Woody Vegetation Extent and Change 1972-2019

Western Australia – Whole State Product

2019 Update of the Land Monitor II Project

S. Furby

DATA 61 | CSIRO

December 2019

2

Woody Vegetation Extent and Change 1972-2019

Western Australia – Whole State Product This report briefly describes the image data files and the methodology that have been used to produce the vegetation monitoring product known as “Woody Extent and Change 1972- 2018” for the Land Monitor II Project. The data product is the Western Australian extent of the National Inventory Land Cover Change Program. This product is a series of classifications of the extent of perennial woody vegetation provided as twenty-eight 1-band raster files, one for each time period. The time periods are 1972, 1977, 1980, 1985, 1988, 1989, 1991, 1992, 1995, 1998, 2000, 2002, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016, 2017 and 2018. Also provided is change in the extent of perennial woody vegetation between each time period as twenty-seven 1-band raster files. The time intervals are 1972-1977, 1977-1980, 1980-1985, 1985-1988, 1988-1989, 1989-1991, 1991-1992, 1992-1995, 1995-1998, 1998-2000, 2000-2002, 2002-2004, 2004-2005, 2005-2006, 2006-2007, 2007-2008, 2008-2009, 2009-2010, 2010-2011, 2011-2012, 2012-2013, 2013-2014, 2014-2015, 2015-2016, 2016-2017, 2017-2018 and 2018-2019. This report summarises the methodology used to produce this product and its limitations. The file contents of the standard data products are described in section 6.

3

1. The Area

The area covered by the products is shown in figure 1. The region includes the whole of the state of Western Australia as well as adjacent areas of South Australia and the Northern Territory that fall within the 1:1,000,000 map sheet tiles processed.

Figure 1: Coverage of the woody vegetation products.

The grid represents the 1:1,000,000 map sheets tiles over the region. Products are formed by 1:1,000,000 map sheet tile. The extents of these map sheet tiles are as used by the Australian Surveying and Land Information Group (AUSLIG) for distribution of their 9 second digital elevation model (DEM) product.

2. The Satellite Image Data

The satellite data used to form these products was sourced from the United States of America (1972-1979) and Australian archives (1980 to the present). The 1972 (1972-1975), 1977 (1976-1979), 1980 and 1985 time periods are created from Landsat MSS imagery and 1988 to the present are created from Landsat TM/ETM+/OLI imagery. From 2000 to the present, complete coverage has been obtained. Prior to 2000 the treeless desert areas and treeless

4

grasslands of inland Australia were excluded from the scene selection search area. In descriptive terms these areas roughly comprise substantial parts of the following deserts: Great Sandy, Gibson, Great Victoria, Simpson and Tanami. It also includes parts of the Nullarbor Plain. The figures in Appendix 1 show the image data coverage for the time periods prior to 2000. From 1972 to 2002 optimal image dates were selected as those closest to January 1 that were as clear of cloud, smoke, noise and other problems as possible, except in 1989 where the preferred date was December 31. From 2004, the optimal dates are listed below. The division into northern and southern Australia is shown along with the map sheet names and boundaries in figure 2. The optimal image dates in northern Australia are:

1. August, September, October 2. July, June, November 3. May, April 4. December, January, February, March

The optimal image dates in southern Australia are:

1. January, February 2. December 3. March, April 4. November 5. May, October 6. June, July, August, September

Figure 2: Regions for optimal image date selection. The grid represents the 1:1,000,000

map sheets tiles over the region.

5

Individual image dates can be discovered from the vector and image date boundary information supplied as part of the ancillary data. The 1972 – 2013 data were ortho-rectified using the AUSLIG 9 second DEM and calibrated to the AGO Year 2000 base (Furby, 2002). The 2014 and 2015 data was supplied by Geosciences Australia already ortho-rectified and calibrated to on-ground reflectance (ARG25- version 1). The 2016 data was also supplied by Geosciences Australia already ortho-rectified and calibrated, but using different GCP chips (ARG25 – version 2). The 2017, 2018 and 2019 data was also supplied by Geosciences Australia already ortho-rectified and calibrated. The GCP chips used to create the ortho-rectified image are the same as for the 2016 update, but the calibration processing has been extended to include terrain-illumination correction. The methodology section describes how these disparate data sources were combined for this product. Final products are provided in datum GDA94, projection GEODETIC (latitude and longitude).

