to produce accurate statewide maps of wooded extent and foliage projective cover (FPC). Approaches that use single date satellite images are limited if photosynthetic herbaceous foliage is present at the time of image acquisition. Therefore an approach was developed that uses an annual time series of Landsat-5 TM and Landsat-7 ETM+ imagery from 1986 to 2008, which consists of over 2500 images for Queensland and NSW. In order to be insensitive to perturbations such as fire, robust and ordinary least-squares linear regression were used to model temporal trends in an annual FPC index. Thresholding of derived temporal statistics was used to classify wooded extent. Thresholds were optimised using Queensland-wide field data. Estimates of FPC were derived from the regression model trend estimates. The classification was stratified using ancillary datasets on the spatial extent of cropping and forestry plantations. Quantitative validation of the final product showed good agreement 2 2 with field (r 0.78) and lidar (r 0.93) estimates of FPC. Work is underway to improve the classification of wooded extent using radar imagery and the extraction of temporal trends in FPC using the full Landsat archive. Introduction The Queensland and New South Wales Governments require large area mapping and monitoring of vegetation cover for regulation, compliance, restoration, and other management actions. The metric of vegetation cover

1

adopted in many Australian vegetation classification frameworks is foliage projective cover (FPC) (Sun et al., 1997). FPC is defined as the horizontally projected percentage cover of photosynthetic foliage of all strata (Specht, 1983). However in this paper FPC refers to the vertically projected percentage cover of photosynthetic foliage from tree and shrub life forms only. Since Australian plant communities are dominated by trees and shrubs with sparse foliage and irregular crown shapes, FPC is a more suitable indicator of a plant community’s radiation interception and transpiration than crown cover (Specht and Specht, 1999). Approximately 11% FPC corresponds to 20% crown cover (Scarth et al., 2008), which is the vegetation cover metric used in the definition of forest extent according to Kyoto Protocol rules (AGO, 2003). In this paper, wooded extent refers to the presence of FPC (1−100%). Operational mapping of wooded extent and FPC requires an efficient and automated method due to the large volume of satellite and ancillary data that require processing and interpretation. Previous research has developed regression modelling and spectral unmixing-based approaches to estimating FPC, however using spectral information alone makes separating woody and herbaceous cover difficult over large areas (Armston et al., 2009 and references therein). More success has been achieved over large areas by time series decomposition of data from multi-spectral, high temporal resolution, coarse spatial resolution sensors such as the Advanced Very High Resolution Radiometer (AVHRR) and the Moderate Resolution Imaging Spectrometer (MODIS). This is simply due to evergreen wooded vegetation generally exhibiting less seasonality in time series of vegetation indices compared to herbaceous cover (Lu et al., 2003; Gill et al., 2009). The Statewide Landcover and Trees Study (SLATS ) required a method for an annual time series (from 1986 to date) of dry-season Landsat-5 Thematic Mapper (TM) and Landsat-7 Enhanced Thematic Mapper Plus (ETM+) as these data are at a resolution and extent suitable for end-users. The annual sampling of these data means techniques developed for AVHRR and MODIS time series can not be directly applied. A small number of studies have used spectral unmixing and linear trend analyses to assess an annual Landsat image time series (Röder et al., 2008; Hostert et al., 2003). However these techniques have not been implemented and validated with ground data in Australian environments. Additional challenges with operationally using an annual Landsat time series consisting of over 2500 images include the removal of time series contamination caused by water, fire/smoke, cloud and cloud shadow (Liang et al., 2002; Hostert et al., 2003). This paper presents an overview of the approach used to operationally map wooded extent and FPC across Queensland and New South Wales, Australia. Initially, the algorithm and various datasets used are described. The calibration and validation of the method using extensive field and lidar data are then presented. Finally, the limitations of the product and its operational
1

