Abstract

Background Information

Over the past 10 years numerous studies have been carried out incorporating the use of airborne remote sensing systems. Digital Multispectral Video is fast becoming the forefront of airborne systems and has proved to be an accurate and effective tool providing immediate, high-resolution results at a realistic cost. Airborne videography emerged in the 1960s and early 1970s, with works by Robinove and Skibitzke (1967) and Mozer and Seige (1971), but it was not before the 1980s that main research and operational applications in mapping and resource management began (Mausel et al., 1992; King, 1995). Airborne videography has been tested and applied for detection and mapping of crop types and conditions, including weed infestation and diseases, and the effect of soil salinity (Richardson et al.,1985; Wiegand et al., 1994; Cloutis et al., 1996; Wiegand et al., 1996). Other applications in agricultural areas include identification of site-variability in cropped areas (Yang and Anderson, 1996; Moran et al., 1996). These works have demonstrated the potential of airborne videography as a non-invasive and rapid method of generating geo-referenced maps that help, for instance, the control of GPS-equipped mobile fertilising units, or the identification of within field variability. Forestry applications of videographic data comprised prediction of habitat complexity (Coops and Catling, 1996); mapping forest species (Everitt et al., 1986; King and Vlcek, 1990, Thomasson et al., 1994), detection of signs of stress and structural damage in natural and revegetated areas (Levesque and King, 1996 and 1999).

Aims and Objectives

This study aims to visually compare Digital Multispectral Video images with Landsat TM and aerial photography to determine their appropriateness for delineation of remnant vegetation.

Determine whether Digital Multispectral Video data can delineate between landuses using classification methods within INTERGRAPH Image Analyst software.

Determine whether Digital Multispectral Video data is an appropriate remote sensing technique for the detection of canola, pasture, remnant vegetation and wheat.

Study Area

The research project was carried out within an area known as the Toolibin Lake Catchment. The Toolibin Lake Catchment covers 44,760ha and consumes a majority of the shire of Wickepin, which forms part of the South West of Western Australia. The catchment lies 4 km south east of the township of Wickepin and 240 km south east of Perth.

The climate consists of warm to hot summers and a cool wet winter. Rainfall varies from 375 mm to 425 mm from east to west of the catchment. The catchment lies within the ‘Zone of Ancient Drainage’ described as “broad flat valleys of low gradient with salt lake chains at their lowest point, gently sloping valley sides, some rock outcrops and large areas of yellow sandplain”. Small stands of native vegetation occupy 14.6 per cent of the catchment (Baxter, 1996).

The dominant landuse is dryland agriculture based on crops (wheat barley, oats and lupins, while canola is being planted more frequently) and pastures mainly for sheep. Toolibin Lake Catchment has been used extensively for research purposes including the Toolibin Catchment Revegetation Manual (Baxter, 1996) thus providing a library of applicable land degradation and revegetation issues.

The Digital Multispectral Video (DMSV) data was acquired over a 2km-band running 17km in a Northeast direction from Lake Toolibin. The area was chosen for it’s diverse agricultural and natural land uses, geophysical and hydrological features.

Figure 1. Study Area Location.

Data Sets

Remotely Sensed Data

The SpecTerra Systems Digital Multi-Spectral Video system (DMSV) is a four CCD camera imaging system designed primarily for the acquisition of high resolution digital multi-spectral imagery of terrain, vegetation, water bodies and coastal environments (Honey, 1995).

The four CCD cameras record 578 lines of 740 pixels per line and the system is capable of recording up to 70% stereo overlap. The resolution (pixel size) and total frame size are dependent on the focal length and flying height of the airborne system at the time of data acquisition. By incorporating optional logging of GPS navigation and time data the images can be positioned accurately providing the ability to cross reference field survey information.

The system was mounted on a small aircraft and the imagery acquired at 10000ft above ground level (11000ft above sea level) giving a 2m pixel resolution. Interchangeable narrow band-pass interference filters were used to generate images for Blue-band 1(center wavelength 450nm), Green-band 2 (center wavelength 550nm), Red-band 3 (center wavelength 650nm) and NIR-band 4 (center wavelength 750nm).

The DMSV data- was acquired on the 16 of September 1998 clear skies prevailed over the Lake Toolibin Catchment. Images were captured for two runs 17km in length, the initial at a compass bearing of 45^o and the final 225^o. The images were geo-referenced to an orthorectified aerial photography of the area during the mosaicing process.