3. Methodology

The single-date woody cover classification methodology used for 2019 data in the 2019 Update to the National Inventory differs from that employed in the previous Updates. Automation is being progressively introduced to the data processing workflow with the ultimate goal of full automation. It has been necessary to adapt some of the steps in the workflow to better suit this goal. So far only the classification of the 2019 imagery has changed. However, as part of the ongoing continuous improvement of the National Inventory the accuracy of the woody extent and the split between the ‘forest’ and ‘sparse woody’ classes is being reviewed. Where necessary and feasible the woody base extent will be updated, and all historical years will be reprocessed through the ‘new’ methodology. This will also include adding additional years into the time series so that annual data from 1988 to the present is used. Map sheets SH50, SH51 and the adjacent areas of map sheets SG50, SG51, SH52, and SI51 will be reviewed and reprocessed in the first quarter of 2020. Revised products will be provided as soon as they are available. The National Inventory mapping uses the following class definitions:

Forest: Tree cover with >20% canopy cover with expected height at maturity >2m. Sparse Woody: Woody vegetation cover from 5-20% canopy cover. Non-woody: Any woody cover <5% canopy cover and all other vegetated and non-vegetated ground cover types.

The sparse woody cover class was introduced into the product during the 2006 update (Furby et al, 2007), but is mapped historically back to 1988. This cover type can only be reliably discriminated in the higher spatial resolution Landsat TM/ETM+/OLI imagery and so only appears as a class from the 1988 time period. It may also include more dense shrublands that do not meet the height requirement of the forest class. This creates the following change classes:

0 = no change 1 = non-woody to sparse

6

2 = non-woody to forest 3 = sparse to forest 4 = sparse to non-woody 5 = forest to non-woody 6 = forest to sparse

The steps used to produce the series of vegetation maps are listed below. This is the standard methodology used for the National Inventory product. 1972 -2013 (Furby, 2002 and Caccetta et al, 2012):

1) Co-register the images to a common map grid based on the 2000 Landsat ETM+ imagery. This allows ground sites to be traced through time and the satellite data to be compared with ancillary map data.

2) Calibrate the ortho-rectified images to a common radiometric base (2000 Landsat

7 ETM+ base).

3) Mask all areas of cloud/shadow and other data noise and irregularities and mosaic into four quadrants for each 1:1,000,000 map sheet tile. Mosaics are formed in MGA projection in zones 50 to 56.

4) Produce individual ‘forest probability’ maps from each date of the masked Landsat

imagery using the CSIRO index-threshold approach that produces measures of confidence of ‘forest’ and ‘non-forest’ vegetation at each date. Locally optimal indices are used within stratification zones. A CSIRO program was used to derive thresholds to match the probabilities within image date and stratification zone boundaries using the previous update forest probabilities as the base. Data values that fall between the thresholds were given a probability (0-100%), and scaled linearly with distance relative to the thresholds.

5) Produce individual ‘sparse woody probability’ maps from each date of the masked Landsat TM/ETM+/OLI imagery (1988 to the present) using the same index-threshold approach as for the forest class (Furby et al, 2007). An additional index based on the texture of the primary woodiness index from the forest mapping is included in this process. These probabilities are combined with the forest probabilities to form 3-band, 3-class probability images (forest, sparse and non-woody).

2014:

1) Compare the ortho-rectification of the 2014 scenes to the 2000 ortho-rectification base using the regular Master Check GCPs. Use the matched GCPS to derive a transformation for each path/row from the old ortho-rectification base to the new ARG25 base. A separate linear transformation for each path/row was found to be sufficient.

