Ball Advanced Imaging & Management Solutions (AIMS) has been involved in the acquisition and processing of CASI-2 hyperspectral imagery for vineyard applications since 1999. The image products developed so far provide an additional tool in aiding vineyard managers and growers alike in the application of precision viticulture practices. Atmospherically corrected high-resolution imagery is used to produce a number of image maps. Namely; graded vineyard greenness index, vine-canopy gap identification and quantification, vine-variety classification and identification and Normalized Difference Vegetation Index (NDVI). These image maps can either individually or collectively help vineyard managers and wine-makers alike to make more informed management decisions regarding the high-value vineyard plantations.
The graded vine canopy greenness image presents the observable vine vigour (greenness) variations within each vine block. Separation of a vineyard into discrete management units facilitates better targeting of vineyard practices such as irrigation, fertilization, spraying, pruning and harvesting. The result of targeted management practices through prior knowledge of the location and the condition of a management unit can lead to significant savings in resources, i.e. water, chemicals, fertilizers and effort, as well as providing the means to improve the vineyard uniformity and hence yield and quality consistency. Another significant advantage of the use of such imagery is the ability to harvest segments of a vineyard block at optimum times in terms of ripeness which may be dependent on the condition of a given management unit. Partitioning harvesting times according to vine condition also avoids the mix of higher quality grapes with the lesser ones increasing the overall value of the harvest, which can be significant in the premium wine market.
The application of satellite imagery to precision crop management (PCM) for high value crops is limited due to restrictions in spectral and spatial resolutions, and time-at-overpass constraints. In contrast, high spectral and spatial resolution airborne sensors such as the CASI-2 enable timely acquisition of imagery (i.e. during optimum growth periods) for the assessment of the health and vigour of plantations. The high spatial resolution of the CASI-2 imagery (0.60 m up to 5.0 m in un-pressurised aircraft) also enables the capture of fine features in plantations such as the vine rows in a vineyard. The high spectral resolution, defined by the number of bands and the bandwidths at which the sensor is able to capture the imagery, also provides the means to more readily analyse the health status of plants and vegetation. This is due to the ability to more precisely characterise the spectral reflectance of the biophysical properties of the plants at different stages of phenology.
Crop monitoring is a time critical task due to the dependence of management practices on a given phenological state (Moran et al. 1997). Advances in farm management technology through the global positioning system (GPS) and geographic information systems (GIS) enable the application of variable rate input applications (Pearson et al. 1994). Field implementation has been enhanced more recently with the abolition of GPS selective availability, enabling simpler and more cost effective non-differential GPS methods to be utilised. Variable rate input can be used for the application of fertilisers, herbicides, insecticides, and seeding in general agricultural, horticultural and viticultural practices. To date, the major limitation has been the production of geo-referenced maps of the plantations, defining areas of weed infestation and crop health deficiencies within a management unit. Airborne remote sensing, in conjunction with purpose specific data processing algorithms can provide the means for the production of such maps using captured image data.
From the farm managers’ perspective, such an approach is aimed at limiting the application of various inputs to areas that require them. The objective is to maximise output while minimising input. In some instances over application of inputs may reduce yield (eg. excessive nitrogen application can cause excessive leaf development at the expense of fruit production [Pearson et al. 1994]).
In the viticulture industry PCM can be used to target harvesting strategies according to crop condition. The premium wine market requires grape quality information at harvest to isolate the good quality grapes from the lesser ones during the grape crush and storage. Grape quality assessment is usually done on a block-by-block basis, which, currently does not take into account the variability of the quality of grapes within each block. If this variation is identified and mapped within a given block, then it is possible to apply precision harvesting techniques, and thus improve the overall value of that harvest since the high quality grapes usually carry a value that is substantially greater than the lesser quality for the same quantity.
As another example of the application of PCM, the premium wine market may require information to meet certain export labelling standards limiting the cross varietal mixture of grapes. The mixture of the off-types can be identified and mapped within a vineyard block using remote sensing, and the application of precision harvesting can ensure the varietal purity at the crush by avoiding the harvesting of those rows and sections of vines in that block.
Airborne digital data is acquired using a CASI-2 sensor mounted in a fixed wing twin-engine light aircraft. A band set in the visible-near infra red (VNIR) portion of the electromagnetic spectrum is configured and programmed into the sensor unit. The band set is designed to characterize the spectral properties of the vineyard (this is application dependent and requires prior knowledge about the target in question).
