Previous PageTable Of ContentsNext Page

Hyperspectral remote sensing for the provision of land & water management data

Sedat A. Arkun, Iain J. Dunk and Stephen M. Rans

Ball Advanced Imaging and Management Solutions (AIMS)
193 South Road
Mile End, South Australia, 5031
+61 8 8354 3300 , +61 8 8354 3355
Email address:,


Ball Advanced Imaging & Management Solutions (AIMS) has acquired and processed CASI-2 airborne hyperspectral imagery of the Upper Onkaparinga River Catchment region of the Mount Lofty Ranges, 30 km east of Adelaide, South Australia, for the purpose of land-cover classification and hence land-use identification. The region comprises a variety of agricultural, horticultural, viticultural and broad scale grazing activities, existing alongside a number of native and plantation forests. The project was undertaken to determine the utility of high resolution CASI-2 hyperspectral imagery in providing a cost effective and timely provision of information towards strategic planning and management of the land and water resources in the greater Mount Lofty Ranges.

The acquired hyperspectral image data was processed for geometric, radiometric and atmospheric corrections and submitted for advanced processing for land cover feature extraction. Extensive ground truthing was also undertaken to assess the accuracy of the final classification.

This project demonstrates the capability of high resolution remotely sensed image data in mapping the spatial distribution, type and extent of land cover classes in an environment of intensive land use. Furthermore, the atmospherically corrected image data can facilitate the rapid detection and assessment of land-use change and hence enable auditable monitoring and more effective management practices to be enacted in a timely fashion. The technology can be readily adapted for land cover classification of different agricultural regions within Australia or indeed around the rest of the world.


As part of a study for the application of airborne hyperspectral image for the provision of land and water management data, Ball AIMS collected imagery across the Upper Onkaparinga River Catchment area of the Adelaide Hills. The image data was processed for geometric, radiometric and atmospheric corrections, then advanced processed for land use classification, identification and feature extraction.

High resolution remotely sensed image data can provide valuable information about the water catchment areas in peri-urban regions by mapping the catchment condition indicators. This is especially true when the classified image data is integrated with ancillary GIS data, which can either be gathered from existing sources or can be collected according to requirements.

The classified image data provides the necessary information regarding the types of crop and vegetation covers, irrigated areas (i.e. pasture), dams and other water bodies in raster or vector format. This information can then be used in determining:

  • Irrigated and unirrigated cropped areas and related water usage,
  • Capacity, area and location of farm dams,
  • Potential sites for pesticide, sediment, nutrient and biological pollution,
  • Changes in land and water use by temporal comparisons.

The image classification was based upon target features determined by Ball AIMS with input from the EPA (SA). The classification features are listed below:

1. Orchards – Pome Fruit and stone fruit

2. Vineyards – Grapevines

3. Vegetables – Potatoes

4. Grazing Lands – Pasture (Irrigated and Unirrigated)

5. Native Vegetation – Eucalyptus

6. Pines – plantation

7. Brassicas – Cauliflower, Broccoli

8. Farm Dams

For this project, Ball AIMS proprietary techniques and other processing methods were applied to the CASI-2 airborne hyperspectral image data. A number of different methods were used for final land cover classification. This was necessary since there was no single technique that provided a comprehensive and accurate classifier for identifying and extracting all the required features. The overall approach was to ‘build’ the final classification layer by extracting features on a layer-by-layer basis and then to union these into a single layer of land-cover classification.

It was necessary to apply atmospheric correction to the image data before processing to enable; a) temporal image data comparison and, b) the application of spectrographic methods of image processing that require the removal of atmospheric effects from the image data.

Image analysis and interpretation, and hence classification, is an integral part of an iterative process. Initial classifications are checked against collected ground data and subsequently re-processed, with more input derived from the ground truth data (site specific as well as spectral data), to produce more effective and accurate classification results.

Remote sensing imagery forms a coherent two dimensional spatial coverage of a snapshot of the current conditions of a given landscape and is a powerful complement to in-situ based sampling of the various land uses and their condition. The resultant image data is in digital format for rapid inclusion into a Geographic Information System (GIS) to provide levels of information which have only recently been made available to decision makers (Gulinck et al. 2000).