2) Use the transformations identified in 1) to ‘reproject’ the 1972 to 2013 3-class probability images maps to the new ARG25 ortho base.

3) Compare the calibration of the 2014 scenes to the 2000 calibration base using the

regular invariant targets. The target information was pooled across all path/rows in

7

a map sheet to derive a single set of gains and offsets for an approximate conversion from the 2000 calibration base to the ARG25 base for each map sheet.

4) Use the conversion identified in 3) to adjust the normalisation of the indices and provide starting threshold estimates for the ‘matching’ process to determine forest and sparse woody vegetation probabilities.

5) Mask all areas of cloud/shadow and other data noise and irregularities in the 2014

data and mosaic into the 1:1,000,000 map sheet tiles.

6) (Same as steps 4 and 5 above) Use a CSIRO program to derive thresholds to match the probabilities within image date and stratification zone boundaries using the previous update forest and sparse woody probabilities as the bases.

2015:

1) Mask all areas of cloud/shadow and other data noise and irregularities and mosaic into the 1:1,000,000 map sheet tiles. From 2015, all mosaics are produced in GEODETIC projection.

2) (Same as step 6 above) Use a CSIRO program to derive thresholds to match the probabilities within image date and stratification zone boundaries using the previous update forest and sparse woody probabilities as the bases.

2016:

1) Compare the ortho-rectification of the 2016 scenes to the 2015 AGDC v1 images

using the regular Master Check GCPs. Use the matched GCPS to derive a transformation for each path/row from the ARG25-version 1 base to the new ARG25-version 2 base. A separate linear transformation for each path/row was found to be sufficient.

2) Use the combined transformations identified in 1) and 2014 to ‘reproject’ the 1972 to 2013 ‘perennial vegetation probability’ maps to the new ARG25-version 2 ortho base.

3) Use the transformations identified in 1) to ‘reproject’ the 2014 and 2015 ‘perennial vegetation probability’ maps to the new ARG25-version 2 ortho base.

4) Mask all areas of cloud/shadow and other data noise and irregularities remaining in the automated composites of the 2016 data.

5) (Same as step 2 above) Use a CSIRO program to derive thresholds to match the probabilities within image date and stratification zone boundaries using the previous update forest and sparse woody probabilities as the bases.

2017 and 2018:

1) Mask all areas of cloud/shadow and other data noise and irregularities remaining in

the automated composites of the relevant image.

8

2) (Same as step 5 above) Use a CSIRO program to derive thresholds to match the probabilities within image date and stratification zone boundaries using the previous update forest and sparse woody probabilities as the bases.

2019:

1) Mask all areas of cloud/shadow and other data noise and irregularities remaining in

the automated composites of the relevant image.

2) Combine the ‘forest’ and ‘sparse’ classifications in a single 3-class classification

using a Random Forests classifier within stratification zone boundaries using the previous update forest and sparse woody probabilities as the bases (Furby, 2019).

All:

1) Process the sequence of 3-class woody cover probabilities from all dates (1972–

2018) using the “Conditional Probability Network” (CPN) approach (Caccetta 1997, Kiiveri and Caccetta 1998). This approach uses the probabilities from neighbouring dates to modify the probabilities of each pixel. The effect of the method is that it ‘smooths out’ sudden changes (e.g. from cultivation), and reduces uncertainty and errors in the individual dates. The result is a series of modified probability images for each date.

2) From the probability images produced above, ‘forest / sparse woody / non-woody

masks’ for each date are formed based on the class with maximum posterior probability from the multi-temporal (CPN) classification.

