CSIRO Mathematics, Informatics and Statistics
Detecting and Monitoring Salt-affected Land
Detecting and Monitoring Changes in Land Condition Through Time using Remotely Sensed Data
CSIRO Mathematics, Informatics and Statistics
Agriculture Western Australia
This report summarises the work and findings of the salinity component of the LWRRDC-funded project, 'Detecting and Monitoring Changes in Land Condition Through Time Using Remotely Sensed Data' (project CDM1). Salinity mapping and monitoring methodology has been developed and evaluated in representative areas of the Western Australian wheatbelt.
The aims of the salinity study are:
The outputs from the project are:
The results show that good regional maps of salt-affected land can be produced by choosing appropriate image dates from within the growing season. Using August and October data, salt- affected land has been mapped with 99.5% accuracy near Moora and with 75% accuracy (85% bare salt, 70% marginal salt-affected classes) in the lower-rainfall Xantippe catchment north- east of Moora. In both cases, 20% of poor crop and pasture sites were incorrectly mapped as salt-affected (errors of commission). By combining data from two seasons with landform maps to reduce the confounding of seasonal and management effects, the errors of commission have been reduced to about 2%, at the expense of a decrease in the ability to accurately detect salt- affected land. These image maps can be compared through time to show where and by how much the salt-affected regions are changing. Examples are given in section 4 of this report.
On a more local scale, a series of calibrated images can be used to show the sequence of changes that occur as known salt-affected regions expand. These can be used to develop spectral indices to highlight the changes and suggest potential problem areas. Examples are given in section 6 of this report.
The results of the mapping and monitoring studies have been presented to the farming
communities at workshops and clinics held in the study areas and to the remote sensing
community through workshops and conference presentations. A list of these activities is
given in section 10 of this report.
It is widely acknowledged that dryland salinity is a major threat to the resource base of many rural industries around Australia. Its impact is felt not only on the farm; salinity generates effects that affect a wide cross section of the community. Currently more than 1.2 million ha of productive land are affected by salinity across Australia, with a further 1.6 million ha at risk. It is estimated that $243 million a year is lost in agricultural production (National Dryland Salinity R, D & E Program, 1994). In Western Australia, the Australian Bureau of Statistics (ABS) conducted a farmer survey in 1989 which reported that 443000 ha (2.8%) of the land cleared for farming in WA was now saline. This represented an annual increase of about 18000 ha per year since the previous survey in 1979. Some shires reported that up to 9% of the previously arable land had been lost, while individual property owners reported that up to 40% of their property was salt-affected (George, 1990).
Farmer-based surveys such as these are likely to underestimate the problem. New property owners do not always know when areas on their recently-acquired properties became saline. Many emerging saline areas support a cover of salt-tolerant species such as barley grass (Hordeum geniculatum) and saltbush (Halosarcia spp.). These make adequate pasture if fresh water is available, and some farmers do not include these areas in their estimates of salt-affected land.
These surveys provide statistical summaries of the amount of salt-affected land on a shire basis. They don't produce maps of where the salinity is, where it is spreading, or how fast.
Air photo interpretation has been used to map salt-affected land in some areas (Nulsen, 1981). Work by CSIRO in 1982-83 showed that multi-temporal Landsat MSS data could map some degrees of salinity (Campbell and Modawski, unpublished report). Hick and Russell (1990) used high-resolution field and multi-spectral airborne scanner data to show that near-infrared and shortwave infrared wavelengths were important for mapping saline areas.
More recently, salt-affected land has been mapped in four test areas typical of different parts of the WA wheatbelt using a single Landsat TM image (Wheaton et al, 1990 and 1992). Severely salt-affected land was mapped accurately in all four regions. Marginally saline land and salt-affected land with a good cover of salt-tolerant species could only be mapped in two of the areas, where the accuracy was 80%. These areas are spectrally confused with areas in poor condition due to other factors, such as poor or heavily grazed pasture or areas subject to wind erosion.
In 1993, a workshop on 'Remote Sensing Methods for Identification of Discharge Areas'
was held as part of the National Dryland Salinity Research, Development and Extension
Program. The emphasis in the workshop was on the identification of methods / approaches
which can be applied now. In this context, the workshop concluded that the use of
multi-temporal Landsat Thematic Mapper (TM) data was considered the most appropriate
method for mapping dryland salinity at the regional (~ 500 000 ha area) and catchment (~
30 000 ha area) scales. At this stage, Landsat TM imagery was seen as the only
cost-effective data source for monitoring land condition.
The aim of this study is to develop and evaluate methods of using multi-temporal Landsat TM imagery to detect and monitor changes in land condition in test areas in the Western Australian wheatbelt. Sequences of images from 1987 to 1994 are analysed to develop procedures, and to identify optimal image combinations, for discriminating between land in persistently poor condition due to salinity and land exhibiting a short-term decrease in productivity caused by factors other than salinity. These procedures are used to produce more accurate salinity maps and maps of change in land condition.
Earlier studies have indicated that paddocks with poor volunteer pasture growth, that have been over-grazed or that have been left fallow, cannot be spectrally discriminated from marginally salt-affected sites supporting salt-tolerant indicator species using a single date of imagery. By using multi-temporal satellite images, the different management and seasonal trends over these cover types can be used to help determine the correct condition label.
