Any spatial accuracy of the final product will depend to a large degree on the spatial accuracy of the georeferenced basemap. Digital orthophotos provide a reasonable accurate and cost effective method of providing this spatial information providing that the error associated with the individual orthophoto is known. That is the error inherent within each orthophoto and between orthophotos when more than one is used.
Where spatial accuracy is to be greater than that of the available basemap then other methods should be investigated. Satellite imagery with a small instantaneous field of view (IFOV) such as SPOT can also be used to georeference DMSV where it is acquired at similar ground resolution. The spatial accuracy of this mosaic is satisfactory for the purpose of this study. Spatial accuracy is an important consideration for the mosaicing procedure, ground truthing or comparison of other image data sources, for example other vegetation maps.
The time spent in mosaicing of many sub images contributes greatly to the processing cost of DMSV imagery. The approximate breakup of time can be seen in Appendix C. If DMSV imagery was georeferenced during acquisition the mosaicing process could be greatly reduced.
To correctly classify any remote sensing data and images, class lists must be:
1. exhaustive, define all possible classes in the image;
2. definitive, not spectrally overlap other classes; and
3. informative, classes must have some information value.
During class seed selection and refinement it was evident that portions of the images had classes which were not defined in the initial class definitions. One particular portion of the Black Lake chain had an algal bloom at the time of data acquisition. This is evidenced by the high reflectance values in band 4 (NIR) over the water body. This part of the image was defined as algal bloom but it was spectrally similar to Tuart class and this led to further refinement of both classes to eliminate spectral overlap. Similar problems existed for Casuarina/Melaleuca classes. These problems required time consuming revision of classes for all images.
To eliminate classes that were not subjects of the study, the images were subset to include just vegetation that was bounded by the relevant waterbodies (Peel Inlet, Harvey Estuary, Harvey River, Serpentine River and Goegrup Lakes, Black Lakes, and Murray River). To avoid misclassification, the larger water bodies were also excluded from the images. Essentially the images classified, contained only vegetation and some farmland with urban and water from the rivers and lakes. It was not possible to achieve the aim of rapid response and correct classification given the data set acquired for this task. There would appear to be insufficient spectral separation to correctly define vegetation classes as devised for this study. As the classes used are artificial this should not be taken to mean that vegetation classification is not possible, rather that separation of the classes was not achieved as the classes themselves are not spectrally distinct. To achieve classification required the subseting of the images as explained above and this substantially added to image processing time.
It is not always feasible to classify all possible vegetation classes desired. “Weeds” are such a case, where the very definition of a weed is disputable. Weeds can be made up of diverse species from such broad categories as grasses, trees, and flowers. This makes their grouping under such terms as spectral properties impossible for classification purposes as used in this project. Mapping invasive species would require a separate classification process to encompass every possible grouping of flora. It would be a worthwhile extension to this project, given that invasive species will most likely dominate any areas that are disturbed by the changing hydrology.
Colour contrast in an image is very important when making a composite mosaic. The contrast of any imagery used is important in the aesthetic value of the final composite mosaic. Large tonal changes between image segments detract from any visual interpretation. This is most apparently so where imagery is used that was flown for specific purposes such as high water penetration and the incident solar angle results in hot spots or large shadow effects in areas of vegetation. Where large areas of the image are water and specular effects are apparent then this can detract greatly from the final aesthetic value of the mosaic. This may be avoided where large overlaps are available. The requested overlap for the flight runs does not appear to have been provided for a large number of the images. This has resulted in no overlap between frames in some instances which is unacceptable.
It is evident from the data that some contrast or light balance adjustment was carried out on the instrument during flight and/or between runs. This makes rapid comparison between scenes difficult. This can be attributed in part to differing sun/sensor geometry given the differing flight times and days. This also seemingly occurred during major landcover transitions such as water boundaries. It is particularly evident in the image at Austin Bay. For the purpose of accurate spectral classification, a fundamental premise is that the sensor remains constant in response for the duration of the data acquisition, or that its variation can be determined, such as degradation of the sensor over a long period of time. An adjustment of the sensor between/during runs renders rapid comparative assessment of run images questionable, particularly where the sensor parameters/variation are not known. Where possible calibration targets should be included to determine instrument variations in light balance. The instrument should not be altered during data acquisition. Alternatively, the images should be radiometrically corrected before supplying to the client.