3) The ‘woody cover’ masks for consecutive time intervals are then compared to form

the vegetation change classes identified at the beginning of this section. The newly adopted Random Forests classifier has changed how decision boundaries are formed between the cover classes. Random Forests (RF) is a decision-tree based classifier, with many trees being derived and the results are drawn from averaging the results from the individual trees. Conceptually RF can be thought of a very large number of hyper-planes for class separation instead of simple multi-dimensional boxes of the index and threshold method used previously. It also allows estimation of probabilities for all three classes at the same time instead of step-wise and is readily extended to any number of classes of interest. More complex decision boundaries effectively reduce the uncertain spectral region and can increase the probabilities for previously less certain spectral signatures. As the 'sparse' class was always the one that was the most uncertain, this class is affected more than forest and non-woody. Generally, it is an increase in accuracy, so it’s a good thing from most perspectives, but it is not as consistent with the results from previous years as usual. RF also treats all three classes simultaneously using all available data (both spectral and texture). As a result, there is more finessing of the forest/sparse boundary with some shift of pixels between the forest and sparse classes. Again, this is not as consistent with the results from previous years as usual. The question of whether the accuracy is improved is less clear between these two classes.

9

In recognition of the different nature of the 2019 RF probabilities, the CPN parameters were adjusted to compensate for the observed shifts and produce a result that was more consistent with previous updates. The issue hasn’t been removed completely, but it has been significantly reduced. Ideally, the approach would be to reprocess the historical epochs using the new methodology rather than adjusting the CPN parameters, but that will only be performed as part of adjusting the woody base extent.

4. Accuracy and Limitations of the data

The perennial vegetation masks are derived from reflectance signals detected by Landsat TM, and depend on a contrast between woody vegetation and other cover types (soil, crops, bare rocks etc). The thresholds for classification of forest and sparse woody have been derived from interpretation and comparison with aerial photography and high resolution (1m) satellite image samples. This classification as ‘perennial vegetation’ relies on the spectral contrasts of cover types resulting from physical differences on the ground, and effectively requires a certain density of vegetation. Hence thin, scattered vegetation with a high proportion of bright soil background may be omitted. Certain dense but highly reflective vegetation types may also be omitted. In particular, bare or very thin areas within bush remnants will not be classified as ‘perennial woody vegetation cover’. Common examples are tracks, rocks, fire-scars, and salt-affected vegetation. Hence the areas mapped as vegetation at particular dates will not necessarily correspond to administrative definitions of reserves etc. There is a time lag in detection of re-vegetated areas, which varies with region and vegetation type. Re-vegetated areas will not be mapped until the vegetation achieves a sufficient density. Hence some recent, slow-growing, or sparse re-vegetated areas will not be detected. Errors of commission may occur when other land covers give a similar spectral response to perennial vegetation. The temporal smoothing of the CPN removes most of the transient cultivation effects that might cause these errors. However, there are cases where errors of commission may remain after the CPN processing. Examples include cleared areas with persistent dark soil, and in also lake fringes and normally dry lake surface where changes in water level have dramatic effects on the cover and reflectance. Errors that are incurred in these areas may result in incorrect mapping of change in lake edge vegetation. However, these errors have not been removed by manual digitising as some may be real vegetation change.

5. Limitations and Liabilities

The information contained in these vegetation maps is necessarily based in part upon various assumptions and predictions. The Land Monitor II Project (comprising the Western Australian State Government agencies, Western Australian Land Information Authority (Landgate), Department of Water and Environmental Regulation, Department of Primary Industries and Regional Development, Department of Biodiversity, Conservation and Attractions, Department of Lands, Planning and Heritage, Department of Fire and

10

Emergency Services, the Water Corporation of WA, and the Commonwealth agency CSIRO Data61) accepts no responsibility for any inaccuracies in these vegetation maps, and persons relying on these maps do so at their own risk.