The CASI-2 imagery utilized in this paper was acquired with a ground resolution of 0.70 metres and programmed to capture 18 narrow spectral bands in the VNIR. Ball AIMS used a band set that was specifically designed for viticulture image data capture, with the spectral bands positioned between 447 and 937nm. The bandwidths for the 18 bands ranged between 6 to 11.8nm. This band set was designed in collaboration with CSIRO using existing spectral libraries and field spectral observations to optimise detection of the reflective and absorptive properties of the vine plants in different phenological states.
Following data acquisition, each flight line is radiometrically and geometrically corrected using ITRES and Ball AIMS proprietary software. Radiometric correction converts pixel digital numbers (brightness values) to spectral radiance units by applying pre-determined calibration parameters to the CCD array in the CASI-2 sensor (ITRES 1996). This is an important factor for the application of advanced image processing since the effectiveness of image classification or analysis of the image data relies on the radiometric uniformity of the image, especially if an atmospheric correction is required to be applied.
The CASI-2 utilises a high performance position and orientation system (Applanix POS/AV 310) to collect attitude (roll, pitch, yaw) and position (GPS) data simultaneously with image data. The system provides dynamically accurate, high-rate measurements of the full kinematic state of the aircraft. The positional data is differentially corrected using a base station. The processed position and attitude data is then used in a proprietary geo-correction process which applies geometric correction to each flight strip and geo-references the data to a standard coordinate system to an accuracy of within 3-5 pixels. If available, a digital elevation model can be utilised to create the ortho-rectified image strips, which eliminates the positional inaccuracies that result from the terrain height variations. This further improves the level of achievable positional accuracies.
The individual flight strips can be colour balanced to negate the illumination variations between the each strip. This is usually not necessary since each adjacent image strip is flown sequentially ensuring there is minimal time passage between each acquisition. This helps to minimise the irradiance variation due to the sun’s movement. The movement of the sun also affects the length of shadows cast by the vine rows. Depending on row orientation, the shadow change can be significant between adjacent image strips if the time lag is too long. However, the Ball AIMS practice of flight headings in the direction of the sun (i.e. into or away from the sun) in conjunction with sequential image strip acquisition, helps to eliminate or minimise the cross track brightening as well as the illumination variations.
The resultant radiometrically and geometrically corrected image strips are then mosaicked to produce a seamless image tile that is geo-referenced and hence is suitable for advanced processing (Figure 1).
Figure 1: The CASI-2 false colour image mosaic of a vineyard in Coonawarra after radiometric and geometric corrections
The mosaic image is then atmospherically corrected using the empirical line calibration method (ELM) converting sensor radiance units to estimates of ground reflectance. Two spectrally contrasting pseudo invariant features (PIFs) bracketing the range of reflectance values from vegetation targets are used to develop prediction equations applied for calibration. The spectral characteristics of these targets, which are free from the effects of the atmosphere, are then measured with a field spectrometer and using the same target features in the image, are enforced on the radiometrically corrected image data to perform a low cost atmospheric correction. The PIF’s can be seen in Figure 1, coordinates: 487 050 E, 5 862 700 N (approx.).
The correction is applied to the full mosaic with the assumption that there is no significant variation of the atmospheric effects across the image. Following atmospheric correction, an independent validation of the atmospheric correction is undertaken using other spectrally flat features in the image such as gravel and bitumen road surfaces at known locations and measured reflectances.
Each vineyard block is defined in a vector layer, on a post to post basis with access tracks or headlands and other features being excluded. All algorithms are applied at the block level that constitutes a vineyard management area, which exclude other features. For the greenness index to provide optimal results, the desired vineyard conditions are:
• Absence of strong groundcover growth between vine rows;
• Vine canopies are fully developed and spread across the trellising;
• Vine understory free of weeds and grass; and
• Uniform structural and morphological existence of the vine rows (eg. vine row spacing, trellising etc.)
Vine rows are isolated from the rest of the imagery using a combination of band math, band ratio and NDVI thresholding designed to exploit the spectral differences between the vines and non-vines including weeds and grasses. This removes the background data attributable to weeds, grass, soil, rock and litter and provides a mask for the vine rows (Figure 2). The combination of these algorithms then produces an image that helps to confine the analysis to vine plants only.
This is important to ensure any subsequent processing to assess the health and vigour of the vines is not attenuated by the existence of non-representative plants and vegetation (or soils/shadows etc.). The NDVI imagery is then truncated using this mask and the mean vine NDVI value is substituted in the inter-rows to remove the variability attributable to the background “noise” (eg. weeds, grasses). This provides NDVI values more representative of the vine-canopy condition throughout the vineyard.