Aerial Photography- the aerial photo mosaic consists of 1:25,000 color photos captured on the 13 of October 1996. These were scanned at 400dpi and rectified to a digital cadastre database using mostly 3^rd order polynomial transformations (a few 2^nd order) and mosaiced with a 3 meter pixel resolution (Robinson, 1997).

Landsat TM- bands 5,4 and 2; acquired on the 3^rd of February 1996 with a 25-meter pixel resolution was displayed in red, green and blue respectively.

Field Data

The field data was collected on the 16^th and 17^th of September coinciding with the DMSV flight. Color Aerial Photography at 1:25,000, which was captured on the 13 of October 1996 (the same photos used to generate the Aerial Photo Mosaic) along with an analogue copy of the cadastre was used for geographic location purposes and land use data recording.

Method and techniques

Ground data collection

During the field inventory the visible (some areas were difficult to access due to boggy conditions and a 2WD vehicle) land uses characteristics (e.g. crop type, remnant vegetation and lakes) across the entire study area were recorded and color photographs were taken using an Olympus automatic camera. The position and direction of the photos were recorded and used for reference during analysis.

The land uses intensively investigated included canola, lupins, wheat, pasture and pasture on stubble, that is, paddocks harvested in season prior to 1998. At the time of the DMSV data acquisition, most of the crops were in early flowering. The attributes recorded for these land uses were location, height, maturity, and dry weight equivalent (DWE).

The location of the elected land use was referenced to the aerial photography from 1996. Following this, the height and maturity of the land use was documented by removing a plant and measuring it against a 1-meter ruler, where a color photo was taken. An agronomist then assessed the maturity status with reference to the photographic information.

To determine the dry weight equivalent (DWE), 1/10^th m² quadrant was placed over three or four sample locations within a paddock in order to gather an average DWE for that land use. All green matter within the quadrant was cut at ground level with a scalpel, while trying to avoid gathering excess dirt and animal manure as this would need to be removed prior to drying. The samples were placed in a bag and the vegetation density noted for each sample (by eye in comparison to the paddock average that is high, medium or low). The samples were brought back to the work shed and washed thoroughly by rinsing in clean water, spinning dry and manually picking out erroneous matter. The clean samples were then placed in ovens and dried over 1 or 2 days. The dry vegetation samples were weighed and each weight recorded in grams per 1/10^th m², and converted to an average tonne/hectare, as shown in Table 1. (Bryant, 1998).

Table 1. Land use samples and their associated Dry Weight Equivalent.

LANDUSE & Maturity	Vegetation Density	DWE g per 1/10th m2	Tonne/Ha	Average T/Ha
Lupins: Early flowering	Low	22.06	2.2	4.4
	Medium	40.82	4.1
	High	69.85	7.0
Canola: Early flowering	Low	9.68	0.97	2.8
	Medium	30.77	3.1
	High	49.24	4.9
Wheat: Early flowering	Low	63.34	6.3	7.2
	Medium	63.68	6.4
	High	90.95	9.1
Pasture on land cropped prior to 1998	Low	7.9	0.8	2.9
	Medium	24.66	2.5
	High	36.41	3.6
	Very High	46.25	4.6
Pasture	Low	24.96	2.5	4.2
	Medium	43.82	4.4
	High	56.84	5.7

Field data was collected concurrently with the time of flight, enabling the ground data to be subsequently used to generate realistic analysis during digital image classification.

Visual interpretation of remnant vegetation

Because of its large resolution (2m), the DMSV image was used as a reference to create a common boundary for the remnant vegetation area calculation. All three images (DMSV, Aerial Photography and Landsat TM) were then referenced to the common boundary and the area of remnant vegetation determined for each image by screen digitizing. A simple methodology was implemented to ensure the same technique was employed for the three images.

• Using place smart line within MicroStation remnant vegetation areas were lassoed.

• All Lake Toolibin and Lake Walbyring were included as remnant vegetation as there is an abundance of tree and scrub species that lie within these lakes which are of very high conservation value. Other waterways and dams were excluded from the area,

• If any two areas of remnant vegetation were more than 20 meters apart two different lassoes were made,

• Any single tree or small group of trees with a diameter equal or larger than 20metres was included as remnant vegetation,

• All roadside corridors were included as remnant vegetation,

During the screen digitizing visible minor shadows were included, while visible large areas of shadow were excluded.