Salt-affected land usually remains so from year-to-year, perhaps dependent on the ground water level, whereas crops are rotated and grazing patterns change. If the low productivity is a temporary condition caused by management effects, the paddock will be cropped in a subsequent season and have good green vegetation cover. At that time, the paddock is likely to be easily separated from salt-affected sites. Images from consecutive seasons are used in section 4.1.3 to investigate the temporal trends of such sites to develop methods for improving the salinity mapping.
Temporal trends of sites with low productivity during a growing season are investigated in section 4.1.1, using a sequence of images to determine the optimal image dates for mapping and monitoring land condition.
Knowledge of the position of a site within the landscape also contributes to our understanding of its condition. Poor vegetation cover at sites on hilltops or slopes is more likely to be caused by grazing or erosion than salinity. Associations between salt-affected sites and landform units are investigated in section 4.1.3 and used to improve the accuracy of the salinity maps produced.
Knowledge of the soil type at a site is also expected to contribute to our understanding of its condition. Although consistent soil maps are not generally available at a regional scale, relationships between soil type and land condition are investigated in section 4.2.2 for the Esperance region and in section 8 for a small catchment in the Moora study area.
Associations of salinity, or potential salinity, with other ancillary data sets are being investigated as part of the LWRRDC-funded project, 'Integrating Remotely Sensed Data with Other Spatial Data Sets to Predict Areas at Risk form Salinity', being carried out in the Kent River Catchment, in the south-west of Western Australia.
The study areas are described in the next section. The individual steps in the salinity mapping and monitoring procedure are described in section 3, and the results are discussed in section 4. The methodology has been developed based on analyses of data from the Moora - Kalannie Goodlands study area and has then been applied, evaluated and refined in the other study areas.
Discriminant analyses have been performed to investigate temporal trends in the data and to choose optimal image dates. The results are reported with the salinity maps in section 4. Discriminant analysis procedures have also been applied to look for differences and similarities within cover types between different years and between the study areas. The results and their implications for simplifying the training steps in the classification are discussed in section 5.
Sequences of images over known expanding salt-affected sites have been examined. These results are presented in section 6. Longer-term changes, from 1981 to 1993, have been investigated using MSS images. This shows the clearing history as well as changes in the most severely salt-affected regions. These results are presented in section 7.
The results of incorporating additional datasets, such as soil types, into the salinity mapping procedure are presented in section 4 for the Esperance region and in section 8 for the Xantippe catchment in the Moora-Kalannie Goodlands study area.
The analyses were directed towards producing salinity maps and maps showing areas of changing condition by addressing the following questions:
2. The Data
2.1 The Study Areas
The salinity monitoring work has been conducted in two study areas in the south-west agricultural region of Western Australia:
The salinity mapping aspect of this study has been carried out in these two regions as well as in:
The locations of these study areas are shown in figure 1.
Figure 1: Location map showing the study areas
2.2 Image, Ancillary and Ground Data
2.2.1 The Moora-Kalannie Goodlands region
A sequence of Landsat TM images over this region has been assembled and analysed. Landsat data are received every 16 days. An archive of TM data from 1988 is held at ACRES. Some images are available from 1986 and 1987. Images have been obtained for each year from 1987 to 1993 to investigate changes in the area affected by salinity and from each month in the 1990 growing season to investigate the optimal image dates for salinity mapping and monitoring. Earlier MSS scenes have been used to look at longer-term changes in land condition.
The dates of imagery used are:
Digital height data were acquired as contours (10 and 20 metres) from the Department of Land Administration (DOLA). These data were gridded to form a digital terrain model over the study area, from which landform units, such as hilltops and broad valleys, were derived using water accumulation models. These procedures are described in Wheaton et al (1994).
Roads and cadastral boundaries have been obtained digitally from DOLA. Image pixels along these boundaries are spectrally mixed and have been labelled as saline in previous mapping exercises. Roads can also act as barriers to natural water flow, leading to outbreaks of surface salinity.
Farm plans, cropping histories and locations of salt-affected and changing sites have been supplied by farmers for twenty-three properties across the study area. Additional saltbush and barley grass sites were identified and located using GPS coordinates as part of the previous salinity mapping work. Field officers from the Department of Agriculture (DAWA) have supplied additional ground information and have assisted in the validation of the image maps produced. The local landcare groups have also assisted with this validation process.
Other digital data acquired as part of the project were only available for a few individual properties or for small catchment or sub-catchment regions. These data sets include soil maps and maps of salt-affected and potentially salt-affected areas. These data have been used as part of the validation process.
2.2.2 The Esperance Region
The sequence of Landsat TM data acquired over this region has been limited by extensive cloud cover on the days of the satellite overpasses. Instead of acquiring images from optimal dates during the growing season, we have acquired the only images available. All of the images used in the study of this area have cloud over some part of the study area. It has not been possible to acquire a sequence of cloud-free images during a growing season.
The image dates that have been used are:
A soil landscape map was provided by the Department of Agriculture. It covers a large part of the Esperance study area.