Figure 2. Vine row extraction from an area of weed/grass infestation. The NDVI vineyard segment is shown on the left; and the extracted vine row mask from the same segment on the right (yellow: vine rows, blue: background).
Vine vigour grading is based on the NDVI ratio, which is used as a surrogate indication of bio-physical status and productivity. In healthy crops the index is positive having a maximum value of 1.0. The visible red wavelengths are sensitive to foliar pigments such as chlorophyll and carotenes, while the NIR is an indicator of plant cell structures and biomass (Davenport & Nicholson 1993). General patterns of vine growth variability evident within the raw NDVI image (Figure 3) are more readily quantified into management units (or regions) in the graded image (Figure 4).
The greenness index is based on a smoothed form of the NDVI image, that has had inter vine row variability removed. Two results from the same data set are produced, firstly a relative greenness index image, and secondly an absolute greenness index image. The relative greenness image grades the smoothed NDVI values such that each grade covers equal area. This enables the mapping of greenness variability throughout the vineyard. In contrast, the absolute greenness image has the function of assessing the greenness conditions of a given vineyard in relation to any other vineyard both in temporal and spatial terms. This is made possible by the application of atmospheric correction to the imagery which then defines all imagery in terms of spectral reflectance. Smoothed NDVI values are segmented through a histogram equalisation process and binned to invariant thresholds to provide “like” comparisons on a temporal basis enabling longitudinal change monitoring. The number of grades to be applied can be nominated by the client in keeping with their application requirements.
A number of smoothing algorithms are applied to the transformed image to generalise the NDVI data on a local area basis, producing spatial regions more easily identified in terms of management units than is possible with the unsmoothed data. Integration of the smoothed data with the spectral data is completed for visual impact and spatial orientation for the interpreter. Graded vine blocks are typically overlaid on a geo-referenced greyscale image (client nominated map or grid datum and projection system) and show discreet vine rows for in-situ referencing on a row-by-row basis for the end user. The overlaying can be performed on any colour imagery the client may request (i.e. true/false colour etc).
Figure 3: The normalised difference vegetation index (NDVI) image. Plants are shown in pink-red and non-plant in blue-black
Figure 4. Five class relative greenness index image overlaid on a monochromatic CASI mosaic. Each plantation within the block is labelled with grape variety and area.
Vine variety mapping is performed by the combination of two supervised classification techniques – spectral matching and maximum likelihood method. The hyperspectral nature of the data set is central to the success of the varietal discrimination since 18 narrow bands create a more complete and discrete spectral signature for each variety than a four band image of low spectral resolution (eg. Landsat TM). This is attributable to the different reflective/absorptive properties that the different varieties exhibit through the VNIR spectrum.
The classification of plant species is not a trivial task due to several species having quantitatively similar spectra and the significant spectral variation within a species. The spectral separability is described by a few variables affecting the visible wavelengths, namely chlorophyll a, chlorophyll b, carotenoids, and xanthophylls, while the number and configuration of cells and cell structures within the leaf are influential in the NIR (Price 1994, Cochrane 2000). Plant species discrimination is found to be most successful at spatial resolutions less than the target size (Cochrane 2000). Hence the high spatial and spectral resolution of the CASI-2 imagery enables the discrimination of the vine varieties using the segregated vine-canopies. This also limits the effects of confounding background influences on the vines such as soils, rocks, grass and ground litter (Elvidge et al.1993). It also allows the use of spectrographic methods to be applied on a pixel-by-pixel basis using either library spectra or image-derived spectra.
The two methods of supervised classification were applied to the 18 band image data resulting in a single layer of classification for each. These were then combined to form a single classified image enhancing the overall classification result (Figure 5).
On a large scale, the varietal mapping enables the regional mapping of vine varieties, which can be used for planning or inventory purposes. At the individual block scale, clusters of misplantings or off-types can be identified which help to avoid un-intended grape mix at harvest. In Figure 5 Shiraz misplantings have been discriminated within a Cabernet Sauvignon plantation. The isolated single pixels classified as Shiraz and Merlot (Figure 6), are likely to be misclassifications due to spectral similarity.
