Assessment of Change in Remnant Vegetation Area and Condition

Detecting and Monitoring Changes in Land Condition Through Time using Remotely Sensed Data

Introduction

This report summarises the work and findings of one component of the LWRRDC-funded project 'Detecting and Monitoring Changes in Land Condition Through Time Using Remotely Sensed Data'. The project is funded by the Land and Water Resources Research and Development Corporation, project CDM1.

The results show that satellite data can provide indicators of change in vegetation condition which can be applied over broad areas. The study shows that the optimal indicators vary with vegetation type, and that the season of the imagery is important. The trends through time can be represented as maps in ways which will separate gradual long-term change from the short-term impacts such as those from fire. Graphical plots of trends through time can be used to compare the response of sites of particular interest. Examples of these output products are contained in this report. The implications for operational application are discussed in Section 6 of this report.

1. Background

The clearing of native vegetation in agricultural regions has resulted in a decline in both production and conservation values. The maintenance and improvement of remnant vegetation is important for its role in water use and maintenance of water quality. Grazing may have a major impact on the density, composition and long-term viability of vegetation remnants. Various schemes have encouraged the fencing of remnants to control grazing. Other negative impacts arise from weed invasion, frequent fires, salinity due to rising saline ground waters, logging, insect damage, rubbish and vermin.

Effective assessment, planning and management of remnant vegetation depends on a sound knowledge of the distribution and condition of vegetation types at local and regional scales, and a capacity to detect changes over time. Current mapping methods are costly and time- consuming, and there is no effective method for monitoring change in condition on a broad scale. Ground-based assessments of condition at a particular time are rare, and exist only for sites within areas of special interest. Thus there is an incomplete knowledge of the area and vegetation resource in remnants in agricultural areas, and no effective broad-scale system for providing information on the vegetation condition, or the change in condition through time in response to management practices.

Satellite imagery provides repeated regional coverage and is the only available data which may provide this broad-scale information on condition and change. The project aimed to use these data to develop and demonstrate methods for mapping remnant vegetation area, type and condition; and particularly for mapping indicators of change in condition through time.

Mapping of vegetation types from remote sensing has been attempted and reported in a range of environments from forests to rangelands around the world. In particular, this project builds on earlier work in mapping vegetation types in the WA wheatbelt using Landsat MSS data (Hobbs et al 1989). The work demonstrated that broad structural types could be mapped, and that use of data from two seasons was necessary to discriminate between certain vegetation types. A similar study using Landsat Thematic Mapper data has shown improved mapping of good-condition floristic associations in the Kellerberrin area. The work is summarised in a separate report (Lambeck and Wallace 1993). One finding of their work was that a large proportion of the remnant bushland in the area has spectral signatures which are different from those of known pristine or good-condition vegetation. Ground assessment has confirmed that the majority of these areas have been affected by severe disturbance and are in poor condition, or are areas comprised largely of bare rock. These findings indicate that major disturbance in condition does affect the spectral response of vegetation.

The aims of this study were to investigate the differences in satellite spectral data which are associated with differences in vegetation condition, and to examine changes in the spectral signature over time which may indicate change in condition. The results can be applied on a broad scale, but it is noted that the spatial resolution of the data limits the kind of information which can be derived. The satellites are measuring an integrated response at the pixel scale (approximately 0.1 ha for Landsat Thematic Mapper TM). The signal will be affected by major changes in structure, composition, density and exposed soil. Satellite data will not detect minor changes in the presence or diversity of species within a vegetation class.

2. Study areas and ground information:

The work was conducted in two study areas within the southwest agricultural region of Western Australia:

Agricultural land use in the Upper Kent Catchment is predominantly grazing. The area has been identified by LWRRDC as a focal catchment for on-going studies. The interest in clearing and vegetation quality in the catchment arose from recognition of increasing salinity of the Kent River, which had been gazetted as a potential water supply by the Water Authority of WA (WAWA). Clearing controls, administered by WAWA, were brought into effect in the late 1970's. WAWA had also commissioned a survey into the quality of remnant vegetation in the Kent (True et al 1992).