Farm plans, cropping histories and locations of salt-affected and changing sites have been provided by twenty-four farmers in the area. Officers from the Department of Agriculture have also provided valuable advice based on their knowledge of the area.
2.2.3 Narrogin Area
Salinity maps for this area were produced using Landsat TM data from:
Ground information was available in the Narrogin area from an earlier salinity survey conducted by the Cuballing LCDC group. Further information was supplied by farmers through the Blackwood Catchment Coordinating Committee.
3. Salinity Mapping and Monitoring Methodology
The steps used to produce the salinity maps and salinity change maps produced during this study are:
The final class -- inconsistent condition labels -- does not correspond to changes that actually occur on the ground. Errors in the salinity mapping at each multi-season interval mean that this combination of image labels might be obtained for a small proportion of the region. By treating these errors as a separate class, the affected areas are highlighted and can be investigated further.
Salinity maps are produced from the probabilities in step 6. Maps of change in condition are produced from the probabilities in step 7. Other map-based ancillary datasets can be included in the network used in step 6 when these are available.
The method used to combine the land condition data from two seasons, landform and, where available, soil types is a Bayesian network reasoning model. In its simplest form, this can be thought of as a series of rules governing the interpretation of the intersection of cover and landform labels. For example, one rule might be 'if cover in year 1 is salt and cover in year 2 is salt and landform is local or broad valley, then condition at time 1 is salt-affected'. This very simple formulation, however, ignores some important sources of information.
One source of information is the probability of class membership. The classification procedure applied to the image data in each year calculates probabilities of belonging to a number of cover classes. The label assigned to each site is the cover class for which it has the highest probability. Consider two sites, both labelled as salt-affected by the classification, but the first has a probability of being salt-affected of 0.95 and a probability of being in poor condition of 0.05 and the second has a probability of being salt-affected of 0.6, a probability of being in poor condition of 0.3 and a probability of being in good condition of 0.1. We might have more confidence that the first site is salt-affected than the second and want to incorporate this into our rules. Our confidence in the label is incorporated into the network model by using the probabilities instead of the label.
A second source of information that is ignored is the accuracy of the classification. Ground reference sites can give an indication of the accuracy of the individual classifications. If it is found that the salt class correctly identifies salt-affected sites 80% of the time and misses 15% of all salt-affected sites, we want to modify our rules to reflect this. This is done by expressing the rules themselves as probabilities. For example, the new rule might be 'if label equals salt, then the cover is salt with probability 0.8'.
The final two sources of information that are ignored in the simple example above relate to terrain information. The first relates to the likely occurrence of salt on hilltops or slopes. Although poor condition land in these parts of the landscape are more likely to be caused by erosion, hill side seeps are an indication that while the probability of salinity is low, it is not zero. The rules are expressed in terms of the probabilities of the occurrence of salt-affected land on each landform unit. The landform units themselves are also subject to uncertainty. The units are derived from a digital elevation model calculated by gridding contour data. If the contours are widely spaced in the horizontal direction, the edges of valleys cannot be accurately defined. A site labelled as in a valley may actually be on the slope at the edge of the valley. Instead of using a landform label, a landform probability is used.
The Bayesian network reasoning model provides the framework for combining these probabilities. The same model is used to combine salinity probabilities from two-season time intervals. The final output is a series of probabilities for the classes of consistently good condition land, consistently salt-affected land, emerging salinity and improving vegetation cover. The labels derived from these probabilities form the salinity change maps.
4. Regional Results
4.1 Moora - Kalannie Goodlands region
This study area provided the most complete set of image and ancillary data. The south-western and central portions of this study area are typical of a large portion of the Western Australian wheatbelt. The data and techniques were developed and evaluated over this study area and then applied in the remaining regions.
A sequence of images is available during the 1990 growing season. Figure 6 shows the image dates superimposed over the vegetation cover index from figure 3. The images cover each stage in the crop growth from seeding, through anthesis and curing, to a sequence of post- harvest images. Previous salinity mapping work (Wheaton et al, 1992), using the September 1990 image, concluded that salt-affected areas covered in barley grass were not spectrally separable from non-saline pasture areas that had been heavily grazed.
Cloud-free August images are available in this study area in only three out of the seven years of data studied. Further south, the proportion of cloud-free spring images decreases. Since the Spot satellite can be directed to look sideways as well as straight down, providing a greater number of potentially cloud-free repeat views of a scene, the analyses were repeated using only TM bands 2, 3 and 4 to simulate Spot image data. These analyses showed that an increased number of Spot scenes was required to maintain the discrimination between the salt- affected and pasture sites and that any image combination must include an August (anthesis) image.
Further details of these analyses can be found in Furby (1994).
Table 1 : Adequate Image Combinations for Discriminating Between Barley Grass, Saltbush and Heavily Grazed Pasture
4.1.2 Stratification of the Study Area
Variations in colour can be seen in the image of the study area in figure 2. While different colours indicate different cover types in the image, there is a general shift from very bright reds (good green vegetation cover) in the Moora region in the south-west to duller reds and more blues (less complete green vegetation cover and more bare soil) in the Kalannie Goodlands region to the north-east. A below-average wheat crop in the Moora region, perhaps caused by salinity, might be considered as average vegetation cover, not salt-affected, in the Kalannie Goodlands catchment. These variations in the amount of vegetation cover across the study area are related to factors such as rainfall, soil type and geology.