Identification of intra-row vine gaps is achieved through the application of textural analysis combined with filtering methods. The technique is based on the processing of the NDVI vineyard data. The vine gap identification (Figure 7) enables the quantification of the number and the extent of the vine gaps within a given vine block, which can then be converted into total area of missing vines. This can be used to assess the potential for re-planting or other management changes and the benefit of such choices, allowing the management to make informed decisions about vineyard improvement.
Figure 5. Five class relative greenness index image overlaid on a monochromatic CASI mosaic. Each plantation within the block is labelled with grape variety and area.
Figure 6. Enlargement from Figure 5 showing the clusters of misplantings
This technique does require a modified application in the case of non-trellised (old) vine plantations where there are regular gaps between each plant. This is also true for instances where certain complicating conditions exist (eg. full weed infestation of the inter-rows) which may warrant additional processing to achieve the same results (i.e. classification and separation of the weeds from vine plants).
Figure 7. Vine row gap identification (enlargement) overlaid on a monochromatic CASI background (yellow: vine rows, red: vine row gaps).
Temporal comparisons of the greenness index can be made to assess changes in the vineyard condition as a result of implemented management practices. This is made possible by atmospherically correcting the image data and converting it to spectral reflectance values. The absolute greenness index images can be used for periodic change detection, whilst the relative greenness imagery can be used to track the spatial nature of the variability across a block from season to season.
The image products described thus far can be overlayed on a digital elevation model to map features across the topography. Draping processed imagery over a DEM can be beneficial for image analysis and interpretation, especially where vigour variation and irrigation anomalies bare a relationship with topography. The vineyard shown in Figure 8 clearly illustrates increased greenness in the troughs and decreased greenness at the ridge. This variation may be due to a number of factors including the irrigation system causing over-watering in the low areas (due to increased water pressure) or the existence of better soils as result of run-off over time, or a combination of these factors.
Figure 8: 3D perspective of graded vine greenness image draped over a digital elevation model.
The utilisation of processed high-resolution remote sensing imagery for farm management enables both static and dynamic quantitative crop assessment in relative and absolute terms. Precise area measurements allow a variety of input and output statistics related to yield and expenditure to be calculated. Management practices with the objective of maximising within block uniformity can be implemented through appropriate resource allocation adjustments. Where uniformity cannot be achieved due to natural causes, harvesting practices may be altered and undertaken in more targeted manner than previously possible.
Issues to consider during image capture and processing
For any form of change analysis, it is essential to apply atmospheric correction to the acquired image data, which enables the developed image processing algorithms and techniques to be applied to temporally different image data. Otherwise the methods will have to be modified to yield similar processing results using image data that is not radiometrically nor atmospherically corrected. To make valid the annual comparisons of vineyard health through greenness level analysis, several conditions must be met during data acquisition. The phenological vine state should be comparable between years as NDVI values are absolute in the sense that they are derived directly from the reflectance values, which in turn are dependent upon both growth state (biomass) and condition. The image data acquisition should also be conducted at approximately the same period of the year and same time of day for similar sun-sensor-target conditions to exist to yield consistent at-sensor radiances to be observed.
An integral part of Ball AIMS flight planning is the calculation of the sun azimuth and altitude (or zenith distance) for the period of data acquisition. It is important that the sun altitude is sufficiently high during image acquisition to avoid unacceptable levels of shadowing occurring either on the ground between the rows, or worse, on the adjacent vine rows. Excessive shadowing can lead to significant ‘darkening’ of the vine canopies due to the point spread function (defined as the finite spectral response - or brightness, that is integrated into a central image pixel from the neighbouring pixels). Since the shadowing will be dependent on the vine row orientation, the shadowing effects will be uneven across an image with varying vine row orientation. This results in unrepresentative at-sensor canopy radiances that will manifest as spatial variations in greenness levels. This was found to be true during advanced image processing despite the fact that band ratio calculations (i.e. NDVI) largely negate the effects of such brightness variations.
Also, as mentioned earlier, the brightness variation across an image is significantly reduced by flying in the direction of the principle plane of the sun (i.e. into or away) which is determined from the calculated sun azimuths, that further helps to improve the image quality.
Another factor to consider during advanced image processing is the effect of the bi-directional reflectance distribution function (BRDF), which is defined as a function of viewing and illumination geometries on the reflectance properties of vegetation canopies. The BRDF effects are most significant at the swath extents of any remote sensing imagery due to the more extreme sun-sensor-target geometry experienced at these points as well as the greater path radiances (Shibayama & Wiegand, 1985). This is especially true for frame grabbers such as aerial photography and some multispectral video imagery where the sensor viewing geometry ranges 360° across an image and hence is affected by the changes in the sun (azimuth and zenith), sensor and target geometry leading to observable variations of at-sensor radiances across the one frame of imagery.