1

http://www.derm.qld.gov.au/slats

2

implementation are discussed along with the research currently underway to address them. Data and Methods Field, image and ancillary data The sampling design, field methods and acquisition of greater than 1900 SLATS field observations of stand basal area (SBA) used for calibration of the wooded extent and FPC products are described by Lucas et al. (2006) and Armston et al. (2009). Coincident field and Optech ALTM3025 airborne lidar data were used for the validation of FPC predictions. Nineteen lidar surveys were acquired with a total of 47 field sites collected from 16 of the 19 survey sites within one month of acquisition. The acquisition and processing the field and lidar data is described by Armston et al. (2009). TM and ETM+ imagery were acquired with Australian Centre for Remote Sensing (ACRES) Level-5 processing. Image dates were restricted to dry season months (May to October inclusive) for Queensland to minimize cloud cover and photosynthetic herbaceous cover. No restrictions were placed on NSW image dates but most were between September and March. All TM and ETM+ images were geometrically registered to an orthorectified ETM+ image mosaic (Gill et al., 2010a). Radiometric calibration (de Vries et al., 2007) and an empirical bi-directional reflection distribution function (BRDF) correction (Danaher et al., 2002) were applied. The NSW image data also had a topographic correction applied (Gill et al., 2010b). Reflectance values affected by cloud, cloud shadow, smoke, water or topographic shadow in any image date were removed using Landsat-derived masks (Shaw and Gillingham, 2006; Danaher and Collett, 2006). Ancillary data used to stratify the wooded extent and FPC algorithm included plantation and crop masks sourced from the Queensland Land Use Mapping Project (QLUMP) for 1999 (Witte et al., 2006). Crop classes of “Cropping”, “Perennial and Seasonal Horticulture”, “Irrigated Cropping”, and “Irrigated Perennial and Seasonal Horticulture” were combined to form a binary mask of crop/non-crop. Commission errors due to creeks, windbreaks and roads were manually edited out. For the binary forestry/non-forestry mask, forestry classes included “Plantation Forestry” and “Production Forestry”. The ALUM (Australian Land Use Mapping) Version 6 dataset 2 was used to produce plantation and crop masks for NSW. The Shuttle Radar Topography Mission (SRTM) 90 m Digital Surface Model (Farr et al., 2007) was used to generate a slope raster using a 3x3 kernel. Fitting trends to the annual time series Time series of an annual FPC index were generated based on a multiple linear regression (MLR) model developed using pre-processed Landsat imagery, a

2

http://adl.brs.gov.au/mapserv/landuse

3

climatological ancillary variable, vapor-pressure deficit (VPD), and field calibration data collected from over 1900 sites in Queensland (Lucas et al., 2 2006). Cross-validation showed this model had an adjusted R of 0.80 and an RMSE of 9.9%. Armston et al. (2009) independently validated the MLR model using FPC derived from field and airborne lidar data. The model provided a prediction accuracy of <10% RMSE but showed greater than 10% bias in plant communities with a high level of photosynthetic herbaceous cover. The next step was to separate the trend from the variance in the annual FPC index time series. The annual FPC index values are therefore assumed to be a combination of herbaceous and woody cover with an error component. Up to 22 near-annual Landsat-derived FPC images between 1986 and 2008 were available for each of the 87 scenes covering Queensland and 44 scenes covering NSW. To account for the irregular sequence of image dates and annual sampling, least-squares linear regression between FPC and the image date was used. This is consistent with similar studies using Landsat time series (Röder et al., 2008; Furby et al., 2009). Least-squares linear regression is very sensitive to outliers and a method was required to identify unusual observations caused by fire, unclassified cloud or other extreme events that perturbed an otherwise stable time series. Iteratively re-weighted least-squares and a Huber weighting function with a tuning constant of 2 was used to fit a robust regression via M-estimation (Huber, 1981). If the resulting residuals were greater than or equal to MAR/0.6745, where MAR is the median absolute residual, then the corresponding observations were labelled as outliers. Multiple change events or extreme values in a time series, e.g. that caused by cropping management practices, were not labelled as outliers using this approach. In the case of the final date in the time series, it was also important to test for a major difference between the fitted value from the robust regression and the actual value for the final year of the time series, a “switch” statistic (S) was calculated by
S= FPC n − Yn , ˆ σ

(1)

where FPCn is the annual FPC index value and Yn is the fitted robust regression ˆ value for the final date in the time series, and σ is the standard error of the least-squares linear regression fit. S was used to switch between the Landsat time series with and without outliers removed in order to account for events such as clearing that occurred late in an otherwise stable time series. In the case of land clearing or a similar major disturbance event followed by woody regrowth, it was invalid to fit a single regression line to the entire time series. Therefore, a method was developed to split the time series by calculating the most likely split point based on a t-test of the regression fit (Figure 1). The method iteratively calculates regression lines from the end of the time period back to each date and then selects the point with the highest tvalue as the split point.