The approach was applied on the data sets as follows:

DMSV (1998, 2m resolution): the NIR band was viewed in black and white and reference was made to the Red, Green and Blue bands viewed as a color composite image. The black and white NIR band provided the best enhancement for detection of remnant vegetation.

Aerial Photo Mosaic (1996, 3m resolution): the Toolibin.tif image was viewed in color, which provided an adequate background for delineation between remnant vegetation and other landuses.

Landsat TM (1996, 25m resolution): a color composite of bands 5,4 and 2 displayed as red, green, and blue respectively. The color contrast was enhanced using a Linear Clip (Intergraph, 1999) of each band independently. This provided the best distinction between remnant vegetation and other land uses. The areas of remnant vegetation computed for each image are shown in Table 2.

Table 2. Area of Remnant Vegetation.

Image Type	Total Area (ha)	Remnant Vegetation (ha)
DMSV	2704.46	1015.20
Aerial Photo	2704.46	1010.00
Landsat TM	2704.46	1078.04

Digital Classification of the DMSV

With detailed field data collected at the time of flight it was possible to perform supervised digital image classification. Four methods were tested: Minimum Distance, Maximum Likelihood, Parallelepiped and a Density Slice of the Normalized Difference Vegetation Index (NDVI).

Supervised Training

To generate the Minimum Distance, Maximum Likelihood and Parallelepiped classified images 10 training sites were selected within the four known landuses (classes) namely, Canola, Pasture, Remnant Vegetation and Wheat.

A composite image of the Blue, Green and Red bands was enhanced using a linear clip, and the 40 training sites were selected by the Region Growing method (Intergraph, 1999). The Region Growing method allows the user to define a seed point on the image and the neighbouring pixels that will be included in the training site if certain user-defined criteria are met. Three different methods of training site selection were performed and training set statistics were calculated on the Blue, Green, Red and NIR band. After viewing the statistics for each training set, the method which resulted in the lowest average total standard deviation of the pixel values for each class in each band was chosen, as compact clusters with small standard deviations generally reduce the chances of erroneous labelling during the classification procedure.

The chosen training set used Center Point; (where only the selected pixel on the image become the seed value) with a Pixel Difference, (which allows the user to allocate a numerical value by which a pixel value can vary from the seed value to determine the pixels inclusion) of +/- 5.

Image Classification

When performing image classification, Image Analyst allows the user to work with a chosen training set, select an appropriate classification algorithm and DMSV Bands for computation.

Using the pre-defined training set (CP-5.00.trn) Minimum Distance, Maximum Likelihood and Parallelepiped classification algorithms on all 4 Bands (Blue, Green, Red and NIR) were computed.

The Minimum Distance classifier computes the Euclidean distance (distance measurement of one pixel to other pixels, based on the Pythagorean theorem) from an unknown pixel to the mean vector of each class, and assigns the pixel to the class to which it is closest. For computational efficiency, Image Analyst considers only the pixels within a specified number of standard deviations of a class mean vector for classification. The default threshold is 3 standard deviations from the class mean (Intergraph, 1999).

The Maximum Likelihood method classifies the input pixels into one of the training classes by using the Bayes decision function for normal patterns with an equal class probability of occurrence or a set of apriori probabilities that the user assigns, they may be determined from external sources of information about the scene. The algorithm quantitatively evaluates the variance and correlation of the class spectral response patterns when classifying an unknown pixel. The discriminant function of a given pixel, being a member of a particular class, is computed. The pixel is then assigned to the class that has the largest discriminant function. The default confidence threshold for Null class is 95% (Intergraph, 1999).

The Parallelepiped method classifies unknown pixels based on the category of spectral ranges in which they lie. Once a pixel is identified as being in a class, it is blocked out of any other overlapping classes. Therefor the behavior of the algorithm is influenced by the order of classes specified. The order that the classes are evaluated can be specified using the Reorder option, a variable in Image Analyst. If a pixel lies outside all class ranges, it is interpreted as unknown or Null class. The default Null class threshold value is 3 standard deviations from the class mean (Intergraph, 1999).