The study area was stratified into two zones. The zone boundary is shown in yellow in figure 2. Within each zone, there are no marked regional variations in the spectral response of sites within each cover type.
The Moora zone, in the south-west of the study area, has a higher average rainfall (~480mm annually) than the Kalannie zone (~340mm annually). The vegetation index in figure 3 shows a higher average green vegetation cover in the Moora zone compared to the Kalannie zone. This green vegetation cover difference can also be seen in the average annual wheat yield in these regions. The Moora zone, in the Moora shire, has an average annual wheat yield of 1.77 t/ha and the Kalannie zone, contained in the Dalwallinu shire, has an average annual wheat yield of 1.54 t/ha (Cooperative Bulk Handling (Ltd) of WA). Although the zone boundary shown here was digitised from inspection of an image formed by averaging a spring image from each year from 1987 to 1993, it shows a close association with rainfall gradients and soil and geological maps of the region.
Multivariate discriminant analyses show that some of the major cover types vary spectrally between the zones. Figure 10 illustrates the spectral separation. Red and green symbols are used to show sites from the two major zones. The remnant vegetation and salt-affected sites are similar in both zones, but there is a shift in the well-vegetated sites -- pasture, wheat and lupins. Since the aim of this study is to map salinity, the differences in the cropped sites could be ignored since they are both well-separated from the salt-affected sites.
Figure 12 is an example of a salinity map produced using maximum likelihood classification procedures applied to data from August and October 1990. The training data for this process were obtained from farm plans and cropping histories. Approximately half of the sites obtained were reserved for validating the salinity maps. The red area through the centre of the image corresponds to a saline watercourse. Salinity is spreading into nearby paddocks. At the top and bottom right of the image, the red areas, labelled salt-affected, are regions of poor vegetation cover in pasture paddocks that are not caused by salinity.
Table 2(a): Accuracy Assessment of the 1990 Salinity Map in the Xantippe Catchment
Table 2(b): Accuracy Assessment of the 1990 Salinity Map in the Moora Region
An alternative approach that has been adopted here is to use the data from another season to update the labels. Poor vegetation cover can be caused by a number of factors other than salinity, such as heavy grazing, below-average rainfall, or wind erosion events early in the season. Causes such as salinity will persist over many seasons, whereas grazing rates, cropping rotations and rainfall vary from year to year. Paddocks with poor vegetation cover due to such causes will return to good cover as these conditions alter. Thus a paddock labelled as salt-affected in two consecutive years is more likely to be salt-affected than a paddock labelled as salt-affected in one year and as good crop in the other year.
Associations between land condition and position in the landscape were also investigated. Landform units were derived by stratifying a water-accumulation map derived from the digital elevation model. This processing is described in Wheaton et al (1994). The sites used in the comparison are the validation sites for which the cover -- salt-affected, barley grass or non salt-affected -- is known. The probabilities listed in table 3 were calculated by counting the coincidence of the landform and cover labels.
Table 3 : Probability of Being Salt-Affected for Each Landform Unit
Figure 13 shows an example of the salinity map produced by combining the probability of salinity from the 1990 and 1991 seasons with the probability of salinity based on the landform unit of each pixel. Only the areas in red are labelled as salt-affected by this process. These correspond to regions with very poor or no vegetation cover in both seasons. Most of the regions that were labelled as salt-affected in 1990 in paddocks at the top and bottom right of figure 12 correspond to regions of poor vegetation cover in pasture paddocks. In 1991 most of these regions produced a good crop and were not labelled as salt-affected. The data combination process has correctly relabelled these regions as not salt-affected. Tables 4(a) and 4(b) show the results of the validation of the combined salinity map. The regions are the same as for the validation of the single-year salinity map in tables 2(a) and (b). The data in the tables show a large reduction in errors of commission (good condition sites labelled as salt-affected), but with an increase in the errors of omission (salt-affected sites labelled as good condition).
4.1.4 Changes in Salt-Affected Land Through Time
The salinity maps produced from the pairs of seasons can be combined through time to produce maps of change in salt-affected areas. Figure 14 is an example of the combination of salinity maps from 1987/88, 1990/91 and 1992/93. The regions in red, including the salt lake chains, have been salt-affected for the whole of the study. Areas in dark and light blue have become saline during the time period of this study.
Green regions have improved over the time of the study. They include regions which have been fenced off and consequently the cover of saltbush or other salt-tolerant species has greatly increased, regions that have been consistently wind-blown early in the study and are now improving, as well as marginally saline areas where changing water levels have lead to improved crop growth.
Yellow indicates regions where the salinity mapping has been inconsistent. Such regions have been labelled as good, salt-affected, then good or vice versa. The majority of these areas are not salt-affected but are on deep sands, or have been wind-blown on several occasions with a slow recovery time. These regions are land in poor condition, but the cause is less likely to be salinity. A small proportion of such sites may be emerging salt-affected areas that will still support a reasonable crop in drier years.
