Mapping Salinity in the Loddon and Campaspe Catchments. in Victoria

Transcription

1 Mapping Salinity in the Loddon and Campaspe Catchments in Victoria A report from the LWWRDC project Mapping Dryland Salinity (CDM2) S. Furby CSIRO Mathematical and Information Sciences CMIS 98/115

2 2 Summary This report summarises the Victorian component of a project funded by the Land and Water Resources Research and Development Corporation (LWRRDC) on Mapping Dryland Salinity (CDM 2). The aims of the overall study are: to map the extent and severity of dryland salinity in three catchments using multi-temporal Landsat TM data; and to provide guidelines to support the use of the approach nationally. This report describes the findings for the Loddon and Campaspe catchments in Victoria. The methodology used has been extended to include digital elevation data as well as Landsat TM imagery in the processing, since information derived from elevation data has been found to improve salinity mapping accuracy in Western Australia. This study has shown that using Landsat TM data from two separate seasons (1992 and 1995) cannot adequately map salinity in these catchments. Areas of low productivity that are being mapped include salt-affected land as well as areas of poor pasture, waterlogged land and rocky outcrops. Approximately 30% of the non-saline land is being mapped as salt-affected. Including digital elevation data, processed to identify landform units such as hilltops, slopes and valleys, improves the salinity maps produced. The value of the elevation data are that they permit the identification of areas of persistently low productivity that are in parts of the landscape with a low salinity risk, such as hilltops and slopes; these areas can be relabelled accordingly. The commission errors in the salinity map are reduced to approximately 20%, but there are still too many low-productivity areas being labelled as salt-affected that are due to waterlogging, poor pasture and urban features for the final product to be a useful salinity map. Only about 55% of the salt-affected land in the catchment is being detected. The most common omission is marginally saline areas supporting sea barley grass, a salt-tolerant species of grass. Better training data and images from at least three consecutive seasons may improve the mapping accuracy. Better training data would include more examples of all the types of salinity present in the region, in particular, areas of sea barley grass. Data from consecutive seasons would allow better discrimination between short-term management effects, such as poor pasture due to over-grazing, and long-term land condition effects, such as salinity and waterlogging. This study also shows that the resolution of the DEM relative to the variation in the terrain is important for calculating useful landform summaries. Landform units could not be derived for the flatter parts of the combined catchments from a DEM on a 100m grid. The findings in New South Wales and South Australia have been reported separately and a summary report comparing the results in each state with those for Western Australia has also been prepared.

3 3 Introduction In August 1993, the Land and Water Resources Research and Development Corporation (LWRRDC) in collaboration with the Murray Darling Basin Commission, the Department of Primary Industries and Energy, and the States reviewed remote sensing with respect to dryland salinity, its research, development, and impact on land management in Australia. Three workshops were held. The most relevant of these was The Use of Remote Sensing in Saline Discharge Identification. At this workshop, through a process of review and discussion, specific objectives were established for various methods of remote sensing for farm, catchment and regional scales. The emphasis was on the identification of methods / approaches which can be applied now. The workshop concluded that the use of multi-temporal Landsat TM imagery was appropriate for mapping dryland salinity at regional (~ ha area) and catchment (~ ha area) scales. The current project was supported by LWRRDC to assess the transferability of methodology developed in Western Australia to other parts of Australia and to develop national guidelines for the application of the approach. The original aims of the study were: to map the extent and severity of dryland salinity in three catchments using multi-temporal Landsat TM data; and to provide guidelines to support the use of the approach nationally. During the course of the project, it was recognised, from work being carried out concurrently in Western Australia, that including digital elevation and using images from consecutive seasons rather than multiple images within a season improves the accuracy of the salinity maps. This approach has been adopted to some degree in two of the three regions considered. The areas chosen for the study were: the Upper South East catchment in South Australia; the Loddon and Campaspe catchments in Victoria; and the Liverpool Plains catchment in New South Wales. Each is a LWRRDC focus catchment with different land types and land use characteristics. The findings in New South Wales and South Australia have been reported separately and a summary report comparing the results in each state with those for Western Australia has also been prepared. 2. The Study Area The Loddon and Campaspe catchments are part of the Murray Darling Basin, and drain into the Victorian side of the Murray River. For the purpose of this project, the study area is the dryland salinity part of the catchments, the approximate location of which is shown in figure 1. The southern edge of the irrigated agricultural districts as defined by the Waranga Western Channel forms the northern boundary of the study area; the Great Dividing Range forms the southern boundary; the Goulburn River catchment forms the eastern boundary; and the Avoca catchment forms the western boundary. The Campaspe catchment has an area of hectares and the Loddon catchment covers hectares, within a 120km by 230km region.