The three image classification algorithms were implemented using the default settings mentioned above and the landuse classes displayed in suitable distinguishing colors.

The Normalized Difference Vegetation Index (NDVI)

The NDVI, often referred to as the greenness index, is derived from a ratio of the NIR and RED bands via the algorithm.

NDVI = (NIR-RED)/(NIR+RED)

Performing a complex arithmetic equation using the NIR and RED bands created a NDVI image (ndvi.cpx). Complex Arithmetic is a spectral operand within Image Analyst that allows the user to create complex arithmetic equations with user-defined parameters. The appropriate bands and an extensive list of operators (e.g. +, x, cos, sqrt etc) are available for selection in order to create a desired expression. The expression can be saved for use later.

A new training set (ie. ndvi.trn) was created with a copy of the location of training sites and they allocated parameters from CP-5.00.trn (that is, the training set using a seed value, center point, and Pixel Difference of +/-5.00). The statistics were calculated using the ndvi.cpx created image for the four classes (canola, pasture, remnant vegetation and wheat)

When viewing the statistics it becomes evident that it would be difficult to separate pasture and wheat using the NDVI results, as they posses a similar mean greenness value while canola and remnant vegetation present very different pixel values, as illustrated in Table 3.

Table 3. The NDVI training set class statistics using ndvi.cpx computed image.

CLASS	MINUMUM	MAXIMUM	MEAN
Canola	139	200	173.31
Pasture	169	255	252.80
Remveg	7	97	45.08
Wheat	225	255	254.78

Because of the distinctive NDVI values for canola and remnant vegetation, a density slicing of the NDVI was performed to identify the spatial location of these two land uses. To replace the existing pixel values with a new single value the Image Analyst Tool Replace Values was used. The Replace Values tool allows the user to replace a single pixel value, multiple values, or a range of values with a new allocated value in a selected image.

The full image range of pixel values (0-255) were replaced according to the chosen class intervals, see Table 4, which allowed for accurate classification to be performed.

Subsequently a Density Slice, which allows the user to isolate pixel values and assign them a color, was performed on the creating a new image (densli_ndvi.cpx). This meant that the canola and remnant vegetation areas of the image were clearly highlighted

Table 4. New Replacement Values allocated to Class pixels.

Existing Value	Class	Replacement Value
0-6	-	0
7-97	Remveg	50
98-138	-	0
139-200	Canola	150
201-255	-	0

NDVI Classification

The NDVI Density Slice image (densli_ndvi.cpx) was required to be a classified image to ensure that accuracy assessment could be determined.

Using the CP-5.00 training set and methodology, three training sites within remnant vegetation and canola were selected. The minimum distance classification algorithm was applied on the densli_ndvi.cpx, which produced the resultant classified image (ndviclass.cpx).

Ground truth assessment

To perform ground truth assessment within Image Analyst all the images to be compared must be classified images. Therefore it was required that the ground truth data be converted into a classified image which could subsequently be used to assess the accuracy of the supervised classification results.

The ground truth classification image was created using the information gathered at time of field data collection. The image was reclassed using Replace Values tool and then a Density Slice was performed on the image with the colors chosen to simulate the classification outputs. The class colors are represented in Table 5.

Table 5. Replacement pixel values, Density slice Color and training set name for the ground truth landuse classes

CLASS	Pixel Value	Color
Canola	200	Yellow
Pasture	150	Red
Remveg	100	Blue
Wheat	250	Green
Lupins	50	Orange
Other	25	Purple
Outside	75	Grey
Unclassed	-	Black

Comparing Classified images for Accuracy Assessment

The Compare classmaps command within Image Analyst computes a confusion matrix that compares two classified images (one being the ground truth image) that have been produced with different classification parameters or methods. The confusion matrix is a report that contains a matrix of the number of pixels assigned to each class in a particular classification. All classified images (Minimum Distance, Maximum Likelihood, Parallelepiped and NDVI) were independently compared with the ground data classification and the confusion matrix computed.

Discussion and Results

Delineation of Remnant Vegetation

The three data sources have been ranked according to their appropriateness for visual identification of remnant vegetation in Table 6. The 25m resolution of the Landsat TM image produces coarse accuracy when dealing with the data visually in comparison to the 2m resolution DMSV, and 3m resolution Aerial Photo mosaic.