The final accuracy assessment, table 7, corresponds to a region in the Kalannie Goodlands catchment between the two salt lake chains in the top right hand corner of figure 2. The classification accuracy here is quite low, particularly for salt-affected sites covered in barley grass. The reasons for this are two-fold. The rainfall in this region is quite low, sandy soils predominate, and it is one of the more marginal areas along the northern edge of the wheatbelt. The vegetation cover is much less here, even in a good year. Cropping tends to occur one in every three or four years, rather than one in every two or three years in the other regions. Combining data from two seasons is less likely to include a good crop. The terrain data are also less accurate in this region. East of Lake Goorly, only 20m contours are available digitally. In this relatively flat terrain, accurately defining the edges of valleys and local rises is impossible. Cloud cover has also caused problems in this region.
Table 5 : Accuracy Assessment of the Salinity Change Map for the Xantippe Catchment
Table 8 : Percentage of Moora and Xantippe Regions Labelled as Salt-Affected
Inspection of the histograms of counts of poor condition labels suggests that the majority of sites that are labelled as poor condition in four or more years are salt-affected. Relabelling pixels based on the number of poor condition labels should also increase the accuracy of the salinity maps, particularly in the Kalannie Goodlands region.
The data combination process was extended to include data from three seasons for the Kalannie Goodlands region. Occurrences of three poor condition labels in consecutive seasons were assumed to have a high probability of corresponding to salt-affected land. Only a small incremental improvement was obtained in the accuracy of the salinity maps produced.
The most obvious difference between the condition image and the way the data combination network operates is that the former is based on counts of condition labels whereas the latter operates by combining condition probabilities. A salinity map was derived from the condition labels image as follows:
The validation results in table 9 show that if the latter three condition classes are assumed to correspond to salt-affected land, almost all of the barley grass covered sites are detected as salt-affected. The accuracy of the salt-affected class with uncertain cover has also improved.
Table 9 : Accuracy Assessment of the Salinity Map from Condition Labels for the Kalannie Region
Combining the condition labels from all years in this way has increased the accuracy of the salinity map produced. However, because data from so many seasons are required to accurately assign a condition label, the time interval of data required to separate actual change in condition from varying vegetation response caused by management and seasonal effects also increases. For example, a sequence of condition labels of good condition, good condition, poor condition, salt-affected and salt-affected may indicate a site becoming salt-affected, as for some of the bare salt-affected reference sites in 1993. It may also indicate a site cropped early in the time interval and left as pasture in the later, perhaps poorer, seasons, as for a small proportion of the pasture reference sites in 1993. If the condition labels continue to be poor and / or salt-affected, the site would be considered to have become saline. If a good condition label occurs in the next season or two, it is more likely that the site is in good condition, but a combination of seasonal and management effects has temporarily caused poor vegetation growth.
In the higher rainfall Moora and Xantippe regions, condition change can be separated from management and seasonal effects after two to three years. In these regions crop rotation patterns are typically three years. Data from four years are sufficient to produce maps of condition change. In the Kalannie region, data from at least four and up to six seasons is required to accurately determine land condition. In this region, crop rotations are typically four years. The lower rainfall and sandier soils also contribute to poorer vegetation cover. Data from more than six years are required to map changes in land condition in this region.
This study area was chosen to validate and refine the salinity change detection methodology. A very wet year in 1989 (674mm) caused large salinity problems to emerge.
Frequent cloud cover over this region has severely limited the amount of image data available. The images used have been the only partially cloud-free scenes available from each year and analyses to determine optimal image dates could not be performed.
4.2.1 Stratifying the Study Area
Regional differences are visible across the study area in figure 4. The study area was stratified based on visual inspection of an image formed by averaging the cloud-free spring images; close association with soil and landform zones was seen. Separate classifications were not performed for each zone as insufficient ground data were available in the northern, or transition, zone. It was observed in the classification images produced that the actual ground cover corresponding to some of the spectral classes varied between the zones. For example, a particular spectral class might contain very poor pasture and marginally salt-affected land in the higher-rainfall sandplain zone and average pasture (non salt-affected) in the transition zone.
4.2.2 Salinity Maps
Salinity maps for individual years have been produced in the same manner as for the Moora - Kalannie Goodlands study area. Extra refinements were made to the individual classifications, as cloud cover makes the combination of data over consecutive seasons more difficult.
Figure 17 shows the salinity map produced by this process for a typical region in the southern or sandplain zone for 1993. As in the Moora and Kalannie Goodlands regions, several paddocks with poor pasture cover have been labelled as salt-affected. Table 10 shows the accuracy of the map, assessed using reserved validation sites. Only 72% of the salt-affected sites are correctly labelled and 15% of the crop and pasture sites are incorrectly labelled as salt-affected.
Table 11 : Associations of Salinity with Landform at Esperance
Table 12 : Accuracy Assessment of the Combined 1991 & 1993 Salinity Map in the Sandplain Zone
Table 13 : Accuracy Assessment of the 1993 Salinity Map in the Transition Zone
Table 15 : Associations of Soil Categories with Salinity at Esperance
Figure 18 shows the salinity map produced by combining data from 1990, 1991 and 1993 with the soil data. Assessment against the reference sites, shown in tables 16 and 17, shows that salt-affected sites are detected with 90% accuracy, with little adverse effect on the errors of commission.