This paper provides a description of the methods employed for image acquisition, base and advanced image processing resulting with a land cover classification of the trial site. The classification results for each of the types are discussed in turn with appropriate levels reference made to imaging and classification issues influencing the final outcomes and achieved accuracies.

Data Acquisition Methodology

The image data was captured on the 24th of February 2000 under clear skies between 11:04 am and 12.47 pm (CSST) covering a site of 8800 ha in area. The site dimensions are approximately 9.6 by 9.2 km square. The CASI-2 image mosaic is shown in false colour in Figure 1. The CASI-2 sensor was programmed to capture imagery using a configuration of 30 bands ranging between 419nm and 950nm, and bandwidths of 3.0 nm to 5.0 nm full width half maximum (FWHM).

The bands that were selected up to 800 nm were specifically configured for land use mapping using a combination of CSIRO and Ball AIMS derived band sets. The bands beyond 800 nm were added for other unrelated research work. The bands were selected to optimise the detection of general vegetative pigment and chlorophyll reflective/absorptive properties, whilst also providing the capability to detect non-vegetative matter such as soils and water.

The image acquisition was performed at two flying altitudes. The full extent of the study area was flown at an altitude providing a 3.5 metre ground resolution. Four image strips were then acquired at a lower altitude providing 2-metre resolution imagery. The 2 metre data were acquired for the purpose of classification comparison and assessment against the 3.5-metre data to evaluate the effectiveness of applying similar methods to different resolution imagery (both spectrally and spatially).

Radiometric Correction

The CASI hyperspectral sensor is a calibrated system where the values for the calibration parameters (gains and offsets) for each CCD are determined during an instrument calibration process that is conducted on an annual basis.

Post flight image processing utilises these calibration parameters to radiometrically correct image data. This enables consistent image interpretation where the radiance values for each pixel for each band are represented in quantities of spectral radiance units (μW cm-2 sr-1 nm-1).

Figure 1: False colour image mosaic of the study region

The other advantage of radiometrically correcting image data is that it enables the application of effective atmospheric correction. This is a crucial factor when considering the image data for temporal (i.e. periodic/seasonal) comparisons. The image data was also colour balanced to facilitate a radiometrically seamless mosaic (i.e. no discernable brightness variations between image strips).

Geometric Correction

The CASI-2 includes the Applanix POS/AV 310 high performance position and orientation system in addition to the standard vertical Gyro. This system comprises the following:

  • Single frequency Novatel 3151 GPS card & antenna for position measurement,
  • Inertial Measurement Unit (IMU) for attitude measurement and inertially aiding position measurement,
  • Instrument Control Unit for data conditioning & recording,
  • Laptop computer for instrument control, and
  • Suite of post processing software.

CASI-2 image geo-correction process involves the extraction of the aircraft roll, pitch, yaw and GPS data. Proprietary post processing software is used to differentially correct the GPS position data. This data is then combined with the strap-down inertial navigation IMU data using Kalman filter algorithms to produce a smoothed best-estimate of trajectory (SBET) solution for the location of the CASI-2 sensor head during flight. The system provides dynamically accurate, high-rate measurements of the full kinematic state of the aircraft. The expected positional accuracy of CASI imagery as a result of the geo-correction process is typically 3 pixels RMS when a digital elevation model (DEM) is incorporated for ortho-correction or the terrain captured is flat (ITRES 1996).

To verify positional accuracy of the image mosaic a number of ground control points (GCP) were surveyed during the field data collection using a differential GPS (DGPS) unit resulting in GCP positional accuracies of 2.0 m. The 3.5 metre image data coordinates were compared against the surveyed GCP’s and were found to have the RMS values shown in Table 1:

Table 1: The image geometric accuracy assessment against DGPS surveyed GCP’s.