6. Products Contents: Woody Extent and Change

All geo-referenced files are provided in map projection GEODETIC (latitude / longitude) with datum GDA94. The pixel resolution is 0.00025 decimal degrees. Raster Files are provided in ERMapper raster format with standard header files (.ers). For reasons of convenience, data products are provided by 1:1,000,000 map sheet units. The map sheets provided are SD52, SE51 and SE52 (northern dry season image date preference) and SF50, SF51, SF52, SG50, SG51, SG52, SH50, SH51, SH52, SI50 and SI51 (southern dry season date preferences). The contents for each map sheet are: 1) Woody vegetation extent files:

prefix_(1988, …, 2019)_NCAS_mapID_woody_geo.ers [ERMapper header file] prefix_(1988, …, 2019)_NCAS_mapID_woody_geo.hdr [ARC header file] prefix_(1988, …, 2018)_NCAS_mapID_woody_geo.bil [BIL format image file] Twenty-eight 1-band raster files for each region, each image corresponding to one of the time periods between 1972 and 2019. CLASSIFICATION CODES:

0 = not woody 1 = sparse woody cover 2 = forest cover

The ‘mapID’ in each file name is the 4 character 1:1,000,000 map sheet identifier as in the list above and shown spatially in figure 2 in section 2 of this report. The ‘prefix’ before the image year in each file name describes the image data source as follows:

General Description

Values Value Description

l1,2,3,4,5,7,8 Satellite name and number (note l is lower case L)

Where the constituent images were acquired by a mixture of satellites, the predominant satellite number is used

at, am, em Source of data and type

at am em

Australia (GA) Landsat TM/ETM+/OLI Australia (GA) Landsat MSS USGS (EROS) Landsat MSS

m, t, p, s Sensor type m t

multi-spectral (no terrain correction) terrain corrected multi-spectral

11

2) Woody vegetation change files:

(1972-1977,…)_NCAS_mapID_chnge_geo.ers [ERMapper header file] (1972-1977,…)_NCAS_mapID_chnge_geo.hdr [ARC header file] (1972-1977,…)_NCAS_mapID_chnge_geo.bil [BIL format image file] Twenty-seven 1-band raster files for each region, each image corresponding to one of the time period intervals between 1972 and 2019. CLASSIFICATION CODES:

0 = no change 1 = non-woody to sparse 2 = non-woody to forest 3 = sparse to forest 4 = sparse to non-woody 5 = forest to non-woody 6 = forest to sparse

3) Combined ‘Ever Woody’ mask:

mapID_1990-2019_everwoody_mask_geo.ers [ERMapper header file] mapID_1990-2019_everwoody_mask_geo.bil [BIL format image file] mapID_1990-2019_everwoody_mask_geo.hdr [ARC header file]

This image product combines the outputs from 1990 to 2019 to identify areas that have ever been woody – forest or sparse - (or conversely never been woody). The data for the initial MSS years is considered less confident and so has not been included in the processing. This mask is used in the display algorithm. CLASSIFICATION CODES:

0 = woody (forest or sparse) in at least one time period 1 = never woody

4) ERMapper algorithm:

mapID_VegChange_1990_2019.alg This algorithm is based on the virtual dataset mapID_woody_1972-2019_mga.ers. Each virtual dataset points to the twenty-eight woody extent files in the corresponding map sheet. The algorithm is used to produce colour displays of woody vegetation and change for the period 1990-2019. It can be readily modified by ERMapper users to alter the dates for display: The colours in the display are: Green: Woody (forest or sparse) in both 1990 and 2019 Cream: Non-woody in both 1990 and 2019 Red: Woody (forest or sparse) in 1990 and non-woody in 2019 Orange: Forest in 1990 and sparse in 2019 Dark blue: Non-woody in 1990 and woody (forest or sparse) in 2019 Light blue: Sparse in 1990 and forest in 2019.