The salinity maps produced from three seasons have been combined through time to produce maps of change in salt-affected areas. Figure 19 is an example of the combination of salinity maps from 1987/88/90 and 1990/91/93. The regions in red have been mapped as salt-affected for the whole of the study interval. Areas in light blue have been mapped as becoming salt- affected during the time period of this study.
Table 18 shows the 'Mahalanobis distances' between the cover classes used in the classification for August 1993 and October 1988. In particular, the spectral class for poor crop and pasture is much closer to the partially vegetated salt-affected spectral classes.
Table 18 : Mahalanobis Distances Between Spectral Classes in the Esperance Study Area
(a) October 1988
Table 19 : Accuracy Assessment of the Esperance Salinity Change Maps (a) Sandplain Zone
Table 20 : Percentage of the Esperance Regions Labelled as Salt-Affected
Initial salinity maps have been produced using the 1993 and 1994 TM data, and the methodology for combining results over years which is described above. Landform summaries have been used to remove some errors of commission. Large-scale hard copy maps have been produced for five study areas of 30km by 24km. These maps are being evaluated by farmers and local catchment groups. Communication with these groups will continue. Information from this evaluation will be used to refine the present mapping and to produce accuracy and error estimates. The results for this region will be reported separately.
5. Using Data from Other Years and/or Regions to Simplify Processing
Supervised classification techniques are labour-intensive in the training and iteration stages. A potential benefit of sequences of calibrated images is that training signatures ought to be able to be applied to a sequence of images of a region or to different regions with similar rainfall and soil conditions.
Data from the Moora - Kalannie Goodlands study area have been used to examine whether training data can be transferred from year to year. Figure 20 shows a sequence of plots illustrating the spectral overlap between data from different dates within the Moora stratification zone. The first plot shows August 1990 and 1993 data. The only temporal trend observed is a group of crop sites in 1993 that do not seem to have spectral counterparts in the 1990 training data. 1993 was a better season than 1990 and these represent very good crops, better than were obtained anywhere in 1990. These variations do not affect the accuracy of the salinity mapping process, and so August training data can be used interchangeably between years.
Data from both the Moora - Kalannie Goodlands and Esperance study areas were used to investigate whether training data can be transferred between regions. Images of these regions can be calibrated to each other via a sequence of intermediate, overlapping scenes, although the calibrations are not as repeatable as between different images of the same region.
Figure 21 illustrates the comparison of the training data between these regions in September 1990. Comparison of the green vegetation cover curves in figures 3 and 5 suggests that the timing of the growing season is similar in the two regions. Figure 21(a) shows the comparison of data from the Kalannie zone with data from Esperance. Trends are evident in all cover types. Figure 21(b) compares data from the higher-rainfall Moora region to the data from Esperance. Again differences are evident for all cover types.
6. Using Image Sequences to Monitor Changing Sites
Sequences of calibrated images can be used to monitor sites where changes are known to be occurring. A site from the Xantippe catchment where salt is spreading into a paddock is used here to illustrate some of the possible products.
Figure 22 shows a sequence of images over the site. A salt lake chain passes through the property just above the north-west corner (top left) of the highlighted paddock. That corner of the paddock is salt-affected; since 1987 this region has been extending towards the centre of the paddock. Shades of red in these images indicate crop and pasture. The brighter the red, the more complete the green vegetation cover when viewed from above. Lupins or lush pasture correspond to the brightest red. Bare areas appear as white or light blue.
The 1989 image at the top right of figure 22 shows a 'red halo' along the edge of the salt- affected region. 1989 was an unusually dry year and many slightly salt-affected regions supported good vegetation cover. This flushing of good vegetation around known salt-affected sites has been observed in images from drier years and from late October. It appears to indicate areas with access to good groundwater that persist well after rainfall events. As the groundwater level rises, such areas are next in line to become salt-affected. Several farmers have reported very high yields just before an area becomes salt-affected.
Samples of good condition, poor condition and salt-affected land from this paddock were used to derive a condition index, displayed as an image for 1988 and 1993 in figure 23. Bright red areas correspond to little or no green vegetation cover and bright green areas have very good vegetation cover.
7. Longer-Term Changes in Land Use and Condition
Landsat MSS data are available in August 1980 for the Esperance study area and August 1981 for the Moora - Kalannie Goodlands study area. These data have been compared to the Landsat TM data from the 1990s to look for changes in land condition over the ten-year interval. Landsat MSS data have a spatial resolution, or pixel size, of 60 metres by 80 metres. Only changes in area by at least 60 - 100 metres will be detected in the comparison.
Figure 25 shows clearing maps for the Moora and Xantippe regions. Areas in green are remnant vegetation in 1993. Areas in orange were covered by remnant vegetation in 1981; the vegetation has been lost due to clearing or salinity.
The right hand image in figure 25 is typical of the Xantippe and Kalannie Goodlands catchments. Large-scale clearing of the area for agriculture occurred about thirty years later than in the Moora region. Several small bush blocks have been cleared in the past twelve years. 4.2% of the Xantippe catchment is remnant vegetation, with 0.8% cleared since 1981. In the Kalannie Goodlands catchment, 6.5% of the area remains as remnant vegetation, with 2.8% of the region cleared since 1981.