Problems of colour balance between individual images may occur if all imagery is not obtained using the same instrument parameters. For rapid analysis of data, colour balance problems needs to be eliminated. Differing colour balance between images is largely avoided if all images are obtained rapidly without alteration of sensor characteristics. Images acquired on differing dates and times have the further imposition of being obtained during differing albedo levels adding to processing times for data. Whilst it is possible to adjust the images to account for this it requires further image processing which adds to the overall data handling times. Where possible, all data for a particular task should be acquired at the same time with the same instrument characteristics applied, or as soon as possible after. Where data are acquired for the same purpose during different seasons, its collective value as a “representative snapshot” is lost, particularly where season may be determinant in classification outcomes such as vegetation classes. As little time as one month may produce spectrally different signatures from the same vegetation where it comes into flower or is relieved of water stress.
Jernakoff et al (1996) found similar problems
with the DMSV instrument and recommended protocols
to improve the performance of the DMSV instrument.
This study experienced similar problems and so the protocols are reiterated:
· Flight should coincide with sun angles of less than 45 degrees to avoid solar
glint.
· 50% overlap should be used to minimise the effects of sun glint and reduce data
loss due to lack of overlap.
· Calibration targets should be included in every mission and gains and offsets not
altered during data acquisition.
· Camera alignment should be precise to minimise post processing of band
registration.
The DMSV instrument is suitable for rapid acquisition of data, but if used in its existing configuration, it is unsuitable for rapid data classification and quantitative analysis where large areas of land cover are involved, and hence many image frames are required. The instrument should not be used to gather datasets of small ground resolution over large areas. This type of data acquisition leads to massive data sets that are difficult to store and manipulate. The spatial resolution of data acquired for this task was too detailed and indeed may have added to the misclassification of vegetation (mixed pixel effect). Certainly this spatial resolution is unnecessary for synoptic water quality assessment unless the objective is to identify a point source of pollutants in a known locality such as a drain input along a stream line. When compared to SPOT data for the same area, a comparison of costs of data acquisition and compilation to a single scene, lend weight to the argument that acquisition of a single SPOT scene is more cost efficient. This assumes that SPOT data is available for the time period required.
Some cyclic electrical interference was evident in the images particularly those images taken over water. This detracted from classification in some portions of the mosaic where the interference patterns were evident in the pixel class allocation (Figure 11). Whilst it is possible to smooth the image data to remove the interference pattern, this alters the basic value of the data.
The subset images of the water bodies of Peel and Harvey Estuary’s were density
sliced to provide images that could be related to water depth and clarity. Density
slicing revealed the electrical interference patterns and were not particularly
successful in the Harvey Estuary due to large contrast differences. Sources of
electrical interference need to be isolated from the instrument to provide data which is
“clean” of interference patterns.
Vegetation loss due to changing hydrology and other factors has already observed
by catchment management authorities and others in the Peel region (George and Bradby, 1993; McComb
et al, 1995). It is important to the ongoing management of
the waterbodies that a greater understanding of the effect of a changing hydrological
regime on vegetation is gained. This study provides a baseline for the
future comparison of vegetation changes. The study however is limited
in the sense it is by nature synoptic in view. It can be enhanced
by more detailed observations in the transect locations described in Figure
1. Future studies need to utilise the synoptic value of remote sensing
to extrapolate vegetation change which occur in the transect locations. To enhance
the ground truthing of the remote sensing data, these further locations are
suggested as possible vegetation transects:
· Black Lake
· Murray River
· Amarillo
Future classification of vegetation may be enhanced by the use of multitemporal data sets. The existing data set can be merged with future data to improve classification accuracy by increasing spectral separability between the classes within the data. The composite data set would for example then consist of a total of 8 bands for two dates involved.
The objective of any data acquisition should be paramount in the planning of the flight mission. Frequently compromises are reached when incompatible objectives are grouped to reduce costs. The solar angle is critical to classification accuracy as the shadow component can greatly affect spectral signatures. When missions have diverse objectives such as water penetration and vegetation assessment, the solar angle is often incompatible for both objectives. Water penetration requires comparatively low solar angles to reduce specular effects and sunglint. Vegetation assessment generally favours high solar angles to reduce the shadow component of the scene. Where compromises are made, then the end users of classifications and assessments need to be aware that the overall classification accuracy is reduced.
When acquiring remote sensing data it is important to concurrently acquire calibration data. Jernakoff et al (1996) suggest a window of 30 minutes for water quality assessment. For large diverse areas such as the Peel-Harvey then it may require the concurrent use of a number of ground parties to gather calibration data, particularly when water quality information is required given the large water body. Close coordination of the flight and ground parties is manifestly important in achieving a successful outcome.