The Kellerberrin area has been the focus of long-term study by the CSIRO Division of Wildlife and Ecology (DWE). Excessive clearing (over 90% of the area) has led to fragmentation of the native vegetation to isolated small remnants. DWE is studying the ecological dynamics of this highly-disturbed system and attempting to define strategies for conservation and revegetation. Salinisation due to rising water tables in the area has also focussed attention on the role of remnant vegetation in water use.

The Upper Kent study area examined in this project is 90 km by 60 km (approximately 540,000 ha), which includes all of the Upper Kent catchment.

Landsat TM image of the Upper Kent Catchment area, April 1993. TM bands 3, 5 and 4 are displayed in blue, green and red. The area shown here is approximately 60km by 40km.

The Kellerberrin study area is 100 km by 90 km (approximately 900,000 ha) and includes the whole region covered by the CSIRO Division of Wildlife and Ecology (DWE) study.

Landsat TM image of the Kellerberrin study area. January, 1993. TM bands 1, 2, and 3 in blue, green and red. The area shown is approximately 60km by 40km

3. Landsat and ancillary data

The project has assembled and analysed sequences of Landsat TM data over the two areas. Summer and spring images were selected subject to cloud cover and availability. Landsat data are acquired every 16 days and an archive of TM data from 1988 is held at ACRES. Some images are available from 1986 and 1987. Earlier MSS images were used to provide long-term clearing histories. The earliest available images were used for this purpose, but no attempt was made to acquire a comprehensive sequence of MSS data. These early MSS images have also been used to look at longer term condition trends.

MSS	19	September	1977
MSS	4	March	1979
TM		September	1986
TM	13	February	1988
TM	17	April	1988
TM	19	May	1988
TM	7	August	1988
TM	19	March	1989
TM	10	August	1989
TM	14	September	1990
TM	1	September	1991
TM	24	February	1992
TM	15	April	1993
TM	22	September	1993

A number of these images were partially cloud-covered over the study area. Summer-spring pairs of scenes from 1988, 1989 and 1993 were cloud-free.

MSS	31	December	1973
TM	6	November	1986
TM	8	October	1987
TM	28	January	1988
TM	23	August	1988
TM	14	September	1990
TM	19	December	1990
TM	5	February	1991
TM	21	February	1991
TM	9	March	1991
TM	25	March	1991
TM	21	October	1992
TM	9	January	1993
TM	22	September	1993

Terrain height data were acquired as contours from the Department of Land Administration (DOLA). These data were gridded to produce a digital terrain model over the two study areas. Associations of vegetation type and terrain were examined qualitatively in this study, but the results were inconclusive and are not reported here.

Ground knowledge of the condition of specific sites was available in both areas and was used to direct the analysis of the satellite image sequences. There were few objective ground records of changes in condition over time, though sequences of aerial photographs exist for both areas. A comprehensive vegetation map was not available for either area.

4. Methodology

A brief overview of the methodology used is described here. Separate technical reports describe the image processing and analyses in more detail.

The related study on the mapping of healthy vegetation types in the Kellerberrin area has been reported elsewhere (Lambeck and Wallace 1993).

5. Results

This section briefly describes the results of the analyses which examined the differences in condition in the spectral data and the derivation of indices related to condition. Examples are presented of the resulting maps of condition and of change in condition over time. The analyses are reported for different vegetation types, as the overall analyses showed that the condition indices and the range of response varied with vegetation type. Examples of clearing history maps are also presented.

5.1 Remnant area estimates and clearing histories

For both study areas, a classification of the 1993 TM data was used to provide a map of existing remnant vegetation, and to estimate the area of remaining native vegetation. Clearing history maps and estimates were produced from this classification of the 1993 data and of the earliest available MSS data.

The following figures show examples of clearing maps from the study areas -- over the periods 1973-1993 for Kellerberrin and 1977-1993 for the Kent River. Portions of both study areas are not covered by the early Landsat MSS scenes. In the Kellerberrin study area (100 km by 90 km), the coverage of remnant vegetation in 1993 was estimated at 7.7%. Coverage in 1973 was 8.7%. The areas of clearing are indicated in orange.

Kellerberrin area. Classification map of remnant native vegetation and clearing from 1973 to 1993. Green areas are classified as remnant vegetation in 1993; orange areas have been cleared since 1973.