12

5) Ancillary Files: cloud_masks [Directory containing cloud masks for image mosaics created during the 2019 Update] date_boundaries [Directory containing individual image date boundaries for each

time period, organised by map sheet] Cloud mask vector files contain the polygons which show the boundaries of masked cloud and noise in each time period. Only those cloud masks created during the 2019 Update are provided (2018 and 2019 for all map sheets and 1988, 2007, 2008, 2009, 2011, 2012, 2013, 2014, 2015, 2016 and 2017 for map sheets SF50, SF51, SF52, SG51 and SG52). The vector files are named as follows:

prefix_ (1988,…,2019)_mapID_mskc.erv [ERMapper vector header file] prefix_ (1988,…,2019)_mapID_mskc [ERMapper vector file]

Image date boundary information is provided in vector form for 1972 to 2005 and in both vector and raster image form for 1988 and 2006 to 2018 and raster only form for 2019 for most map sheets. These data have been created over a 19-year program and file naming and attribution conventions have evolved over this time. Vector files (in MGA projection prior to 2015) contain polygons showing the extent of particular image dates in the map sheet mosaic. The day, month and year of the imagery are mandatory parts of the vector attribute. The attributes for some years may also contain path/row information. The form of the file names is:

prefix_ (1972,…,2018)_mapID_vdatebnd.erv [ERMapper vector header file] prefix_ (1972,…,2018)_mapID_vdatebnd [ERMapper vector file]

The final ‘vdatebnd’ descriptor may or may not be present, or it may be replaced by ‘local’ or ‘extatt’. The raster image date extent images give the day and month for each image pixel in ‘ddmm’ format. The year is implied from the image file name, noting that for ‘south’ images, imagery from November of December will be from the year before that indicated in the image file name. The images for some years may also contain a second image band with path/row information in ‘ppprr’ format. As with the vector files, images will be in MGA projection prior to 2015. The form of the file names is:

prefix_ (2006,…,2019)_level_mapID_rdatext.erv [ERMapper image header file] prefix_ (1972,…,2019)_level_mapID_rdatext [ERMapper image file]

The ‘level’ descriptor refers to the processing level of the source imagery. ‘AGO’ typically refers to the 2000 ortho-rectification and calibration base. ‘NBAR’ refers to AGDC version 1 calibration and ortho-base. ‘NBAR2’ refers to the revised geometric base and can be assumed to imply that terrain-illumination correction was also applied to the data. All these data are in the ‘Ancillary’ directory.

7. Documentation:

LM2019_WoodyExtent_Change_1972_2019.pdf [this report]

13

References

Caccetta P.A. (1997), Remote Sensing, Geographic Information Systems (GIS) and Knowledge-Based Methods for Monitoring Land Condition, PhD Thesis, School of Computing, Curtin University of Technology. Kiiveri, H.T. and Caccetta, P.A. (1998), Image fusion with conditional probability networks for monitoring salinisation of farmland, Digital Signal Processing, October, 8:4, 225-230. Caccetta, P., Furby, S., Richards, G., Wallace, J., Waterworth, R., Wu, X. Long-term monitoring of australian land cover change using landsat data: Development, implementation, and operation. In Global Forest Monitoring from Earth Observation; Achard, F., Hansen, M.C., Eds.; CRC Press: Boca Raton, FL, USA, 2012; pp. 243–258. ISBN: 9781466552012. Furby, S. (2002). Land Cover Change: Specification for Remote Sensing Analysis, National Carbon Accounting System, Technical Report No. 9. Furby, S.L, Wallace, J.F. and Caccetta, P.A., (2007). Monitoring Sparse Perennial Vegetation Cover over Australia using Sequences of Landsat Imagery, 6th International Conference on Environmental Informatics, Bangkok, Thailand, November 2007. Furby, S. (2019). Methodology for the Forest Extent and Change Mapping for the National Inventory: 2019 Update. Technical Report for the Department of Environment and Energy (available early 2020).

Acknowledgements

We would like to thank the Land Monitor II Partner Agencies for support of the Land Monitor II Project. This work has been supported by the Australian Government Department of Environment and Energy, through management and coordination of the National Inventory Land Cover Change Project, details of which can be found at http://www.environment.gov.au/climate-change/climate-science-data/greenhouse-gas-measurement/.

14

Appendix 1: Landsat Image Extents 1972 to 1998

15

16

17

18