Space data was the only data type that met the requirements for conducting this assessment within the resources available; as argued earlier in The View From Space .
A commercially available, geocoded digital mosaic of Landsat MSS data for the entire continent formed the primary dataset for quantifying the location, severity and extent of landcover disturbance. The strategy and associated methodology are now described.
The principal dimensions of this dataset were two - spectral and spatial. There was no temporal dimension. This dataset was taken as a time slice of the continental landcover conditions as of approximately 1991. No change detection was applied to images of different dates. The only time dimension in this analysis was an implied one: all (significant) change detected using the 1991 dataset was assumed to have occurred in the last 200 years. Exactly when this change occurred cannot be directly inferred from this satellite dataset alone.
The project set out to answer one question - what is the current state of the continental landcover?
This we hold, makes a small but significant contribution to a national understanding of the severity and location of biotic erosion. Another significant contribution will come from quantifying the dynamics of landcover change; the rate and location of habitat disturbance across the continent. There are limits to this second question. The high spatial and temporal resolution of landcover change is only possible using satellite data since 1972; eg see Graetz et al (1992), Sivertsen (1994), Wilson et al (1993). The rates of clearing can be determined further into the past using archived aerial photography and surveyor's maps; eg Hobbs et al (1993).
The spatial dimension of this dataset was determined by its resolution; ie by the size of the ultimate unit of data capture and representation, or pixel. The pixel size of the supplied dataset was 0.001 x 0.001°, or approximately 1 hectare. Thus the continental aggregate represented a very large data volume to process. What is more important, this fine spatial resolution was far, far greater than that of the Natural Vegetation dataset. The latter was published at a scale of 1:5,000,000 and appears to have an intrinsic resolution of, at best, 1 km2, or more realistically 25 km2. Thus, this basic dataset was 2500 times more coarse than the Landsat source dataset. Therefore, the Landsat MSS dataset was resampled to 0.005 x 0.005° (˜500 x 500 m or 25 ha) for the analysis and further resampled to exactly 1 x 1 km for reporting.
The de-focusing of the satellite dataset may, at a superficial glance, seem to reduce the acceptance of the findings. Hectare-sized patches of landscape are human-scaled and easily experienced. An area twenty-five times larger is less easily imagined in the mind's eye, and therefore commands less validity. Nonetheless, the use of coarse (25 ha) rather than a finer-grained (1 ha) pixel size introduces insignificant error in the overall assessment.
The 1: 1, 000, 000 scale map sheet BRISBANE covering the city of Brisbane, the nearby coast and inland; total area approximately 120, 000 km2. Even with a pixel size of 25 ha, there is an abundance of detail in the image. Fine details of the dunes and sand blowouts on Fraser Island are detectable, as are subtle and obvious patterns within the landscape as a whole. Overall, a visual appreciation of the extent of clearing and fragmentation of the original vegetation is easily obtained. If you can do this, then you have instinctively learned to assimilate both the spectral (colour) and spatial (grain) dimensions of satellite images. The computer techniques used in this project imitate the analysis and assessment that you have just made.
The final dimension of the satellite dataset - the spectral dimension - was the most critical because it determines what information can be derived. MSS data has four spectral channels, traditionally labelled from #4 to #7, each giving four values for the reflected light from each pixel regardless of its size. For most landcover types, these four wavebands are highly correlated and can be reduced to two: a measure of cover and of greenness. Cover includes both the cover of dark foliage over the brighter soils and the shadows within a canopy. Greenness is a measure of the abundance of young and vigorous leaves within the canopy, be they trees, shrubs or the herbaceous understorey.
Under dry conditions, ie with no preceding rainfall, the cover of live vegetation is linearly but inversely related to the brightness of one channel, the reflected red radiation recorded as MSS #5. In this waveband, the intact woody canopy is dark whereas that of the replacement crop, pasture, or even bare soil, is bright. In addition, for the interpreter digitally interacting with this image, there are both spectral and spatial clues to what is cleared, and what is uncleared woody vegetation. It is remarkably easy for a human to correctly interpret a complex image. It is considerably more difficult to achieve the same result using a computer. The concepts discussed above are illustrated using an image of the Armidale area of eastern Australia.