The three maps, with line work in different colors were overlaid after area calculations were computed. This simple spatial analysis clearly displayed the areas of under and over identification within the Landsat TM scene. The 25m resolution enlarged the lasso boundary for most vegetation regions, while some of the smaller single trees did not exhibit the characteristic tone of remnant vegetation and therefor were not included.

The results of the area computed for Remnant Vegetation (Table 2), highlight the similarity between the Aerial photography and DMSV identification capabilities with only a 5.20ha of difference (greater for DMSV), which can be equated as 0.19% of the total study area. The year of data collection, 1996 for the Aerial Photography and 1998 for the DMSV data must be considered when comparing these results as the vegetation within the DMSV image has a further two years growth. Vegetation patches with a diameter smaller than 20m, which were not included on the aerial photography of 1996, may well have been lassoed in the DMSV image for 1998 due to the age of the vegetation and greater resolution of the data.

Table 6.Image comparison for remnant vegetation delineation.

Data / Technique	Very Good	Good	Fair
DMSV- NIR band in B/W supported by a color composite of Red, Green and Blue.	The 2m-pixel resolution made it possible to selectively identify small single trees and shrub. Re-vegetation strategies implemented throughout the study area were easily depicted and excluded from the calculation. Narrow roads, buildings, thin tree bands, fence lines and dams were recognized.
Aerial Photography- Color photographs 1:25,000 scanned at 400dpi. Displayed as color.	Single trees were lassoed although delineation between trees, shrubs and small buildings increased in difficulty. The color image aided in detection of shadow. The larger scale image 1:25,000, 3m pixels decreased the clarity. Generally an acceptable method for mapping remnant vegetation.
Landsat TM- Band 5, 4 and 2 displayed as Red, Green, and Blue respectively. Color contrast enhancement using a linear clip of each band independently.			The 25m pixels merged landuses together making it difficult to accurately classify vegetation boundaries. Dense over-story put constraints on the inner roads to provide distinct change in reflectance while small trees merged into other landuses. Without considerable knowledge of the area and prior vegetation detection (DMSV and Aerial Photography) legitimate extraction of vegetation would have been extremely difficult. However detection of densely vegetated areas within the lakes would have been possible due to the distinguishing change in reflectance.

Digital Classification Images

A visual comparison of the four classification algorithms with the Ground truth classification was performed. From these a general idea of the suitability of each classification could be drawn. Table 7 highlights some conclusions based on of the visual interpretation.

Table 7. Visual interpretation of the Classification algorithms.

Minimum Distance	Maximum Likelihood	Parallelepiped	Density slice Ndvi
CANOLA Canola area classed very well. Over-classification of pasture.	CANOLA Canola area slightly under-classed. A little over-classification of pasture.	CANOLA Canola area classed O.K. Over classification of pasture.	CANOLA Canola area classed O.K. Major inclusion of pasture and some wheat areas as canola.
PASTURE Pasture area: a lot of underclass. A lot of over-classification of wheat.	PASTURE Pasture area: a lot of underclass. Over-classification of wheat.	PASTURE Pasture area: a lot of underclass. Major over-classification of wheat.
REMVEG Remveg areas: generally good, a lot of under-classification.	REMVEG Remveg areas: generally good, a lot of under-classification.	REMVEG Remveg areas classed O.K. A lot of under-classification.	REMVEG Remveg areas classed well. A fair amount of over classification in pasture areas.
WHEAT Wheat areas are under-estimated. Over-classification of pasture.	WHEAT Wheat areas are under-classed. Small over classification of pasture.	WHEAT Wheat areas grossly under-classed or classed as wheat.

The Lupin areas of the ground truth classification were classed as a mixture of wheat and pasture using the Minimum Distance, Maximum Likelihood and Parallelepiped algorithms. Density Slicing of the NDVI displayed the area as unclassified.

From Table 7, it is assumed that the Minimum Distance and Maximum Likelihood classification are generally suitable, except for the fact that large regions of the study area remained unclassified. The Parallelepiped method had trouble classifying the wheat areas and again a considerable part of the study area was unclassified. A Density slice of the NDVI displayed major over classification of canola and large areas remained unclassified.

Accuracy Assessment

In order to determine the accuracy and suitability of each classification algorithm a percentage value for the number of pixels assigned to each class from the results presented in the confusion matrix was manually computed. Table 8 displays these results while excluding inconsequential values.