4 4 Figure 1: Location map showing the Loddon-Campaspe study area. Figure 2 shows the October 1995 Landsat TM image of the study area, with the catchment boundary shown in yellow. The northern parts of the area shown are predominantly cropped, while the southern areas are a mixture of cropping and grazing. Figure 2: October 1995 Landsat TM image, bands 3, 5 and 4 in BGR. Shades of red and orange indicate good green vegetation cover. Shades of green and blue indicate less vegetation cover, through to bare paddocks in light blue and urban areas in darker blue. The dark green areas are remnant vegetation.

5 5 Agriculture in the area predominantly consists of wool, fat lamb and wheat production. Other crops include oats, barley, lupins, fieldpeas, chickpeas, canola and safflower. Salinity in the dryland parts of the catchments is due mainly to the clearing of native tree cover and the planting of shallow-rooted crops. The combination of the increased water seepage from the cleared recharge areas with the shallow-rooted crops has caused the water table to rise, dissolving soils salts and bringing them to the surface. There is less of a salinity problem in the permeable soils in the region, e.g. alluvial gravels and basalts, as salt is more easily washed out of them. Higher concentrations of salt are found in the low permeability soils, such as clay, where water cannot easily penetrate. The majority of salinity is occurring in the hill country. This is characterised by stream discharge and break-of-slope discharge. This is particularly evident as narrow linear areas associated with stream patterns. Salinity levels range from marginal areas characterised by low-productivity cropping and increasing levels of indicative salt-tolerant plant species, such as sea barley grass and spiny rush, to severe areas dominated by bare, salt-encrusted ground with dead or dying trees. Current methods of controlling salinity in the region include the cessation of further tree clearing and the encouragement of further plantings, particularly in recharge areas, of crops and perennial pastures such as lucerne, phalaris and crocksfoot. 3. Image, Ancillary and Ground Data The Landsat TM images used in this study are: 2 July August October October 1995 The sequence of images through the 1992 growing season was the most recent cloudfree sequence available at the commencement of the project in late August was a wet year with about double the average rainfall from August onwards. Flooding occurred in the later months of the growing season. The flooding is particularly noticeable in the October image. The September 1995 image was added to the processing at a later stage to provide a more recent look at the area and to provide data from another season to possibly allow discrimination between management and land condition effects. The 1995 season was considered more average in terms of rainfall. Digital height data were acquired for this project from CSIRO Land and Water (CLW). Two datasets were provided. They are: a digital elevation model (DEM) on a 20m grid for the Axe Creek catchment, south-east of Bendigo; and a DEM on a 100m grid for the whole of the Loddon and Campaspe catchments. Landform units, such as hilltops, slopes and valleys, were derived from the DEMs using water accumulation models. These procedures are described in Caccetta (1996).

6 6 A digital road mask has been obtained. Image pixels along these boundaries are spectrally mixed and have been labelled as saline in previous mapping exercises. Ground information has been provided in three stages. Early in the project, a 1:25000 saline discharge map was used to identify salt-affected regions within the catchment (Allan, 1994). The map has been compiled by field survey in consultation with field officers and land holders, and classifies saline sites into severity classes according to indicative plant species. Areas delineated in this dataset are labelled as having homogeneous characteristics in regard to salinity. Inspection of these sites within the image data showed that each site is made up of several different land cover types, some of which have quite a range of green vegetation cover. This is shown by the different colours within the yellow discharge polygon boundaries in figure 3. This range of cover types made the saline discharge map unsuitable to use directly as a source of training and validation data. Figure 3: October 1995 Landsat TM image, bands 3, 5 and 4 in BGR, for a region within the Axe Creek catchment. Shades of red and orange indicate good green vegetation cover. Shades of green and blue indicate less vegetation cover, through to bare paddocks in light blue and water and roads in darker blue. The yellow lines are the boundaries of polygons in the 1:25000 saline discharge map. Areas inside these boundaries are supposedly homogeneous with respect to their salinity status.