The 1: 1, 000, 000 scale map sheet ARMIDALE; total area approximately 120, 000 km2. Setting aside the unnatural or false colours of the image, just using the clues of tone (light to dark) and the spatial pattern or texture within the image makes a rapid interpretation of what is undisturbed natural vegetation, and what has been cleared for pastures and crops. Much of the remnant natural vegetation is restricted to areas of significant topographic relief: here in the ILZ, relief is the most frequent protection against future clearing. The intensity of the red colour indicates the level of greenness in the landcover. In this scene, both remnant natural vegetation and some of the pastures were green on the ground and were thus portrayed as red in this image. The several smoke plumes and their location indicate the continuing nature of landcover disturbance.
All spectral channels of the satellite dataset were used to understand and interpret the images for all vegetation types. For the Intensive Landuse Zone, the objective of the assessment was to allocate each pixel into one of five disturbance classes; uncleared, thinned, cleared, indeterminate, and other.
The uncleared class was characterised by having an intact canopy; this being determined relative to all the remnants of any landcover type. The uncleared class may be disturbed by grazing but within the ILZ, this was disturbance that could not be separated from clearing.
For the landcover types with an open and sparse canopy, ie projected foliage cover (pfc) < 30%, it was difficult to separate clearing, the loss of the overstorey, from the effects of acute overgrazing, the loss of the understorey. This difficulty was one of the two principal sources of error; see below.
The cleared class was characterised by the effective replacement of the canopy with a herbaceous layer of crop or pasture plants. Here the overstorey cover was assessed to be reduced to < 5%; ie scattered or no trees. The thinned class was determined to be an intermediate class of tree cover between uncleared and cleared. It was included to increase understanding of the severity of clearing. It is our contention that, in terms of biotic erosion, thinned and cleared are effectively equivalent.
The indeterminate class was assigned by the analyst to two situations. The most common was where an allocation to clearing was not ecologically meaningful; eg for a naturally occurring grassland. More rarely, it was used to indicate that isolated clearing could be detected by the interpreter but not consistently detected by computer analysis. This occurred principally in the Northern Territory where what little clearing has occurred was scattered within a very seasonally dynamic landscape. The other class was reserved for lakes mapped within the Natural Vegetation dataset.
Overall the procedure was straightforward. The allocation decision rules were based on the expert judgement of the analyst (Graetz) after intensive examination of the satellite data in digital format. Decisions were made only after a visual reconnaissance of the scene and a consultation of ancillary data (maps, etc.).
After prolonged exploration of the image data for any one vegetation type, the analyst selected MSS #5 values that separated uncleared from thinned, and thinned from cleared. Those values and this channel alone were used to allocate each pixel into one of these three disturbance classes. Since the Landsat MSS #5 data were continuous in nature with no obvious thresholds, the choice of MSS #5 values varied by landcover type and by climatic region. Thus expert judgement was critical in assigning values to classes. The allocation process was then computerised.
We reiterate the point that even at the finest spatial resolution of 0.001 x 0.001°, a pixel contains a mixture of landscape elements - tree canopies, shadows, understorey and so forth. This was particularly so with the pixel size used here, 0.005 x 0.005°, where the full spectra of mixtures of overstorey and understorey are possible. The category thinned was created to capture the most uncleared part of this spectrum where the canopy was approximately half that of the uncleared state. The category of cleared will also comprise spectrum mixtures from totally treeless to sparse scattered trees. The broad classification was adequate for the focus of this project; ie quantifying that area of landcover that was uncleared.
Because of the nature of the number of decision classes and artefacts within the MSS dataset, there are uncertainties associated with the assessment values derived. These errors are most usefully discussed in relation to the uncleared category as being either errors of commission, where the true value for the uncleared category is inflated, or omission, where actual value for this category is diminished. Both error types were unavoidably included in the analyses reported here.
A large and consistently occurring error of commission results from the cleared grassy landscapes often being green and actively growing at the time of image acquisition. This tends to darken them towards the signature of cleared woody canopies with the result that a proportion of the cleared class was allocated to uncleared. One example of this error of commission is in Armidale image where green actively growing pasture (reddish in colour) occurs in a N-S strip just to the west of the Great Dividing Range. This landcover is obviously non-woody canopy, being quite distinct to the eye from the neighbouring forested upland remnants. Nonetheless, it was impossible to set decision rules that consistently separated this green pasture from the surrounding forest remnants because those to the east were also green and actively growing (reddish in colour) while those remnants to the west were not.