Proj : UTM, Zone 54 South

Datum: WGS-84


Total RMS


1.50 Pixel



East (MGA)

North (MGA)

Image X (Pixel)

Image Y (Pixel)

Predict X

Predict Y

Error X

Error Y





























































































More ground control points were subsequently surveyed providing a GCP distribution that is more even and extensive across the image. The second set of GCP’s were surveyed using a non-differential GPS unit. These measurements were performed after the selective availability (SA) was disabled for GPS signals by the US DoD. This resulted in the achievement of positional accuracies of 4-6 metres RMS (verified against known precise ground coordinates). The second set of GCP’s were also used to assess the image mosaic positional accuracy extending across the rest of the imagery, hence enabling the evaluation of the overall geocorrection and georeferencing achievable by the CASI-2 system. The assessment results of this set of GCP’s are shown in Table 2.

Table 2: The image geometric accuracy assessment results against non-DGPS surveyed GCP’s.

Proj : UTM, Zone 54 South

Datum: WGS-84


Total RMS


1.71 Pixel



East (MGA)

North (MGA)

Image X (Pixel)

Image Y (Pixel)

Predict X

Predict Y

Error X

Error Y





























































































Atmospheric Correction

In strategic data analysis of imagery captured by an imaging spectrometer such as CASI-2, it is most critical that the image data is converted to reflectance or surface leaving radiance so that individual spectra can be directly compared to measured laboratory or field data for identification (Ferrand at al. 1994).

It is also envisaged that the classified image will be used for change detection and monitoring purposes, which typically involves comparison against other images. For an image to be used for temporal comparison it is essential that adequate atmospheric and radiometric corrections be applied to bring all scenes to a common radiometric datum. This way spectrographic methods can be applied to all imagery providing the basis of comparison in the detection of changes in the type and extent of land use for a common scene.

There are various methods available for the removal of radiometric effects due to atmospheric scattering and path radiance. A robust atmospheric correction process would require the measurement of several atmospheric parameters (e.g., temperature, pressure, wind, humidity etc.) characterizing the atmospheric conditions prevalent during image acquisition enabling the application of an atmospheric removal program. This method is very time consuming both in the field as well as computationally and is usually a costly option.

A method, offering a logistically simple means of generating acceptable estimates of surface reflectance (atmospheric correction) is the Empirical Line Method (ELM) widely used in airborne remote sensing operations. This method requires the ground placement, or on site identification and spectral measurement of, at least two spectrally flat targets (bright and dark), which can be used as calibration features. The geographic positions of these objects are also determined in the field to enable image identification and location. The spectral characteristics of these targets, which are free from the effects of the atmosphere, are then measured with a field spectrometer. The same target features are identified in the image data and the field measured spectra are then enforced on the radiometrically corrected image data to perform a low cost atmospheric correction (Smith & Milton 1999).

For this exercise two calibration surfaces were used for the ELM atmospheric correction consisting of a white panel made of a number of dropsheets laid out on the ground in a paddock and the surface of a large unmarked bitumen car park in the nearby township of Lobethal providing the dark object. The atmospheric correction results were checked by comparing corrected image spectra of known targets against measured ground spectra.

Classification Methodology

The radiometrically and atmospherically corrected, colour balanced CASI-2 mosaic imagery was classified using a multistage processing methodology including supervised classification using the matched filtering technique and the maximum likelihood decision rule as well as contextual classification using transformed textural bands. The advanced image processing was conducted using the ENVI (Research Systems Inc.) image processing software. Atmospherically corrected optical bands, transformed textural bands, variance and co-occurrence bands were used as input to the maximum likelihood algorithm. Training sites were strategically selected from regions of the imagery to encompass the spectral characteristics and variability for each feature of interest (Richards 1993). These regions were typically sites that were visited during the ground truth data collection. The field data is assessed before deciding whether a site is suitable for inclusion in a training set. The overall objective of the training process is to assemble a set of statistics that describe the spectral response pattern for each land cover type specified for classification. To yield acceptable classification results, training data must be both representative and complete in terms of ground cover classes.

Following this stage of classification, spectrally separable classes consistent with a single land use were combined into single land use classes. Subsequently, brief field verification was undertaken and training sites re-selected prior to a final classification process to complete the exercise.

The water bodies, potatoes, orchard crops and brassicas were extracted separately through a mixture of maximum likelihood classification, spectral matching (using collected ground spectra and image spectra as reference), textural image transformation and band ratio/band math application. Following this process the individual and collective classification layers were combined to provide a single overall classification image – see Figure 2 below.