4

100%

8

80%

FPC time-series Best trend estimate Lagged t-test

7 6 5 4

Annual FPC index

60%

40%

`

3 2

20% 1 0% 1985 0 1989 1993 1997 2001 2005 2009

Year

Figure 1. Example of the trend-fitting method used to split the annual FPC index time series when a major disturbance event followed by woody regrowth is detected. The green line is the annual FPC index time series and diamonds show the significance of the regression fit from 2009 back in time to any given year. In this example, the time series is split in 1998. The figure is adapted from Danaher et al. (2010).

Classification of wooded extent and FPC In order to discriminate wooded land cover from non-wooded areas such as crop and pasture, a decision tree classification was developed based on thresholds applied to the Landsat time series minimum and a normalised standard error of the fitted trend. The time series minimum (Ymin) is often assumed to represent the persistent or woody component of vegetation cover (Roderick et al., 1999; Donohue et al., 2009). Ymin was calculated here as

min(Y ) if S ≥ x Ymin =  , min(Yo ) if S < x

t-value

(2)

where Y is the annual FPC index time series, Y0 is the same time series with outliers removed and x is a threshold to be optimised. Outliers were removed in the calculation of the minimum statistic to avoid unusually dry images causing sparsely wooded regions to be classified as non-wooded (e.g. due to droughtinduced leaf drop and fire). In Australia, several studies have shown that the temporal variance of a spectral vegetation index for wooded vegetation is less than that of non-wooded vegetation, particularly compared to pastures and crops (Lu et al., 2003; Gill et al., 2009; Donohue et al., 2009). In the case of the Landsat time series here, a normalised standard error (NSE) was calculated as
NSE = ˆ σ Ymin

,

(3)

5

ˆ ˆ where σ is the standard error of the least-squares linear regression fit. σ was used instead of the variance to account for areas with significant trend in woody vegetation, such as woody regrowth cleared early in the time series. The NSE was found to improve the contrast between wooded and non-wooded land cover, especially in areas of low FPC such as open savannas that have high temporal variation in herbaceous cover.

Thresholds were applied to S, Ymin and NSE in the first component of the decision tree shown in Figure 2. If the conditions were true then the classification was nominally non-wooded. Default thresholds for S, Ymin and NSE were derived using the statewide SLATS SBA field dataset. Field sites with a SBA greater than zero were categorized as wooded. Additional sites in non-wooded cropping areas were extracted from Landsat data through manual interpretation of the imagery and aerial photography. These data were then used to optimise the S, Ymin and NSE thresholds by maximizing the Kappa statistic, a measure of omission and commission, using a global optimization technique (genetic algorithm; Yearstretch GASolver 3 Version 1.5.6, 2002). The classification in Figure 2 then proceeds through a series of decisions designed to minimise specific cases of wooded omission or commission error:

Due to the absence of a Landsat radiance-based topographic correction for Queensland and limitations of the NSW correction, decisions were developed to avoid wooded omission error on spectrally bright eastern aspect slopes. The elevation slope threshold was 25° for Queensland and 45° for NSW. The Landsat time series often failed to capture the management cycle of crops such as sugar cane, which often resulted in time series similar to wooded vegetation. Areas within the crop mask were classified as nonwooded. Areas used for forestry plantations are subject to fire, woody regrowth control following clearing, selective and full timber harvesting. This resulted in a low Ymin and high NSE. Therefore the least squares linear regression fit to the annual FPC index time series was split (Figure 1). With reference to available independent data, thresholds to separate wooded and nonwooded land cover were developed though visual interpretation (Figure 2; SE is the standard error of the least squares linear regression fit). In order to detect woody regrowth following clearing within the annual FPC index time series, the least squares linear regression fit was also split for areas remaining nominally classified as non-wooded from the first decision in Figure 2. Due to the different temporal dynamics and management practices that occur in areas of natural vegetation compared to plantations, a different set of thresholds to separate wooded and non-wooded land cover were developed though visual interpretation (Figure 2).