This source of uncertainty was inherent in the dataset and intensive and systematic efforts to remove it failed. Moreover, careful selection of decision rules to minimise this error lead to the converse omission error in the drier, less dense canopies. Overall, a strategy was followed to minimise the size of the error and to stabilise its sign; ie consistently to one of commission or omission. The decision rules for the uncleared category were always set not only to achieve the smallest uncertainty possible in the situation, but also to keep the error consistently one of commission. Furthermore, the emphasis was on getting the values correct for the more mesic landcover types, the forests and woodlands. This strategy was hampered by significant variation within the dataset that was not related to the surface. The net result was that there were both errors of commission and omission in the results reported here. The difficulty is illustrated by the Albany image.
The south-west of Western Australia. This area experiences both high rainfall and cloud cover almost all the year. As a consequence, the cleared areas of crops and pastures, are usually green. Under these conditions, it was far more difficult to separate uncleared from cleared for landcovers near the coast than it was in the drier inland. Significant errors of commission and omission were possible with small changes in the decision rules set for this scene that is a spectacular example of loss and fragmentation of natural vegetation.
The objective for the Extensive Landuse Zone of the continent was to use satellite data to allocate part or whole of each vegetation type to one of the following disturbance categories: slight, substantial, significant, indeterminate, and other.
The allocation was to reflect the contemporary level of disturbance imposed by the two principal agents of landuse; grazing by domestic stock, and to a much smaller extent, repeated burning. The impact of grazing is by far the most influential in generating both systemic and cumulative landcover changes.
The temporary and cumulative impacts of grazing on the landcover can be detected and interpreted using satellite data for most vegetation types. Similarly, even though the dataset used was uni-temporal, some appreciation of the nature of the fire regimes can be interpreted, but this was far less certain than that for grazing.
The critical difference between the assessment of grazing and burning, and that for clearing, was that the former cannot be implemented by computer. The assessment of the disturbance level can only be based on expert decision and opinion; a decision that was reached only after detailed interpretation of the spectral and spatial patterns within the image, and the integration of ancillary data.
The subjective component of this assessment was thus higher than that involved in the assessment of clearing. To maintain a high level of transparency, the source document contains example images that explain and support the decision taken for each vegetation type.
The nature of the evidence and its interpretation is exemplified by the Rawlinna image.
The Nullarbor Plain near Rawlinna, Western Australia, pastorally occupied and a part of the ELZ. The predominant landcover type is a low open chenopod shrubland - bluebush and saltbush country - with sparse mulga inland. The landcover consequences of pastoralism are quite obvious from 900 km in space and with a coarse ground resolution of 25 ha. The landcover change within whole paddocks and around watering points can easily be detected and assessed by eye. The level of disturbance here ranges from substantial to significant. However, when this relatively small area of this landcover type was considered along with the all the others not within this scene, the disturbance level was assessed overall as substantial.
Where natural grasslands (xG) were designated by the Natural Vegetation dataset within the ILZ, they were allocated to the indeterminate class for the clearing assessment and manually interpreted for the ELZ.
Finally, there is no category for undisturbed. There is no vegetation type that was absolutely undisturbed by the direct (landuse) or indirect (feral animal) effects of human activity. Therefore we used a relative scale: by definition landcover types that were designated unused or protected by tenure were used as benchmarks for the assessment and were always set to disturbance category 1, slight.
Furthermore, the disturbance rating that was finally assigned took into account local variations in disturbance level, such as shown in the Rawlinna image above, as well as the condition of the landcover type that occurred within the ILZ but was not actually cleared.
At this point, the reader may be less than convinced that all of this interpretation and assessment of landcover disturbance is possible using satellite data.
It is not easy to illustrate this without drowning the argument with images. Therefore we have included a set of images covering the continent in a section called Exploring Your Country. Each image covers one sheet of the excellent 1:1, 000, 000 map sheet series and are named accordingly. We recommend that you explore your country using each of these images (5° x 6° in size) and try to interpret the type and level of landcover disturbance that you can detect.
For the ILZ, the interesting areas are in the south-western corner of Western Australia, New South Wales and Queensland. For the ELZ, the interesting areas are South Australia, the Northern Territory, Queensland and Western Australia.
If you have a particular interest in the state of the national forests, then examine New South Wales, Victoria and Tasmania.