The area of bush remnants in the Upper Kent Catchment in 1993 was estimated at approximately 21%, excluding the state forest area in the south-west. The area cleared since 1977 represents approximately 7% of the catchment. Outside the catchment within the larger study area, approximately 7% has been cleared since 1977. Maps showing the spatial distribution of vegetation and clearing in the two areas have been produced and supplied to WAWA, DAWA, the Department of Conservation and Land Management (CALM) and DWE.

Kent Catchment area. Classification of clearing from 1977 to 1993. Orange areas are classified as cleared during the period. Remnant vegetation and forest plantations appear dark in the image.

5.2 Condition mapping within vegetation types - Kellerberrin

From the ground information supplied by DWE, three broad vegetation classes were defined:

Initial discriminant analyses showed that the spectral signatures of these classes are different; the condition analyses were conducted separately on these vegetation types. Figure 6 illustrates the spectral separability of these classes. It confirms that good condition sites from each vegetation type are spectrally distinct and illustrates the spectral similarity of some poor condition sites of different vegetation types.

Ordination plot illustrating the spectral separation of Kellerberrin sample sites and groups of sites based on vegetation type and condition. The ordination plot is produced from the analysis of 1993 TM data from 2 dates in summer and spring.

5.2.1 Wandoo and mallee woodland

The discriminant analysis using TM data from January and September 1993 showed clear separation of good condition sites from very poor and grazed sites. Site ordination using all six spectral bands from both dates demonstrated that a one-dimensional condition index could be defined. Band reduction procedures were applied to determine the important spectral bands and to produce a simplified condition index that would be robust across different years.

The important spectral bands were found to be combinations of TM bands 4, 5 and 7 (near and shortwave infra-red) from the summer image with either band 4 or band 5 from the spring image. A weighted sum of the three summer bands and spring band 5, essentially an infrared brightness index, was found to maintain a high proportion of the discrimination between good and poor condition sites over the three spring/summer pairs from 1988, 1990/91 and 1993. Increases in this brightness index are associated with decline in condition. Use of the summer image alone (bands 4+5+7) provides 90% of the separation of this two-season index.

5.2.2 Heath

The discriminant analysis using the same 1993 two-season data showed clear separation of good condition heath sites from poor condition, grazed sites. Band reduction showed that summer is the dominant date for site separation. Bands 4, 5 and 7 from the summer image are important in measuring heath condition. Band 2, from the visible part of the spectrum, is also important. Unlike the woodland sites, the best condition index is based on contrasts between bands 2 and 4 and bands 5 and 7 rather than a sum. Good condition open heath sites are similar to poor condition heath sites using any brightness-based index. The spring image was found to be less important in discriminating between heath sites. Figure 7 is a colour display of the condition index for a portion of the study area.

Colour display of the condition index for heath class in 1993. Green: good condition; Blue: intermediate; Red: poor condition. Non-bush areas have been masked and appear as black. Detail of the Kellerberrin study area, approximately 20km by 10km.

5.2.3 York gum and Jam dominated woodland

Discriminant analysis using the 1993 data showed very little separation of these sites based on condition. This vegetation type is associated exclusively with very rocky terrain. Spectra for image pixels within these sites are a mixture of vegetation and rock signals. The amount or type of rock in the signal is not related to vegetation condition. The analysis suggests that it is not possible to produce a condition map at a particular date for this vegetation type, due to this spatial variation. It is possible however that major changes through time at a location will be detectable, although ground information was not available to verify this.

5.3 Change in condition through time within vegetation types - Kellerberrin

The condition indices above were calculated for spring and summer TM image pairs from different years and the values compared to detect changes in the condition of the vegetation. Figure 8 shows the change between 1988 and 1993 in the heath condition index. Red areas have declined over the five-year time span and include areas affected by fire in the early 1990's. Green indicates areas that have improved, perhaps through reduced grazing pressure, or recovery from fire. Blue areas have been relatively stable over the five-year time span.

Colour display of the change from 1988 to 1993 in the heath condition index. Green: improving condition; Blue: stable; Red: declining condition; Black: non-bush areas. Detail of the Kellerberrin study area, approximately 20km by 10km.