3

http://www.yearstretch.com/gasolver.htm

6

NSE <= x Ymin >= x True False

Within crop mask?

Elevation slope >= x% & Robust fitted Yn > 20% True

True

False

False

Non-Wooded

Wooded*

Within plantation mask?

Wooded*

False

True

Split least-squares regression: Slope >= 0.004 & Fitted Yn >= 25% & SE <= 11% & No. images >= 5 & (Yn >= 15% | Yn = 0%) False True

Split least-squares regression: Slope > 0 & Fitted Yn > 15% & R-Squared >= 0.6 False True

Non-Wooded

Wooded*

Non-Wooded

Wooded*

Figure 2. Flow chart of the decision process followed to classify Landsat annual time series as wooded or non-wooded. Thresholds that are to be optimised or are dependent on State are denoted x. * It is possible for an individual time series classified as “Wooded” to have a value of zero after prediction of FPC.

If the annual FPC index time series had ≥ 3 observations and was classified as wooded in Figure 2, the FPC value (0−100%) was derived from the robust linear regression fitted trend line for the final date in the time series (Yn). The assumption was that the robust linear regression fit successfully separated trends from seasonality (particularly that from herbaceous vegetation) and noise. There were two exceptions to using the robust linear regression fit. Firstly, if the S statistic was greater than or equal to the optimised threshold, then the annual FPC index value (Yn) was used. Secondly, if the least-squares linear regression split shown in Figure 1 was applied, the fitted value for Yn from the result was used. Post-processing Two automated post-processing steps were implemented at the end of the classification process to correct any missing data values and apply a multitemporal water mask. A large proportion of the Landsat image time series over Cape York Peninsula were cloud contaminated. It was therefore difficult to retain at least three valid observations for any given time series, and the resulting pixel value was often the null value. These pixels were corrected by

7

replacing the null value with the most recent cloud-free annual FPC index value. In most instances, the values were from imagery close to the nominal prediction date. As significantly reduced water extent exists in Landsat images acquired during periods of drought or coastal images acquired at low tide, large areas of tidal zones, dams and lakes were incorrectly classified as wooded or non-wooded due to the herbaceous ground cover or bare ground, respectively. For each water mask date, erroneous water pixels were removed if they had been also classified as cloud, cloud shadow, or topographic shadow in the other available masks. The resulting time series of water masks were then used to calculate the number of dates with inundation. If the number of inundation dates were ≥ 7, it was recoded to the null value. This threshold was manually lowered for some scenes depending on the range of tides and conditions present in the time series, and the number of available Landsat images. Validation The Landsat FPC estimates were compared to field and lidar derived FPC. The field and lidar estimates of FPC were derived following the procedures of Armston et al. (2009). The area covered by the lidar data is much greater than the field data, therefore the lidar data was spatially aggregated to Regional Ecosystem (RE; Accad et al., 2006). RE’s are defined as areas of consistent landform, geology and vegetation and are also the basis for vegetation management policy in Queensland. As the lidar data essentially provides many replicates of FPC measurements at the field plot scale within an RE, this validation approach permitted the discrimination between random, systematic and locally systematic errors (Armston et al., 2009).

Results and Discussion Decision tree threshold optimisation The optimised thresholds were 92 (NSE), 103 (3% annual FPC index; Ymin) and 125 (25% annual FPC index; S). The Kappa statistic for the model calibration was 86.86%, which is a good result considering one set of thresholds was used for all of Queensland. Table 1 shows the error matrix resulting from the optimisation using the Queensland-wide SBA field dataset, with an overall model error of 6.46%, non-wooded omission error of 7.20% and non-wooded commission error of 5.90%. It was found that to improve local accuracy upon application of the model the optimised thresholds were manually adjusted for a minor proportion of individual Landsat scenes. S and the thresholds used in the final decision in Figure 2 were the most common thresholds manually adjusted. It is important to note that the manual adjustment of thresholds was done in reference to field observations and aerial photography by trained operators and included several stages of checking to ensure consistency.