The final clearing map, figure 26, shows the eastern portion of the Esperance study area. It shows quite large tracts of land that have only recently been cleared for agriculture. 23.2% of the Esperance area remains bush, while 18.1% has been cleared since 1980.
Image displays created by combining image bands from the MSS and TM images were used to detect changes in the extent of salt-affected areas in the Moora - Kalannie Goodlands study area. In the very south-west corner of the study area, the Moora region, most of the changes which are apparent correspond to loss of vegetation cover along the water courses. Further north-east, towards the centre of the study area, numerous areas of expanding and newly- emerging salinity can be found. The image sequence in figure 27 shows one such example. Further north-east, in the Xantippe and Kalannie regions, no significant changes in salt-affected areas can be observed. This area is only just beginning to show an emerging salinity problem.
8. Using Other Ancillary Data to Map Salinity
One of the most common errors in the salinity and salinity change maps is the labelling of areas that are consistently in poor condition as salt-affected. Examples of this are the yellow areas in figures 14 and 19. These typically correspond to areas that are frequently windblown or areas of deep sands, which rarely support good vegetation cover. This latter error might be corrected by incorporating data on soil type into the processing. The results of such a procedure are described here. For more details of the processing, see Caccetta et al (1995).
Detailed soil maps have been obtained for the Xantippe catchment, just north-east of the centre of the image in figure 2. These detailed soil classifications have been combined into three major types - deep sands, clay flats and 'other'. Associations between these soil types and salinity have been investigated using a salinity map produced by the catchment advisor from extensive field work. These comparisons showed that deep sands have a high probability of poor vegetation cover, but a low probability of being salt-affected. Clay flats, however, are the predominant soil type in the valleys and have a higher probability of being saline.
Another ancillary data set that has been derived from the digital terrain model is the drainage network, from which points at which water flows converge can be derived. Expert advice suggests that these areas also have a higher probability of salinity.
Figure 28 shows an example of the salinity map produced using the satellite data from 1990, 1991 and 1993 in the form of class label images, landform units, soil maps and drainage networks. Also shown are the salinity map produced using just the class label images from 1990 and 1991, reproduced from figure 13, and the 1993 salinity map produced by the catchment advisor. Including the extra soil and drainage information has relabelled many of the 'salt- affected' areas mapped by using only the satellite and landform data. Comparison with the ground-truth salinity map suggests that the new salinity map is more accurate. Using the soil and drainage information has also relabelled some of the 'good condition' areas adjacent to areas mapped as salt-affected. Again, comparison with the ground-truth map suggests that these new labels are correct.
Table 21 : Accuracy Assessment of Salinity Maps Against Ground Mapping (a) 1990, 1991 and Landform Units
This study has shown that more accurate salinity maps can be produced by incorporating multi- temporal satellite images and landform and soil type data into the mapping process. It has also demonstrated the ability to detect and display short-term (2 - 4 years) and medium-term (~ 10 years) changes in land condition.
Earlier salinity mapping studies concluded that land in poor condition due to causes other than salinity could not be spectrally separated from salt-affected land supporting a cover of salt- tolerant grasses, such as barley grass, using satellite image. This study has shown that in the Moora region, such sites show different temporal trends during the growing season and that the combination of August (anthesis) and October (curing) images allows better separation of these sites. Data from both the Moora-Kalannie Goodlands and Esperance study areas show that even when only a single image is used, the choice of image date can greatly affect the accuracy of the mapping.
Using images from more than one season greatly improves the accuracy of the salinity maps produced. Salt-affected land remains so from year to year, perhaps dependent on the ground water level, whereas crops are rotated and grazing patterns, management strategies and rainfall patterns change from year to year. If the low productivity is a temporary condition, the paddock is likely to be cropped or to support good pasture cover in a subsequent season. At that time, the paddock is more easily separated from the salt-affected sites. Data from the Moora region suggests that two seasons are adequate to correctly label all but 1% of the sites in poor condition that are not salt-affected. Three seasons are needed in the Kalannie Goodlands and Esperance regions.
Landform information, where it can be obtained with sufficient accuracy, has contributed to the production of better salinity maps. Poor condition land on hilltops and slopes has been shown to be unlikely to be caused by salinity in the Moora region. Although similar associations are expected in other regions, the coarseness of the original contour data in the Kalannie Goodlands and Esperance regions meant less accurate terrain models could be derived. Better resolution is required in these relatively flat terrain types.
Soil type information has also been shown to be important to the accuracy of the salinity maps produced. Areas of deep sands were shown to have consistently less vegetation cover than surrounding areas, but had a low probability of being saline in both the Esperance region and the Xantippe catchment in the Kalannie Goodlands region. Clay flats, the predominant soil type in the valley floors in the Xantippe catchment, were found to have a high probability of salinity. Including this information was shown to not only help correct the labels assigned to sites in poor condition not caused by salinity, but also to correctly label salt-affected areas wrongly labelled using just satellite and landform data. Such soil maps are not available digitally over the whole of the study areas. Even when the soil maps were available, expert assistance was required to categorise the soil types into classes relevant to salinity mapping.