Figure 9 (below) shows plots through time of the condition indices for sample sites within the study area. Condition estimates for 1989 and 1992 could not be obtained due to cloud cover at the times of satellite overpasses. The higher, or brighter, the condition value in the plot, the poorer the condition of the site. A horizontal line, such as for the good condition site at the bottom of the plot, indicates a site that has not changed over the five-year interval. Sites with curves that decrease have improved. The red curves indicate sites that are in poor condition in 1993. The increase in these curves indicates that these sites are declining. Possible causes in these cases are fire, clearing and grazing.

Condition index plotted against time for 1988-1993 for sample woodland sites. Increase in the index value is associated with a decline in condition.

5.4 Long term change 1973-1993

Except for fire, major changes in vegetation structure, composition and density typically happen slowly. The 1973 MSS image has been used to look at trends in condition over a twenty-year interval for the Kellerberrin area.

MSS images do not have the shortwave infra-red spectral bands equivalent to TM bands 5 and 7. Discriminant analyses were performed on the December 1973 MSS image to determine an appropriate index of condition based on the MSS spectral bands. A condition index based on the sum of MSS bands 1, 2 and 4 was found to be best for both heath and woodland vegetation types. The same index was calculated from the January 1993 image (as the sum of TM bands 2, 3 and 4) and the values compared. Figure 10 shows the changes in vegetation condition over the twenty-year interval. Similar colours apply as in Figure 8. An additional colour, yellow, indicates a second level of improved sites that have not changed as much as the green areas.

Colour display of change in condition index over 20 years from 1973 to 1993. The index is based on MSS (1973) and TM (1993) data. Green, Yellow: improving condition; Blue: stable; Red: declining condition. The area shown is approximately 20km by 10km.

The sequence of air photos in Figure 11 relate the condition images to observed ecological changes. The western edge of this remnant is Wandoo and the rest is heath. The 1993 condition image is based on the heath index, which incorrectly labels the Wandoo woodland as being in poor condition. Table 1 below shows numerical values of the condition index and of changes over the period for sites within the remnant. In this table, the index has been recoded so that values of zero correspond to very good condition and one hundred to very poor. Positive changes indicate improvements in condition.

Sequence of air photos 1962-1992 for a remnant in the Kellerberrin area, together with satellite maps of the heath condition index in 1993, and the change from 1973 to 1993. Green, Yellow: improving condition; Blue: stable; Red: declining condition.

Description	1993 Condition Index	Change from 1973 to 1993
Good Condition Heaths (stable)	52	3
Central Cleared Strips (re-growing)	65	37
Eastern End (re-growing)	48	53
Woodland (western end)	78	19

Numerical values of the heath condition index (1993) and change (1973-93) for areas within the remnant shown in Figure 11. The condition index is scaled from zero (very good) to one hundred (very poor). Positive changes indicate improvements in condition. Note: index values do not apply to woodland.

5.5 Condition mapping within vegetation types - Kent Catchment

Examples of woodland vegetation from the Kent Catchment in good condition (top) and grazed condition (bottom).

From the known ground information and from interpretation of the image data, two broad vegetation classes were defined:

Initial analyses showed that the spectral signatures of these classes were different; the condition analyses were conducted separately on these two vegetation types. The first step in the analysis in each case was a discriminant analysis to determine whether differences in condition were detectable in the spectral data.

5.5.1 Woodland

Three broad descriptive condition labels are used to describe the woodland sites; 'good' condition sites, which have natural diversity of canopy and understorey; 'grazed' sites, which retain substantial tree and canopy density, but where the shrub understorey has been lost through grazing and is replaced by annual weeds (Figure 12); 'very poor' sites, which show both loss of understorey and substantial thinning of tree and canopy density. The discriminant analysis using TM data from April and September, 1993, showed clear separation of good condition from very poor condition sites. There appeared to be some spatial effect on the vegetation signature, with sites higher up in the catchment appearing generally brighter. This may be a change in vegetation with rainfall and soil. Within the lower portion of the catchment, there was shown to be a general separation of good condition sites compared with grazed sites - these differences are much smaller than those between the good and very poor sites.