8

Table 1. Error matrix resulting from optimisation of the wooded extent classification thresholds using the Queensland-wide SBA field dataset.
Observed Wooded Predicted Wooded Non-wooded Total 1238 96 1334 Non-wooded 102 1627 1729 Total 1340 1723 3063

The requirement for manual adjustment of thresholds is in part due to the use of one set of thresholds for all of Queensland, but is also due to the annual sampling of the available Landsat imagery not being of sufficient temporal resolution to: (i) consistently capture the range of soil and vegetation conditions; and (ii) to adequately separate the seasonal and trend components of the time series. As an example, Figure 3 shows: (i) a Landsat-5 TM image subset; (ii) a composite image consisting of normalised standard error (red) minimum (green) and slope (blue), to which the optimised thresholds were applied; and (iii) the resulting wooded extent and FPC product. The composite image highlights areas of high NSE and low Ymin in red, which includes areas of tree clearing, pasture and cropping. Yellow and green areas indicate stable woody vegetation cover (forest) with high Ymin and low NSE.

Figure 3. (i) dry season Landsat-5 TM (RGB – Bands 5/4/2); (ii) RGB composite image of normalised standard error (NSE), minimum (Ymin), and the slope of the fitted trend; and (iii) the 2008 wooded extent and FPC product (the water mask is shown as blue).

Wooded extent and FPC product The Landsat image time series (up to 22 image dates for each individual Landsat scene) were processed to produce the wooded extent and FPC product. The final decision in Figure 2 used for the classification of wooded extent is also retained as a separate product. The woody extent and FPC product was then mosaiced for Queensland and NSW. Figure 4 shows the 2008 wooded extent and FPC product mosaic for Queensland (dry season; May to October) and NSW (September to March; including the ACT).

9

Figure 4. 2008 mosaic of wooded extent and FPC for Queensland, New South Wales and the Australian Capital Territory. Water and missing data are coloured blue.

10

Similar to the annual FPC index and other existing mapping (Armston et al., 2009), Figure 4 shows the confinement of the high FPC (and above-ground biomass) vegetation to closed forests along the eastern Queensland and NSW coastal areas. The large expanses of lower FPC woodlands through the Mulga Lands Bioregion in the south west and tropical savanna in the Gulf of Carpentaria and Cape York Peninsula regions are consistent with field observations of wooded extent. Manual checking by trained operators has indicated that the classification performed well overall. However a number of limitations have been observed in local areas. Figure 5 shows examples of these areas that include: (i) underclassification of mulga (Acacia aneura) in western Queensland due to very low foliage cover (<10%), often related to their narrow pendulous phyllodes and leaf drop during times of drought; (ii) improved pastures of the Wet Tropics, which are perennially green and exhibit similar time series properties to wooded vegetation; and (iii) areas of woody regrowth following clearing, where automated splitting of the time series was unsuccessful due to high inter-annual variance in the annual FPC index.

Figure 5. Examples of wooded extent and FPC product limitations: (i) areas of mulga (Acacia aneura) classified as non-wooded; (ii) areas of improved pastures classified as wooded; and (iii) areas of woody regrowth classified as non-wooded. Top panel shows Landsat-5 TM (RGB – Bands 5/4/2) and the bottom panel shows the wooded extent and FPC product. Arrows highlight a specific example location for each limitation.

11

Validation of FPC estimates Quantitative validation of the FPC estimates has been undertaken using both field and lidar data. A direct comparison of field and Landsat estimates of FPC 2 is shown in Figure 6. Overall the agreement is good (r 0.78) and there is a near 1:1 correspondence between the estimates. At low FPC (<10%), the Landsatderived product appears to underestimate, which is partly due to the omission of sparse mulga as shown in Figure 5. The two observations where the Landsat-derived estimates over-estimate by 25−35% correspond to closed coastal heathland on seasonally waterlogged alluvial plains and plant communities with understorey dominated by spinifex (Triodia spp.), which has evergreen herbaceous foliage. This result is also consistent with validation of the annual FPC index (Lucas et al., 2006; Armston et al., 2009).