Figure 2: The composite land use classification image

Irrigated Pasture

The irrigated pasture classification proved to be quite straight forward but with unintended commission of other healthy grasses that were generally found in such regions as creek beds, orchard and vineyard inter-rows and town ovals etc. This classification was made possible by using a combination of spectral matching and decision rule methods.

Un-Irrigated Pasture

Unirrigated pasture was partitioned into three categories:

  • 20% green cover to 80% grass litter,
  • 30% green cover to 70% grass litter, and
  • 45% green cover to 55% exposed soil.

Green cover to grass litter ratios and exposed soil were quantified from field photographs taken of sample quadrats in three separate paddocks having different stages of health. The un-irrigated pasture was classified using supporting ground data and careful selection of training sites.


As with the irrigated pasture, the potato classification was uncomplicated providing a good level of discrimination against other classes except for some confusion with that of irrigated pasture, due to respective similar spectral signatures (Figure 3), as well as the presence of some quantities of grasses and weeds in the potato plantations.

Figure 3: Image spectra of Potato vs. Irrigated pasture

The potato classification was also made possible by using a combination of spectral matching (image and ground spectra) and decision rule application.


Mature orchards were successfully classified, while immature or sparsely planted orchards were susceptible to misclassification as irrigated pasture due to the spectral integration of the cover grasses commonly found between the rows of trees, with that of the orchard trees. This problem is prevalent even through the dry summer season since the majority of the orchard plantations were found to have above ground sprinkler/irrigation systems providing a healthy ground cover. It could be possible to minimise the misclassifications with more comprehensive and representative ground data collection and applying more refined un-mixing strategies. However, this would require the input of substantial effort in modelling the compositional presence of the various mixtures of target material commonly found in orchards. This approach would also have to take into account the orchard morphology and its influence in orchard classification both statistically and spectrally, further complicating the classification picture. Also, the ground cover proportion of grass to orchard canopy varies widely and can be as much as 3:1 in some instances - especially in the new orchards where the trellising methods encourage a vertical canopy structure, effectively reducing the canopy size against the background.

The combination of these factors lead to reduced spectral variability not only between the different orchard crop types but also orchards from other land cover types. This is complicated further by the different maturity states of the orchards in the area. In the case of the pear plantations it was observed on a number of sites that two or more varieties are planted as alternate rows in each block contributing to the spectral confusion. However, this form of plantation generates a high frequency data that associates spatial properties with pixels (i.e. rows of trees) which lends itself for analysis using spatial context (textural analysis). Hence a number of texture bands were created and added to the VNIR band set.

For the reasons given above an un-mixing classification strategy was not employed for this exercise, instead a combination of decision rules, texture analysis and filtering methods was used, which provided a good classification result in discriminating and mapping the orchards as a general land cover class.


Vineyard classification proved to be quite a challenge due to the presence of a number of spectrally complex conditions. These range from the level of maturity of plantations (affecting overall visible canopy size) to the many varieties of weeds and grasses, usually co-existing in different proportions commonly found in the inter-rows complicating the spectral feature space. When these are in turn combined with a variety of exposed soil types, the spectral definition of a ‘vineyard’ becomes quite complex that often can become confused with many other features in the overall image scene. Also, although the vineyards are planted in distinct spatial patterns (vinerows) the use a contextual classification technique was not suitable as described for the orchards since it was not possible to distinguish these spatial patterns in the 3.5 metre image data.

Since the combination of all of these factors produce a wide variety of “vineyard” spectral responses, it was necessary to utilise a number of different training sites to capture this variation and hence maximise classification whilst minimising omission during processing. Spectral matching field observed vine signatures proved to be only marginally useful due to the difficulties mentioned above and the relatively low spatial resolution of the imagery (3.5m), which is not adequate to capture the vine canopies alone. As briefly explained in the section on orchard classification, for spectral matching and un-mixing techniques to be successful, either the target surface must have a homogeneous ground cover across a site – leading to “pure” pixels in the image data, or we must have comprehensive knowledge of the proportional mixes of materials contributing to total pixel radiance. By establishing the proportional quantities of these materials within a ground area matching that of the image pixel, it is possible to model the expected ground pixel reflectance and apply these to the image data to extract the matching pixels. This then enables the identification of surface targets that consist of a mixture of ground covers.