Table 8. Percentage of total Ground Truth Pixels assigned to each class.

MINIMUM DISTANCE CLASSIFICATION
	GROUND	Lupins	Remveg	Pasture	Canola	Wheat
MIN_DIST
Canola		0.016	0.01	7.84	82.31	0.14
Pasture		48.69	0.03	27.71	0.58	23.01
Wheat		35.31	0.05	0.37	0.07	38.07
Remveg		0.12	58.03	0.09	0.12	0.04
Null		15.86	41.87	63.97	16.92	38.74

MAXIMUM LIKELIHOOD CLASSIFICATION
	GROUND	Lupins	Remveg	Pasture	Canola	Wheat
MAX_LIKE
Canola		0.00	0.01	3.96	73.78	0.02
Pasture		44.01	0.01	25.33	0.33	13.89
Wheat		26.47	0.01	0.40	0.04	36.23
Remveg		0.04	46.63	0.03	0.02	0.01
Null		29.47	53.34	70.27	25.83	49.85

PARALLELEPIPED CLASSIFICATION
	GROUND	Lupins	Remveg	Pasture	Canola	Wheat
PARALL
Canola		0.01	0.01	7.84	82.31	0.14
Pasture		79.49	0.03	27.974	0.56	50.27
Wheat		4.52	0.05	0.11	0.05	10.81
Remveg		0.12	58.03	0.09	0.12	0.04
Null		15.86	41.87	63.97	16.92	38.74

NDVI CLASSIFICATION
	GROUND	Lupins	Remveg	Pasture	Canola	Wheat
PARALL
Canola		5.77	1.86	44.55	78.72	21.54
Remveg		3.59	74.56	8.19	3.81	7.51
Null		90.63	23.58	47.25	17.47	70.95

Minimum Distance, Maximum Likelihood and Parallelepiped

The resultant statistical percentage values above demonstrate that for most landuses classes the errors of commission (ground truth class, classified as other classes. e.g. GROUND Remveg X MIN_DIST Canola = 0.01%) is insignificant. The lupins were included within the ground truth classification to demonstrate its reflectance similarity to other classes. The results show lupins having spectral values similar to other to pasture and wheat, therefore being labelled as such, regardless of the algorithm used.

The largest errors of commission were shown to occur where the existing wheat was classified as pasture. Also to a lesser, but consequential degree, where existing pasture was classified as canola.

Density Slice of NDVI

The NDVI percentage results produced high accuracy for classifying existing remnant vegetation and canola, 75% and 79% respectively. Most of the remaining areas occupied by these two classes failed to be identified, that is 23% and 17% respectively were unclassified.

The Null Result

The major concern in the four methods of classification is a very high percentage of Null (unclassified) values for all the existing landuse classes. Very high Null classes values demonstrate that the training sets do not properly characterize the heterogeneity. Generally, the most appropriate way to reduce the high Null values would be to collect more training sites within each landuse class. For example, training sites for remnant vegetation could be chosen within classes of dense over-story, aged vegetation, scattered vegetation, etc. Similarly crops (wheat and canola) could be segregated into areas differing in density, across different soil types or degree of weed content.

The training set methodology chosen for all classification algorithms was CP-5.00.trn, which demonstrated the lowest average total standard deviation for each class (eg. homogeneity). This restricted the classification algorithms to search for areas with such high constraints. Perhaps if training set with a higher level of standard deviation was employed; lower Null values would results in each classification. However this may result in higher misclassification errors for the exiting landuse classes.

Conclusion

DMSV images have proved the most appropriate form of imagery for visual delineation of remnant vegetation. The 2m-pixel accuracy made it possible to identify and map small patches of vegetation, and with a suitable enhancement it was even possible to distinguish between trees and shadow.

The DVMS data has successfully delineated between landuses and been capable for detecting, Canola, Pasture Remnant Vegetation and Wheat. The project has highlighted that DMSV remotely sensed images, with a 2m-pixel resolution, provide a useful tool for both visual and digital classification methods. It has also demonstrated that DMSV data possesses the capability to distinguish internal crop variability (e.g. density, yield, etc.)