100 80 Field FPC (%) 60 40 20 0 0 20 40 60 80 100 Landsat FPC (% )
Figure 6. Comparison of field and Landsat-derived estimates of FPC.
r2 = 0.78 RMSE = 10.84 n = 47

Figure 7 shows a comparison of Landsat with lidar estimates of FPC (> 0.5 m above ground) that have been spatially aggregated to RE’s. A minimum of 5 observations was required for an RE polygon to be included in the comparison. 2 The relationship for Figure 7 (right) is strong (r of 0.93), however the RMSE error bars show the accuracy of individual predictions within an RE is still quite variable and is the reason for the relatively poor relationship in Figure 6. It also appears that the Landsat FPC systematically under-predicts at > 60%, which is not clearly shown in Figure 6 and is the same result found in the validation of the annual FPC index (Armston et al., 2009). Figure 7 (left) includes RE's where >1% of the lidar derived FPC was below 0.5m above the ground. These RE’s mainly correspond to heathland and other coastal plant communities such as low open Banksia aemula woodland, which are characterised by an understorey

12

of low herbaceous and woody foliage (e.g. Xanthorrhoea spp.). Hence these locally systematic errors with such RE’s included highlights limitations of both the Landsat and lidar estimates of FPC. The Landsat derived FPC estimates are sensitive to evergreen herbaceous and woody foliage below 0.5 m and the lidar derived FPC estimates are unreliable close to ground level, which is largely due to uncertainty in the ground classification (Armston et al., 2009).

r = 0.67

2

r = 0.93

2

80 Lidar FPC (%) 60 40 20

RMSE = 12.22

Lidar FPC (%)

n = 110

80 60 40 20

RMSE = 5.56 n = 94

20

40

60

80

20

40

60

80

Landsat FPC (% )

Landsat FPC (% )

Figure 7. Comparison of lidar estimates of FPC with Landsat predictions. The lidar data has been spatially aggregated to Regional Ecosystems and the error bars show the RMSE. The left plot shows all RE’s sampled and the right plot shows RE’s with less than 1% of their foliage < 0.5 m above the ground.

Conclusions and Future Research This paper has described the data and methods currently used to derive estimates of wooded extent and FPC from an annual time series of Landsat imagery (from 1986 to 2008) in Queensland and NSW, Australia. The wooded extent and FPC product presented has been developed and enhanced over many years and is now considered a robust and operational product. Validation 2 of the FPC estimates showed strong relationships with field (r 0.78) and lidar 2 (r 0.93) derived estimates at the spatial scale of field plots and RE’s, respectively. However bias has been observed in the cases of very sparse FPC, RE’s with evergreen herbaceous foliage, and woody regrowth following clearing. The wooded extent and FPC product provides a core data set for use by State governments in Queensland and NSW. Queensland mosaics have been produced for the years 1999 and 2000 and annually from 2004 to 2009, primarily for use in SLATS reporting. Currently, NSW have only produced the 2008 mosaic as a baseline map of wooded vegetation cover but additional mosaics are planned. While most of the development of the wooded extent and FPC product took place in Queensland, the application of the method to NSW used nationally available input datasets and the image processing framework

13

presented by Schmidt and Gillingham (2008). This has demonstrated that it is possible to implement operationally in other States in Australia. Planned improvements related to operational production and validation of the wooded extent and FPC product described in this paper include:
• • • •

Application of the method to Queensland Landsat images processed with a radiance-based topographic correction (Gill et al., 2010b). Validation of wooded extent estimates in Queensland and NSW using high spatial resolution satellite imagery (e.g. SPOT-5). Extension of field and lidar validation of FPC predictions to NSW. Data acquisitions are currently underway. Correction of wooded extent commission error caused by improved pastures using Advanced Land Observing Satellite (ALOS) Phased Arrayed L-band Synthetic Aperture Radar (PALSAR) imagery (Lucas et al., 2010). A preliminary implementation of this has been completed for Queensland vegetation management policy requirements.