7 7 Additional ground data were provided by field officers from the Centre for Land Protection Research (CLPR) in June These data provided information about representative salt-affected sites in Sites were located directly onto prints of the October 1995 image. In August 1997, some preliminary salinity maps were assessed by CLPR. The locations of the assessment sites were not provided. 124 single-pixel sites were identified and the salinity status identified by available ancillary datasets a tree cover map and the 1:25000 saline discharge map. Thirty sites were visited in the field to assess the ground cover. In June 1998, forty-two additional field sites in the Axe Creek sub-catchment were visited to assess the ground cover. 4. Salinity Mapping Methodology The steps used to produce the salinity maps in this study are listed below. This is the standard methodology being used for salinity mapping work in Western Australia, except that image data from consecutive years are not available. More details on the methodology can be found in Furby et al (1995). 1. Co-register the images to a common map base (here, AMG coordinates at 25m pixel size). This allows ground sites to be traced through time and the satellite data to be compared to the ancillary map data. 2. Calibrate the image data from different dates to a reference image so that digital counts from different image dates can be compared (Furby et al, 1996). 3. Locate ground sites of all the major cover types in each of the images. 4. Stratify the study area into zones within which there are no marked regional variations in rainfall, land-use types or rotations, geology, predominant soil types or visible patterns in the image. If there are strong differences between these zones, they should be processed separately. 5. Apply discriminant analysis procedures, in particular canonical variate analyses (Campbell and Atchley, 1981), to the training data to examine the separation of ground cover types in the TM spectral data and to determine which image dates are most appropriate and to define sensible spectral groupings of ground cover types. 6. Apply neighbourhood-modified maximum likelihood classification techniques (Campbell and Wallace, 1989) to the best individual image dates. This produces probabilities of belonging to each of the major cover classes on each date for each pixel in the images. 7. Combine the cover class probabilities from each single-date classification to calculate the probability of each pixel being salt-affected. A conditional probability network has been used for these calculations (Caccetta et al 1995). 8. Combine the cover class probabilities from each date with position in the landscape hill, slope, valley floor to calculate the probability of each pixel being salt-affected. A conditional probability network has been used for these calculations.

8 8 Although an obvious stratification of the study area could be made into predominantly pasture and cropping zones, see figure 2, this was not done as adequate ground data for the cropped regions was not supplied. Including information on position in the landscape (step 8) was performed only after an initial assessment of the salinity maps produced using only Landsat TM image data (step 7). The assessment showed that the amount of salt-affected land was being greatly overestimated. A large proportion of the areas that were incorrectly labelled as salt-affected were found to be on hilltops or local high areas. Experience from salinity mapping work in Western Australia suggested that these errors could be corrected using landform units derived from a DEM. In its simplest form, a conditional probability network (CPN) can be thought of as specifying a set of links between related datasets and a set of rules governing how those datasets are combined. The final network used in this study, including landform information, is shown in figure 4. Figure 4: The conditional probability network used to produce the salinity map. Underlying the model are the true salinity maps at each available image date (square boxes). It is these maps that are being estimated. At each image date, land-use / condition classification images (upper circles) have been formed. They are a realisation of the true salinity map. With sufficient ground data, the error rates between the true salinity maps and the land condition maps can be estimated. There are also relationships between the true salinity maps. If it is known that an area is salt-affected at one point in time, it can be assumed that it will continue to be salt-affected in the following season with