The site ordination based on the 1993 data (2 dates, 12 bands) suggested that a one-dimensional condition index could be defined. Band reduction procedures indicated that the most important bands were spring band 4 (near - IR) and summer band 7 (mid-IR). The sum of these bands was shown to provide a high proportion of the information separating both good and very poor, and good and grazed sites.

A classification of 1993 woodland condition was produced using the best 4 bands from the two seasons identified by the band reduction procedure. Field visits have confirmed the accuracy of mapping of good grazed and very poor classes at sample sites within the area.

Kent Catchment (detail). Condition mapping for woodland classes in 1993. Green: good; Orange, Red: grazed; Yellow: very poor. Grey areas are mapped from spectral data as 'lowland' vegetation types, while black areas are non-bush.

The chosen smoothed condition index (summer band 7 + spring band 4) was calculated for the 1993 and 1988 data. A colour image of change in this index from 1988 to 1993 was produced to show areas of change in condition over this period. Within woodland remnants, the change in the index was generally small. Forest plantation areas are clearly shown as decreasing in this index, indicating improved condition associated with increased cover density. Indications of negative condition change were noted in some remnants and a large portion of the state forest. The areas of greatest change were found to have been affected by fire.

5.5.2 Lowland vegetation analysis

The analysis of the yate and lowland vegetation from known and screen-selected sites produced results comparable with the above. A greater spectral range was associated with differences in condition, and with changes through time. Increasing salinity and grazing have caused a decline in some areas of this vegetation.

The analysis of the 1993 data indicated that a single brightness index was adequate to summarise the known condition information; the most important bands were bands 5 and/or 7 in summer, in combination with either of the same bands in spring. Spring data alone provided less information than summer data. Band 7 (summer) alone provided adequate discrimination of good and grazed sites and was used as an index to illustrate change over the period 1988-1993.

Kent Catchment area. Representation of change by display of summer TM band 7 from 1988 and 1993 in green and red respectively for remnants within the catchment. Interpretation: dark blue: good condition; yellow: poor condition both dates; red: decline in condition; green: improvement; black: masked non-bush areas.

5.6 Estimation and display of trends through time - Kent Catchment

Index difference images over a period provide only a gross summary of change between two dates. It is generally much more informative to examine the actual patterns of response through time from a series of images through the period. In this way, gradual long term trends can be separated from sudden disturbance events. Areas which have been disturbed and recovered partly or fully can also be identified. Correction of the data to remove seasonal differences is an essential first step. Graphical plots can be used to display and compare trends for particular selected areas. For an area, trends through time can be summarised for each pixel from a series of images. One method for doing this is to estimate the linear and quadratic components of the response through time on a pixel basis.

Time plots 1988-93 of summer band 7 for sample remnant sites from the Kent Catchment. Decrease in brightness on the y-axis is associated with improvement in condition. Colours match those of Figure 16; black lines: stable sites; red: decline; orange: severe disturbance followed by partial recovery. Blue lines indicating dramatic improvement are from bluegum plantations established on formerly cleared land.

For the Kent catchment, linear and quadratic trends were estimated for band 7 from summer images over the period 1988 to 1993. This single band was used as it contains condition information for both vegetation types. The estimated trends are displayed in Figure 16, which highlights areas of different temporal response. Positive and negative linear trends in the index indicate overall decline (Red in Figure 16) or improvement (Blue) in condition. Quadratic trends indicate accelerated change or recovery from disturbance during the period. Green areas in Figure 16 have shown this recovery, while yellow/orange areas show only a partial recovery from disturbance. Black areas are unchanged. The obvious colour patterns in the state forest indicate areas of different fire history and recover. Smaller areas of change are highlighted within remnants throughout the catchment.

Kent Catchment. Colour representation of time trends fitted to summer band 7 from 1988 to 1993. Linear and quadratic trends are displayed in blue, green and red. Interpretation of colours : black: little change; blue: improvement in condition; red: decline in condition; green,yellow: disturbance during the period with full or partial recovery to 1988 condition; Grey: masked non-bush areas.