The pre-classification ground truth exercise for this trial project was brief with respect to the total project size and scope, hence the data collected, although sufficient for some target surfaces, were not always comprehensive enough for this method to be applied to all target classifications. Regardless of these factors the classification of established vineyards (i.e. mature) was quite successful with some commission of non-vineyards evident. The commission occurred across the imagery mainly in regions consisting of a mix of bare ground, weeds and grasses commonly found on roadsides, along streams and creeks, hobby farms and abandoned strawberry plantations etc.

Water Bodies (Dams)

The determination of the number and extent of dam formations across a water catchment area is an essential exercise in order to properly model and assess the catchment indicators (state, condition, pressures etc.).

The classification of the water bodies, which in main consist of farm dams of various sizes, was problematic because of the vast spectral variability found between the dams. The reasons for this lie in the fact that the “water” signatures are significantly influenced by the presence various levels turbidity, algae and vegetation (i.e. reeds) found in the water samples. Careful examination of the 30 band hyperspectral data showed that it was possible to isolate distinct spectral features at particular regions of the spectrum (visible and NIR) that were common to all water bodies regardless of their physical composition of silt, algae or vegetation. This necessitated the development of a completely different approach of classification than those used for the other features, consisting of a combination of band math, band ratio and filtering operations.

Hence, through a series of these processes it was possible to extract the water bodies from the image scene. The hyperspectral nature of the image data was central to the development and application of these techniques in water identification.

Scope for Upscaling the Methodology to Greater Mt Lofty Ranges

It is possible to apply the techniques developed and applied for this trial project using CASI-2 hyperspectral imagery for large scale land use mapping of the greater Mt Lofty Ranges or any other similar regional landscape since the methods are fundamentally the same in their application for general land cover mapping. The findings of this trial project encourage further research and development in refining and enhancing the classification processes described so far, not only for land use classification but also for enabling the assessment and monitoring of the physical as well as bio-physical conditions of the features identified.

2m vs. 3.5m Image Data Classification

Two strips of image data were captured at 2m ground resolution with the same 30 band set (FWHM and band centres). The 2m mosaicked image data was then classified using the same techniques used for the 3.5m data. Since the 2m mosaic only contains a small proportion of the spatial extent of the 3.5m trial area, a direct classification comparison was limited to those regions common to both images.

The resultant classified image shows distinct improvements for most ground covers as would be expected. The cauliflowers, which occur in plantations of long strips, classify markedly better with the 2m data due to the improved spatial resolution which in turn improves the spectral purity of the image pixels - see Figure 4 below.

Figure 4: Cauliflower classification shown for the 2m image data (left) and the 3.5m image data (right). Notice the improved distinction of the rows of cauliflower for the 2m data

The cauliflower rows in the above images can be seen in the photograph shown below in Figure 5.

Figure 5: The cauliflower plantation amongst lettuces and potatoes. The cauliflowers are the centre two rows with potatoes planted to the right and lettuces on the left of the cauliflowers.

Improved strawberry classification was also observed for those strawberry fields where heavy to medium heavy infestation of weeds were observed during the field visit. This again is a result of the higher spatial resolution improving the spectral purity of the pixels, thus enabling more discriminate classification. The strawberry classification for the same site in the 3.5m image data was only marginally successful where significant proportions were classified as a mixture of vines, weeds, grass, potatoes, orchards etc.

Unsurprisingly, spectral matching also performed better on the 2m data than the 3.5m data for some target classes (i.e. irrigated pasture, potato, cauliflower) again due to the improved spectral variation. The 2m data, however did not improve the ability to discriminate the vines from other background classes (weeds, grasses, soils etc.). This was expected since the vine row spacing is generally 3 metres with an average canopy width of around 1.0m resulting in a canopy to ground cover proportion of 1:2. The 2m data does not provide enough additional detail over the 3.5m data to resolve the spectral confusion.

There is unquestionable advantage in the utility of 2m image data over the 3.5m data for land cover classification as well as general quality of definition for mapping. However these benefits must be assessed against a set of requirements and budgetary constraints since the cost of acquiring and processing 2m data is more than for 3.5m data.