9 9 high probability. Also there may be some information on how likely an area is to change from non-saline to saline over the three-year interval. Similarly there is a relationship between landform (lower circle) and the true salinity maps. If the landform type of an area is known, a statement can be made about how likely it is to be salt-affected. Each of the arrows in the diagram represents a relationship between the data sources. The exact nature of the relationships can be specified by expert knowledge; can be derived from ground training data; or can be estimated from the datasets themselves. Conditional probability networks differ from a simple intersection of two or more data layers in two ways. Firstly, the probabilities of belonging to each land-use / condition class are the inputs to the model, not class labels. This allows a measure of uncertainty in the class label to be included in the calculations. For example, consider two sites, both labelled as salt-affected by the land-use classification. Suppose the first site has a probability of being salt-affected of 0.95 and a probability of having poor cover for other reasons of Suppose the second site has a probability of being salt-affected of 0.60, a probability of being in poor condition of 0.30 and a probability of being in good condition of We might have more confidence that the first site is salt-affected than the second site and want to incorporate this into the rules. Secondly, the rules themselves are expressed in terms of probabilities, not hard yes-no rules. For example, salinity in this region is relatively rare on hilltops. However, hillside seeps are found in the catchments. In this case, while the probability of salinity in the higher regions will be low, it is not zero. 5. Results 5.1. Salinity Map formed Using only Satellite Images and Ground Data Figure 5 shows the salinity map produced for a typical region in the Axe Creek catchment, south-east of Bendigo. Table 1 shows the assessment of the map against independent validation data by CLPR. The assessment was performed by: selecting 124 pixels at random (sample sizes within cover classes are proportional to the area assigned to each class in the salinity map) across the study area comparing the sites to the 1: discharge map and a 1: tree cover map to generate omission and commission error rates inspecting 30 of the wrongly labelled sites in the field to establish how the errors might be corrected. It should be noted that the 1: discharge map is not considered an ideal assessment of the salinity status of a particular site. The values in the table are the numbers of sites assigned to each category.

10 10 Salinity Map Label Table 1: Assessment of Salinity Map No Landform Water Reference Sites Forest Reference Sites Salt-affected Reference Sites Non-saline Reference Sites Water Forest Salt - severe Salt - marginal Non-saline Figure 5: 1995 salinity map for part of the Axe Creek catchment formed using only Landsat TM images and ground data. red is severely salt-affected land yellow is marginally salt-affected land green is forest shades of grey are non-saline land, dark areas have little green vegetation cover and bright areas have good green vegetation cover The figures in table 1 show that 2 of the 3 salt-affected sites are detected and 30% of nonsaline sites are being labelled as salt-affected. There are too few saline validation sites for this to be a reliable accuracy assessment.

11 11 Field inspection showed that the non-saline reference sites incorrectly labelled as saltaffected fell into the following categories: rocky outcrops on low rises, surface stone on hilltops (11) poor pasture infested with onion grass (4) rocky granite hills with unimproved pasture (2) waterlogged reedy areas on alluvium (2) scattered trees along roadsides (2) overgrazed low-lying areas on sedimentary bedrock (2) A number of the forest reference sites have also been mapped as salt-affected. These correspond to regions of scattered trees on highly reflective soils. The Landsat TM image of the study area in figure 2 shows that most of the forested regions are surrounded by bright areas. These regions have very poor green vegetation cover. Many of these regions have been mapped as salt-affected using Landsat TM data only, causing the red halo around the forested regions in figure 5. Areas of scattered woodland are also mapped as salt-affected The Contribution of the Elevation Data Several of the misclassified regions identified in the previous section have been noted to be on hilltops or local rises. The probability of salinity occurring in such parts of the landscape is quite low. Figure 6 shows the salinity map produced by adding landform into the processing for the same region in the Axe Creek catchment as shown in figure 5. Table 2 shows the proportions of the Axe Creek catchment assigned to each cover class in the two salinity maps. Table 2: Proportion of Axe Creek Catchment Assigned to Each Salinity Map Cover Class Map Class Salinity Map No Landform Salinity Map with Landform Water Forest Salt - severe Salt - marginal Non-saline Clearly the amount of land being mapped as salt-affected has decreased by including the landform information. The proportion of both the forest and non-saline cover classes has increased. Pixels around the edge of the forested regions that are high in the landscape, and hence assumed to have a lower salinity risk, have tended to be relabelled as forest reducing the halo effect observed in the salinity maps. Pixels in the middle of paddocks in low salinity risk areas have tended to be relabelled as non-saline. The value in using landform information comes from the ability to assign a more appropriate condition label to poor condition land in parts of the landscape that are not prone to salinity, such as bare hilltops or slopes.