Landsat high temporal resolution image time series are now freely available 4 from the United States Geological Survey (USGS) archive and offer an unprecedented opportunity for consistent large area mapping and monitoring of FPC. Current FPC mapping and monitoring research is focussing on spectraltemporal unmixing of woody and herbaceous cover fractions from the high temporal resolution USGS Landsat archive, using in part the techniques developed by Scarth et al. (2010). This is anticipated to not only improve estimates of FPC, but also enable detection of long-term trends including woodland thickening, drought-related tree death, thinning, and woody regrowth following clearing. Acknowledgments All of the SLATS and NSW DECCW teams are thanked for their efforts in completing the Queensland and NSW classifications. Geoff Horn (NSW DECCW) is thanked for assistance in preparing Figure 4. Craig Shephard and Michael Schmidt are thanked for reviewing a draft version of the manuscript. References Accad, A., Neldner, V.J., Wilson, B.A., and Niehus, R.E., 2006, Remnant Vegetation in Queensland: Analysis of Remnant Vegetation 1997−1999−2000−2001−2003 including Regional Ecosystem Information, Queensland Herbarium, Queensland Environmental Protection Agency, Brisbane. Armston, J.D., Denham, R.J., Danaher, T.J., Scarth, P.F. and Moffiet, T., 2009, Prediction and validation of foliage projective cover from landsat-5 TM and

4

http://landsat.usgs.gov

14

Landsat-7 ETM+ imagery for Queensland, Australia. Journal of Applied Remote Sensing, 3: 033540-28. Australian Greenhouse Office (AGO), 2003, Greenhouse Gas Emissions from Land Use Change in Australia: Results of the National Carbon Accounting System 1988−2001. Australian Greenhouse Office, Canberra. Danaher, T., 2002, An empirical BRDF correction for Landsat TM and ETM+ th imagery. Proceedings of the 11 Australasian Remote Sensing and Photogrammetry Conference, Brisbane, Australia. Danaher, T., Scarth, P., Armston, J., Collett, L., Kitchen, J. and Gillingham, S., 2010, Remote sensing of tree-grass systems: The Eastern Australian Woodlands. In: Ecosystem Function in Savannas: Measurement and Modelling at Landscape to Global Scales, Eds. M.J. Hill and N.P. Hanan, Taylor and Francis (in press). Danaher, T., and Collett, L., 2006, Development, optimisation and multitemporal application of a simple Landsat based water index. Proceedings of the 13th Australasian Remote Sensing and Photogrammetry Conference, Canberra, Australia. de Vries, C., Danaher, T.J., Denham, R., Scarth, P.F. and Phinn, S., 2007, An operational radiometric calibration procedure for the Landsat sensors based on pseudo-invariant target sites. Remote Sensing of Environment, 107(3): 414−429. Donohue, R.J., McVicar, T.R. and Roderick, M. L., 2009, Climate-related trends in Australian vegetation cover as inferred from satellite observations. Global Change Biology, 15: 1025–1039. Farr T.G., Rosen P.A., Caro E., Crippen R., Duren R., Hensley S., Kobrick M., Paller M., Rodriguez E., Roth L., Seal D., Shaffer S., Shimada J., Umland J., Werner M., Oskin M., Burbank D., and Alsdorf D., 2007, The Shuttle Radar Topography Mission. Reviews of Geophysics 45: RG2004. Furby, S.L., Caccetta, P.A., Wallace, J.F., Lehmann, E.A., and Zdunic, K., 2009, Recent developments in vegetation monitoring products from Australia's National Carbon Accounting System. Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS), IV: 276−279. Gill, T., Collett, L., Armston, J., Eustace, A., Danaher, T., Scarth, P., Flood, N. and Phinn, S., 2010a, Geometric correction and accuracy assessment of Landsat-7 ETM+ and Landsat-5 TM imagery used for vegetation cover monitoring in Queensland, Australia from 1988 to 2007. Journal of Spatial Science, in press. Gill, T., Flood, N., Gillingham, S., Danaher, T., Shepherd, J., and Dymond, J., 2010b, An operational method for deriving standardised surface reflectance from Landsat TM and ETM+ imagery for Queensland and New South Wales. Proceedings of the 15th Australasian Remote Sensing and Photogrammetry Conference, Alice Springs, Australia.