6. Discussion

The results show that satellite data can provide information on change in vegetation condition over time. The derived indices can be applied to produce "condition change" maps over broad areas which highlight areas of major change over a period. Estimates of different trends through time can be calculated from a series of images and displayed to highlight areas with different dynamic responses. Trends for individual sites can be examined by plotting the time traces for those sites. These can indicate the timing of major disturbance events and recovery from these, or gradual long-term changes.

The study has developed a methodology for deriving and applying spectral condition indices, and reports the results for two particular areas. The results show that:

Operational applications for monitoring must take account of these findings. Where an adequate map of vegetation types exists, this map can be used to mask the image to produce displays of the optimal spectral indices for each vegetation type. Regional summaries of changes in these indices can be readily produced.

Where satisfactory vegetation maps are not available, alternative displays appropriate to different vegetation types can be produced to locate and highlight areas of change. Where these do not agree, checking of the vegetation type in the field or from photographs will be necessary to verify the interpretation.

Extensive field checking may be required to produce meaningful regional summaries. However, where the interest is the identification and location of particular sites which have changed, the approach can be applied practically on a broad scale in association with field checking. Such an exercise was conducted in the Kent Catchment as part of this study; areas of change were visited and interpreted across the whole catchment in one day.

An alternative approach is to apply a "robust" index which is applicable for a broad range of vegetation types, though perhaps not optimal for any of them. Our analyses suggest that simple indices from a sequence of summer images are most suitable in the WA agricultural area. The summer brightness of bands 7 and 5 (or band 7 alone) are effective in a range of vegetation types in the Kellerberrin area and the Kent catchment.

The study has shown the association of spectral changes with differences in condition through time. Ecological interpretation of the numerical differences can be made only loosely due to a lack of objective ground monitoring data. Certainly, large changes such as fire or recovery from fire are readily detected using TM data; the magnitude of changes in the indices through time associated with these major disturbance events suggest that more subtle intermediate stages of decline or improvement will be detected. Over the 5-year period of TM imagery, areas of apparent decline due to grazing were detected in the Kent Catchment area. The 20-year history available using early MSS data in Kellerberrin cover a much greater range of change, though with reduced data quality. The example of Table 1, using the less-than-optimal MSS-TM index over the period, shows changes in index values of over 50 for regrowth after clearing. Interpretation of the photographic record provides some ecological context for changes of this magnitude, and hence for intermediate values for this index.

The present lack of ground-based monitoring information in these relatively well-mapped areas of vegetation highlights the difficulties of providing broad scale monitoring information in future from ground based data alone.

Satellite data can provide broadscale information on condition and changes through time. However, satellite data has limitations in terms of resolution and sensitivity. Some ground information is required to determine the limitations in an ecological sense, and to provide ecological interpretations of the observed changes for different vegetation types.

7. References

Campbell, N A, Furby, S L and Fergusson, B 1994. Calibrating images from different dates. Report to LWRRDC. Project CMD1. CSIRO Division of Mathematics and Statistics, Perth.

Furby, S L 1994. Vegetation condition mapping at Kellerberrin. Technical Report; CSIRO, Division of Mathematics and Statistics, Perth.

Hobbs, R J, Wallace, J F and Campbell, N A 1989. Classification of vegetation in the Western Australian Wheatbelt using Landsat MSS data. Vegetatio 80: 91-105.

Lambeck, R J and Wallace, J F 1993. Assessment of the conservation value of remnant native vegetation in the central wheatbelt of Western Australia using Landsat TM imagery. Report for ANPWS (Now ANCA).

True, D E, Kikiros, G, and Froend, R 1992. Preliminary assessment of the effects of grazing on remnant vegetation in the Kent River Water Reserve. Report No WS106. Water Authority of Western Australia.

Acknowledgements

The authors gratefully acknowledge the assistance and advice we have received at all stages of this study from our collaborators. In particular we thank Dr Graham Arnold, Dion Stephen and Robert Lambeck of CSIRO, DWE for their assistance in providing ground information and interpretation for the Kellerberrin area study. For similar roles in the Kent Catchment we thank Dr Ray Froend of WAWA, Denise True, and Sue Kelly and Dr Don McFarlane of DAWA. The project was carried out with funding provided by the Land and Water Resources Research and Development Corporation.