12 12 Figure 6: 1995 salinity map for part of the Axe Creek catchment formed using Landsat TM images, landform information and ground data. red is severely salt-affected land yellow is marginally salt-affected land green is forest dark grey indicates roads shades of grey are non-saline land, dark areas have little green vegetation cover and bright areas have good green vegetation cover Table 3 shows the assessment of the new salinity map against independent validation data by CLPR. The assessment was performed by visiting representative parts of the catchment and noting the actual ground cover of areas labelled as salt-affected. Nearby areas of salt-affected land that were not detected in the salinity mapping process were also noted. The values in the table are the number of sites assigned to each category. For some of the salt-affected sites, the salinity mapping process labelled part of the area as saltaffected, but under-estimated the full extent of the salt-affected area. Half of each of these sites has been allocated to correctly labelled as salt-affected and the other half assigned to incorrectly labelled as non-saline. Similarly, if the road mask corrected the label for part of a site, the site was split between the appropriate categories.

13 13 Table 3: Assessment of Salinity Formed Using Satellite Imagery and Landform Image Class Salt-affected on Wet (waterlogged) Not salt on ground ground on ground Salt - severe Salt - marginal Non-saline Road The figures in table 3 show that 55% of salt-affected sites are detected. Of the area labelled as salt-affected by the mapping process, only 30% of sites are actually saltaffected. The sites incorrectly labelled fall into the following categories: salt-affected areas supporting a cover of sea barley grass (a salt-tolerant species) are labelled as non-saline; waterlogged areas covered in spiny rush ( a water-tolerant species found on both waterlogged and salt-affected sites) are labelled as salt-affected; and persistently bare areas such as roadside verges, houses and shops and tracks through the forest are labelled as salt-affected. Very few sites of sea barley grass were provided. The sites that were provided were very similar spectrally to poor pasture areas and could not be separated using a single image date. Similar problems have been encountered in mapping salt-tolerant grasses in Western Australia, and the accuracy is generally much lower for such areas than for more severely salt-affected sites. Spiny rush is a species that is associated with wet land in the Loddon-Campaspe catchment area. It is a good indicator of persistent waterlogging, but is equally common on both saline and fresh areas. Spiny rush is not considered an indicator of salinity unless other salt-tolerant species are also present. Several training sites covered in spiny rush were provided as being salt-affected, leading to all occurrences of spiny rush being mapped as salt-affected. In situations where spiny rush is the dominant vegetation cover, it is not possible to discriminate between salt-affected and waterlogged land. Rush species are also an indicator of waterlogging in the Upper South East catchment in South Australia; however, the species that are common in that region are restricted to fresh water areas only. The salt-affected validation sites for which the salinity mapping process labelled part of the area as salt-affected and part of the area as non-saline, all have a region covered in spiny rush surrounded by, or adjacent to, a region covered in sea barley grass. The region covered in sea barley grass is the region not being mapped as salt-affected. The persistently bare sites that have been mapped as salt-affected are all man-made structures that can be expected to be consistently free of vegetation. Such structures are usually mapped as salt-affected in Western Australia also; however, there are fewer such structures in typical Western Australian rural areas. The spectral signatures in the October 1995 image for about half of the bare validation sites are typical of the salt-affected training sites. The signatures for the other bare validation sites are less typical of the salt-affected training sites and more typical of the