The BRDF effects are largely minimised for line scanners such as the CASI-2, where an image is constructed scan line by scan line by the forward movement of the sensor platform where the variation to the view angle is only across track, resulting in negligible forward scatter being observed. This combined with the practice of flying in the principle plane of the sun and the relatively narrow field of view (FOV) of 37.9° all help to largely eliminate the effects of BRDF on the CASI-2 hyperspectral imagery. The relatively narrow FOV also helps to maximise the capture of vine row ‘canopy’ only where a wide FOV could result with proportional vine row canopy/side combinations at the swath edges due to the extreme viewing angles.
Integration with GIS for Total Vineyard Management
Total vineyard management requires the collection, fusion, manipulation and analysis of data from a variety of sources for effective decision making in the improvement of vineyards, and hence grape and wine quality and quantity. Processed remote sensing imagery such as the those demonstrated in this paper provide valuable information that can significantly enhance the ability to analyse and quantify the results of management decisions through effective monitoring of the vineyard conditions in conjunction with the ancillary data sets. These sets of data are best integrated in a capable GIS environment, which would facilitate the fusion and analysis tasks being efficiently undertaken.
Existing plantation areas and other features of interest (eg. dams, vacant blocks, future plantation or improvement sites etc.) can be derived from geo-referenced airborne images for inventory and management. This can help the grower to accurately quantify the application of fertiliser and chemicals, water required for irrigation, cover crop plating, average expected yields from each block, materials required for new plantings (cost estimation), and vine plants for each block.
It has been demonstrated that using high resolution CASI-2 hyperspectral imagery it is possible to develop image products that can be used as valuable tools by the vineyard manager and owner alike for the improvement and enhancement of their operations. These tools enable such improvements to be implemented in a cost effective and efficient manner, thereby potentially increasing the quality and quantity of produce and hence return, as well as increasing the value of the operations and investments by value adding to their existing infrastructure. Careful analysis can establish relationships between all the factors identified from the image data as well as ancillary information hence enabling a better understanding of the critical indicators of plant health, crop quality and quantity. This in turn helps to define the parameters influential in the control and management of high value vineyards leading to more uniform and effective targeting of resources both in effort and materials.
For remote sensing to be effectively applied to vineyard precision crop management, a ground resolution approaching that of the vine canopy width (or better) is essential for vine row delineation especially in vineyards where substantial background ‘noise’ (i.e. weeds, grasses etc.) exist. Most processing methods developed so far require only a few high spectral resolution bands. However, variety discrimination and health status determination require carefully selected hyperspectral data of high spectral resolution. Furthermore, the CASI-2 sensor providing programmable, high resolution spectral and spatial image data, can help facilitate development of prognostic analysis for vineyard health status determination using the principles of the red edge shift. This process is currently a subject of Ball AIMS research and development.
Ball AIMS acknowledges the assistance and collaboration of Rosemount Estates Pty Ltd and Southcorp Pty Ltd in providing valuable input and feedback in the development data products and the CSIRO (Land & Water, Canberra) for their assistance during data acquisition for this project.
Cochrane, M.A. (2000). Using vegetation reflectance variability for species level classification of hyperspectral data. International Journal of Remote Sensing, 21, 2075-2087.
Davenport, M.L. and Nicholson, S.E. (1993). On the relationship between rainfall and the normalised difference vegetation index for diverse vegetation types in East Africa. International Journal of Remote Sensing, 14, 2369-2389.
Elvidge, C.D., Chen, Z. and Groeneveld, D.P. (1993). Detection of trace quantities of green vegetation in1990 AVIRIS data. Remote Sensing of Environment, 49, 304-310.
ITRES Research Ltd. (1996). Compact Airborne Spectrographic Imager software manual.
Moran, M.S., Inoue, Y., Barnes, E.M. (1997). Opportunities and limitations for image-based remote sensing in precision crop management. Remote Sensing of Environment, 61, 319-346.
Pearson, R., Grace, J., and May, G. (1994). Real-time airborne agricultural monitoring. Remote Sensing of the Environment, 44, 271-290.
Shibayama, M, Wiegand, C.L. (1985). View and Zenith, and Solar Angle Effects on Wheat Canopy Reflectance. Remote Sensing of Environment, 18, 91-103.
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.