Recommendations for further investigations

Further research should be conducted to determine a proper way to reduce the number of “unclassified” pixels within the DMSV data. Some suggestions resulting from this study are:

• Create subclasses within a category such as, dense vegetation, scattered vegetation, aged vegetation, etc. These subclasses could be latter merged in one main category e.g. remnant vegetation. In this way, the heterogeneity of the land cover could be characterized, while keeping compact homogeneous clusters during the training phase.
• Perform an increase the amount of statistical analysis in order to determine the most suitable DMSV bands delineation between land cover classes.
• Collect more field data with increased accuracy for the location of samples. Use a GPS for ground truthing purposes.

References

Coops, N. and Catling, P., (1996). Predicting the complexity of habitat in forests from airborne videography for wildlife management. Int Jnal Rem Sens, vol. 18, pp. 2677-2682.

Baxter, A. (1996) Toolibin Catchment Revegetation Manual. Department of Agriculture, Narrogin WA

Bryant, R. (1998) Personnel communication, Methodology for DWE. Department of Agriculture, Narrogin WA

Everitt, J., Escobar, D., Blazquez, D., Hussey, M., and Nixon, P., (1986). Evaluation of mid-infrared with a black and white infrared video camera. Photogrammetric Engineering and Rem. Sens., vol. 52, pp. 1655-1660.

King, D. and Vleck, J., (1990). Development of a multispectral video system and its application in forestry. Canadian Jnal. of Remote Sensing, vol. 16, pp. 15-22.

Honey, Dr F. (1995) Introduction to SpecTerra Systems. Marketing Brochure SpecTerra Systems Pty Ltd, Nedlands WA

Intergraph Corporation (1999) Image Analyst Users Guide January 21^st.Huntsville Alabama 35894-0001

Levesque, J. and King, D., (1996). Semivariance analysis of tree crown structure in airborne digital camera imagery. Proc. of the 26th Int. Symposium on Rem. Sensing of Environment, Vancouver, Canada, pp. 275-278.

Levesque, J. and King, D., (1999). Airborne digital camera image semivariance for evaluation of forest structural damage at an acid mine site. Remote Sens. Environ., vol 68, pp.112-124.

Thomasson, J., Bennett, C., Jackson, B., and Mallander, M., (1994). Differentiating bottomland tree species with multispectral videography. Photogram Eng and Rem. Sens., vol. 60, pp. 55-59.

Robinson, D. (1997) Lake Toolibin-Photo Rectifying and Mosaicing. Department of Agriculture, South Perth WA

Robinove, C., and Skibitzke, H. (1967). An airborne multispectral video system Prof. Paper 575-D. U.SGS, pp.143-146.

Mozer, M., and Seige, P., (1971). High resolution multispectral TV camera system. Proc. 7th Int. Symposium on Remote Sensing of Env., Ann Arbor, Michigan, pp. 1475-1480.

Mausel, P., Everitt, J., Escobar, D., and King, D., (1992). Airborne videography: current status and future perspectives. Photogram Eng and Rem Sens, vol. 58, pp. 1189-1195.

King, D., (1995). Airborne multispectral digital camera and video sensors: a critical review of system designs and applications. Canadian Journal of Remote Sensing, vol. 21, pp.245-273.

Cloutis, E., Connery, D., Major, D., and Dover, F., (1996). Airborne multispectral monitoring of agricultural crop status: effect of time of year, crop type and crop condition parameter. Int. Jnal. Remote Sensing, vol. 17, pp. 2579-2601.

Moran, M., Inoue, Y., and Barnes, E., (1997). Opportunities and limitations for image-based remote sensing in precision crop management. Remote Sens. Environ., vol. 61, pp.319-346.

Wiegand, C., Anderson, G., Lingle, S., and Escobar, D., (1996). Soil salinity effects on crop growth and yield. Illustration of an analysis and mapping methodology for sugarcane. Journal Plant Physiol., vol. 148, pp. 418-424.

Wiegand, C., Rhoades, J., Escobar, D., and Everitt, J., (1994). Photographic and videographic observations for determining and mapping the response of cotton to soil salinity. Remote Sens. Environ., vol. 49, pp. 212-223.

Yang, C., and Anderson, G., (1996). Determining within-field management zones for grain sorghum using aerial videography. Proc. of the 26th Int. Symposium on Rem. Sensing of Environ., Vancouver, Canada, pp.606-611.