14 14 bare training sites. The less typical validation sites were used to retrain the October 1995 classification and a new salinity map was formed using the CPN updating process. An accuracy assessment performed on the new salinity map showed that of the area labelled as salt-affected by the mapping process, 72% of sites are actually salt-affected (up from 30%). The reduction in commission errors is, however, at the expense of detecting salt-affected land; only 42% of salt-affected sites were detected (down from 55%). The persistently bare validation sites cannot be reliably separated from saltaffected land using the two years of image data; however, transient bare sites, due to management rather than man-made structures, may be separated by images from consecutive seasons. The accuracy assessment does not comment directly on the areas mapped as bush or on other non-saline classes that were correctly labelled. To try to estimate the proportion of the non-saline land labelled as salt-affected by the salinity mapping process, an accuracy assessment was performed using the original training sites. The assessment showed that approximately 20% of non-saline sites are labelled as salt-affected. The training data are not independent of the classifications performed, so it is not a true accuracy assessment; however, it does give an indication of the order of the commission errors. A salinity map formed by including landform information has been produced for the Axe Creek catchment only. Adequate landform maps cannot be produced for the flatter parts of the Loddon and Campaspe catchments from the DEM on the 100m grid. Figure 7 shows two landform maps for the Axe Creek catchment. In the map on the left-hand side, the landform units were derived from the DEM on the 20m grid. In the map on the righthand side, the landform units were derived from the DEM on the 100m grid. The former shows much more terrain variation within the low, flat (pink and green) areas in the top part of the catchment. The low flat areas cover 37% of the combined Loddon and Campaspe catchments based on the DEM on the 100m grid. Figure 7: Landform map for the Axe Creek catchment from the DEM on a 20m grid (left) and 100m grid (right). red indicates hill tops yellow indicates upper slopes green indicates lower slopes and local valleys pink indicates valley floors

15 15 These landform maps show that the DEM resolution must match the degree of terrain variation in the region of interest before landform information can be derived. A higherresolution DEM does exist for the combined catchments, but it was not provided for this study. 6. Conclusions This study has shown that using Landsat TM data from a single season cannot adequately map salinity in these catchments. Areas of low productivity are being mapped that include salt-affected land as well as areas of waterlogging, poor pasture and rocky outcrops. Approximately 30% of the non-saline land is being mapped as salt-affected. Including digital elevation data, processed to identify landform units such as hilltops, slopes and valleys, improves the salinity maps produced. The value of the elevation data is that they permit the identification of areas of persistently low productivity that are in parts of the landscape with a low salinity risk, such as hilltops and slopes, and relabels them accordingly. The commission errors in the salinity map are reduced to about 20%, but there are still too many low-productivity areas being labelled as salt-affected that are due to waterlogging, poor pasture or urban structures for the final product to be a useful salinity map. Marginally salt-affected areas covered in salt-tolerant grasses, such as sea barley grass, are not being adequately mapped. This study also shows that the resolution of the DEM relative to the variation in the terrain is important for calculating useful landform summaries. Landform units could not be derived for the flatter parts of the combined catchments from a DEM on a 100m grid. Better training data and images from at least three consecutive seasons may improve the mapping accuracy. Better training data would include more examples of all the types of salinity present in the region, in particular, sea barley grass areas. Data from consecutive seasons would allow better discrimination between short-term management effects, such as poor pasture due to over-grazing, and long-term land condition effects, such as salinity and waterlogging. 7. Acknowledgements The author wishes to thank the following people for their assistance during the project: Sara Hill and Rob Clark for providing ground training data and assessing the salinity maps produced; and Adam Gilbee for performing the initial image data processing.

16 16 7. References Allan, M.J. (1994). An assessment of secondary dryland salinity in Victoria. Technical Report No. 14. Land Protection Branch, Department of Conservation and Natural Resources, Victoria. Caccetta, P., Campbell, N.A., West, G., Kiverii, H. K. and Gahegan, M. (1995). Aspects of reasoning with uncertainty in an agricultural GIS environment. The New Review of Applied Expert Systems, 1, Caccetta, P. (1996). Remote sensing, GIS and Bayesian knowledge based approaches for land condition monitoring. PhD Thesis, Curtin University of Technology. Campbell, N. A. and Atchley, W. R. (1981). The geometry of canonical variate analysis. Syst. Zool., 30, Campbell, N. A. and Wallace, J. F. (1989). Statistical methods for cover class mapping using remotely sensed data. Proc. Int. Geosci. Remote Sensing Symp: Furby, S. L., Wallace, J. F., Caccetta, P. and Wheaton, G. A. (1995). Detecting and monitoring salt-affected land. Report to LWRRDC project Detecting and Monitoring Changes in Land Condition Through Time Using Remotely Sensed Data (CDM1). Furby, S. L., Palmer, M. J. and Campbell, N. A. (1996). Image calibration to like values. Proceedings of 8 th Australasian Remote Sensing Conference, Canberra, Australia.