Combining the labels from the single-year classifications to produce a condition image in the Moora - Kalannie Goodlands study area identifies alike regions that cross paddock and road boundaries. This suggests that the sequence of labels is identifying underlying land condition, independent of particular cover types in particular years.
Salinity change maps can be produced at a regional scale, although only small examples are shown here. Assessment of these maps against validation sites and evaluation by DAWA field officers and catchment advisors suggest that the changes in condition being detected are accurate but that omissions exist. Not all of the changes identified are caused by salinity. Many of the areas mapped as having increasing vegetation cover were in poor condition at the beginning of the study, due to management or persistent wind erosion, and the recovery cycle is 3 - 4 years to reach their current good-condition status. A small proportion of these areas were, and still are, salt-affected but were fenced off early in the study interval and now support a very good cover of salt-tolerant species.
Feedback on the salinity maps and salinity change maps from DAWA officers and catchment group advisors in the Moora and Kalannie Goodland regions has indicated that even though these maps are not completely accurate, they are still very useful products when combined with a recent image of the area. Salinity maps have been produced by DAWA officers for parts of this study area based on extensive field work over a period of twelve months. To create a salinity change map would require another twelve months of field work. Having an image- derived salinity change map, updated by local knowledge and field visits to particular sites of interest, is perceived as a much more time effective means of updating their information base.
The salinity mapping and monitoring methodology presented in this report was developed using image and ancillary data from the Moora region. Using data from two seasons with landform units, salt-affected areas can be detected with 90% accuracy with 2% of other areas mapped as salt-affected. Data from four years are required to produce maps of condition change. This methodology was then routinely applied and evaluated in the Kalannie Goodlands and Esperance regions. Reduced accuracy was obtained.
The methodolgy was then refined in these regions. In the Kalannie Goodlands region, condition labels from 1987 to 1993 were combined. This increased the salinity detection accuracy to 71% (from ~40%) with less than 1% of other sites labelled as salt-affected. Data from more than six years are required to map changes in land condition in this region. In the Esperance region, combining image data from three seasons with soil data leads to salt-affected land being detected with 90% accuracy (from ~75%) with 4% (sandplain zone) and 8% (transition zone) of other areas mapped as salt-affected. Data from six years are required to map changes in land condition in this region.
10. Workshops / Clinics / Conference Presentations
The results of the project have been actively disseminated at workshops and clinics for agricultural advisors and LCDCs, at agricultural shows and at conferences. (Some of the workshops have been run jointly with an NLP-funded project on 'Remote Sensing Training for Resources Officers and Farmers'.)
Campbell. March 1994/95:
Workshops will be held to present the results of the salinity mapping and monitoring work to farmers, land care advisors and DAWA officers in the Moora and Kalannie Goodlands region (at Dalwallinu), Esperance and the Dumbleyung / Narrogin regions.
Project results will also be presented at the 8th Australasian Remote Sensing Conference in Canberra in March 1996 and at the 4th Australian National Workshop on Rehabilitation and Productive Use of Saline Land at Albany, Western Australia in March 1996.
Caccetta, P., Campbell, N. A., West, G., Kiiveri, H. T. and Gahegan, M. (1995). Aspects of reasoning with uncertainty in an agricultural GIS environment. Workshop on Knowledge Based Systems in Resource Management. AI 94. Armidale, Australia, November 1994.
Campbell, N. A., Furby, S. L. and Fergusson, B. (1994). Calibrating Images from Different Dates, Interim Report to LWRRDC from the project, Detecting and Monitoring Changes in Land Condition Through Time Using Remotely Sensed Data (CDM1).
Campbell, N. A. and Wallace, J. W. (1989). Statistical methods for cover class mapping using remotely sensed data. Proc. Int. Geosci. Remote Sensing Symp: 493-496.
Furby S. L. (1994). Discriminating Between Pasture and Barley Grass and Saltbush Using Multi- temporal Imagery. CSIRO Division of Mathematics and Statistics Technical Report.
George, R (1990). The 1989 saltland survey. J. Agric. West. Aust., 31, 159-166.
Hick, P. T. and Russell, W. A. R. (1990). Some spectral considerations for remote sensing of soil salinity. Aust. J. Soil. Res, 28, 417-31.
Nulsen, R. A. (1981). Salt-affected land in the Shire of Wongan - Ballidu, Western Australia, Aust. J. of Soil Res., 19, 87-91.
Wheaton, G. A. and Wallace J. F. (1990). Mapping Dryland Salinity. Interim Report to the National Soil Conservation Programme, Project number 7405 Mapping Land Degradation Using Remote Sensing.
Wheaton, G. A., Wallace, J. F., McFarlane, D. J. and Campbell, N. A. (1992). Mapping salt- affected land in Western Australia. Proceedings of the 6th Australasian Remote Sensing Conference, Wellington, New Zealand 1992, Volume 2, 369 - 377
Wheaton, G. A., Wallace, J. F., McFarlane, D. J., Furby, S. L., Campbell, N. A. and Caccetta, P. (1994). Mapping and monitoring salt-affected land in Western Australia. Proceedings of Resource Technology 94 Conference, Melbourne, Australia 1994, 531-543.
For More Information
For more information on salinity mapping and monitoring contact Suzanne Furby.