15

Gill, T.K., Phinn, S.R., Armston, J.D., and Pailthorpe, B.A., 2009, Estimating tree foliage-cover change in Australia: challenges of using the MODIS 250m vegetation index product. International Journal of Remote Sensing, 30(6): 1547−1565. Hostert, P., Röder, A., and Hill, J., 2003, Coupling spectral unmixing and trend analysis for monitoring of long-term vegetation dynamics in Mediterranean rangelends. Remote Sensing of Environment, 87: 183−197. Huber, P. J., 1981, Robust Statistics, New York: John Wiley & Sons, Inc. Liang, S., Fang, H., Morisette, J.T., Chen, M., Shuey, C.J., Walthall, C.L., and Daughtry, C.S.T., 2002, Atmospheric correction of Landsat ETM+ land surface imagery – Part II: Validation and application. IEEE Transactions on Geoscience and Remote Sensing, 40: 2736−2746. Lu, H., Raupach, M.R., McVicar, T.R., and Barrett, D.J., 2003, Decomposition of vegetation cover into woody and herbaceous components using AVHRR NDVI time series. Remote Sensing of Environment, 86: 1−18. Lucas, R.M., Lee, A.C., Armston, J., Carreiras, J., Viergever, K. Bunting, P., Clewley, D., Moghaddam, M., Siqueira, P., and Woodhouse, I., 2010, Quantifying Carbon in Wooded Savannas: The Role of Active Sensors in Measurements of Structure and Biomass. In: Ecosystem Function in Savannas: Measurement and Modelling at Landscape to Global Scales, Eds. M.J. Hill and N.P. Hanan, Taylor and Francis (in press). Lucas, R., Cronin, N., Moghaddam, M., Lee, A., Armston, J., Bunting, P. and Witte, C., 2006, Integration of radar and Landsat-derived foliage projective cover for woody regrowth mapping, Queensland, Australia. Remote Sensing of Environment, 100(3): 388−406. Röder. A., Udelhoven, Th., Hill, J., del Barrio, G., and Tsiourlis, G., 2008, Trend analysis of Landsat-TM and -ETM+ imagery to monitor grazing impact in a rangeland ecosystem in Northern Greece. Remote Sensing of Environment, 112: 2863−2875. Roderick, M.L., Noble, I.R., and Cridland, S.W., 1999, Estimating woody and herbaceous vegetation cover from time series satellite observations. Global Ecology and Biogeography, 8: 501−508. Scarth, P., Röder, A., Schmidt, M. and Denham, R., 2010, Tracking grazing pressure and climate interaction – The role of Landsat fractional cover in time th series analysis. Proceedings of the 15 Australasian Remote Sensing and Photogrammetry Conference, Alice Springs, Australia. Scarth, P., Armston, J., and Danaher, T., 2008, On the relationship between crown cover, foliage projective cover and leaf area index. Proceedings of the 14th Australasian Remote Sensing and Photogrammetry Conference, Darwin, Australia. Schmidt, M. and Gillingham, S., 2008, Raster data analysis made simple with a scriptable open source framework: PyModeller. Proceedings of the 14th

16

Australasian Remote Sensing and Photogrammetry Conference, Darwin, Australia.
Specht, R. L., 1983, Foliage projective covers of overstory and understory strata of mature vegetation in Australia. Austral Ecology, 8: 433−439. Specht, R.L., and Specht, A., 1999, Australian Plant Communities: Dynamics of Structure, Growth and Biodiversity, Oxford University Press. Shaw, J.K., and Gillingham, S.S., 2006, Effects of cloud and smoke th contamination on woody vegetation time series trends. Proceedings of the 13 Australasian Remote Sensing and Photogrammetry Conference, Canberra, Australia. Sun, D., Hantiuk, R.J., and Neldner, V.J., 1997, Review of vegetation classification and mapping systems undertaken by major forested land management agencies in Australia. Australian Journal of Botany, 45: 929−948. Witte, C., van den Berg, D., Rowland, T., O’Donnell, T., Denham, R., Pitt, G. and Simpson, J., 2006, Mapping Land Use in Queensland – Technical Report on the 1999 Land Use Map for Queensland, Department of Natural Resources, Mines and Water, Brisbane.

17

Sign up to vote on this title
UsefulNot useful