Classification Accuracy Assessment

At the completion of the classification process it is necessary to perform an assessment of the accuracy of the final classification results. This process enables a degree of confidence to be attached to the end results and provides the means to determine whether the analysis objectives have been achieved (Jenson 1986).

The composite land cover classification was in the first instance consolidated by performing a majority analysis allocating spurious pixels within a large single class to that class. The final land cover classification accuracy was then assessed by calculating a confusion (contingency) matrix using ground truth information. In all, 321 regions of interest (ROI) were created covering field sampling sites visited during the project. These were used as ground verification data against which the classification was compared (see Figure 6). Although statistically desirable, the ground truth sites were not selected using a strict random selection method since the majority of the study site is not easily accessible (considerable effort and time is required for access requiring prior arrangement with the vast number of landowners). However this is largely offset by the sheer number of sample sites used. The result of the confusion matrix calculation is shown in Table 3.

The report shows the overall accuracy, kappa coefficient, errors of commission (percentage of extra pixels in class), errors of omission (percentage of pixels left out of class), producer accuracy, and user accuracy for each class. Producer accuracy is the probability that a pixel in the classification image is put into class X given the ground truth class is X. User accuracy is the probability that the ground truth class is X given a pixel is put into class X in the classification image (Richards 1993).

Table 3: The result of the contingency matrix calculation for 11 feature classes.

Figure 6: The classification verification ground truth sites (there are 321 shown in blue) overlaid on an aerial photograph.


The land cover classification demonstrated in this paper shows that advanced spectral analysis methods in combination with traditional analysis methods can be effectively used for detailed land cover, and hence land use classification of the Mount Lofty ranges of South Australia using high resolution CASI-2 hyperspectral image data. The inherent spectral dimensionality of this data has enabled much better characterization and hence higher “separability” of the types of features captured in the imagery. This is reflected by the accuracy assessment results shown above.

There is still scope for further enhancement of the classification processes to improve the individual classification accuracy of complex features such as orchards and vineyards. The inclusion of an airborne SAR data in the analysis process would significantly enhance the detection of the morphology and hence the identification of the different varieties of orchards due to their unique canopy structures. Indeed this holds true for many other features commonly found in this type of intensive agricultural / horticultural landscape.

In addition, the radiometrically corrected CASI-2 imagery enables the application of atmospheric correction, resulting in image data ready for temporal analysis using existing processes as well as spectral libraries. This facilitates the capability to rapidly map and track change detection, make change assessments (Howard & Wickware 1981), and perform auditable monitoring, thus enabling more effective management processes to be enacted in a timely fashion.


Ball AIMS acknowledges the assistance and collaboration of the South Australian Environment Protection Agency (EPA) in providing valuable input and feedback and the CSIRO for their assistance during data acquisition for this project.


Ferrand, W.H., Singer, R.B., and Merwnyi, E. (1994). Retrieval of apparent surface reflectance from AVIRIS data: a comparison of empirical line, radiative transfer, and spectral mixture methods. Remote Sensing of the Environment, 47, 311-321.

Gulinck, H., Dufourmont, H., Coppin, P., and Hermy, M. (2000). Landscape research, landscape policy and Earth observation. . International Journal of Remote Sensing, 21, 2541-2554.

Howard, P.J. and Wickware, G.M. (1981). Procedures for change detection using Landsat digital data. International Journal of Remote Sensing, 2, 277-291.

ITRES Research Ltd. (1996). Compact Airborne Spectrographic Imager software manual.

Jenson, J.R. (1986). Introductory digital image processing. Prentice-Hall, Englewood Cliffs, New Jersey.

Li, R. (1998). Potential of high-resolution imagery for national mapping products. Photogrammetric Engineering & Remote Sensing, 64, 1165-1169.

Richards, J.A. (1993). Remote sensing digital image analysis: an introduction. 2nd ed. Springer-Verlag, Berlin.

Smith, G.M. and Milton, E.J. (1999). The use of the empirical line method to calibrate remotely sensed data to reflectance. International Journal of Remote Sensing, 20, 2653-2662.

Previous PageTop Of PageNext Page