Multispectral image calibration

by | Mar 1, 2016 | Viticulture research, Winetech Technical

Multispectral remote sensing has been in use in the South African wine industry for more than 10 years, with applications ranging from segmented or sequential harvesting, to adapting future block layout, amongst other applications. Its utilisation as a research tool has also been explored in the form of finalised Winetech/THRIP funded projects, as well as a currently running project. In this article it is discussed how the difficult challenge of calibrating imaging to the often diverse environment of the grapevine can be approached, using imaging along with single-vine measurements of vigour.One key difficulty in working with non-calibrated images, or images from different service providers/imaging platforms is to be able to compare images over differing seasons. One approach is to use relative simple measurable grapevine parameters, such as dormant cane mass.

Results from three studies will be presented here, one aimed at investigating the correlation of within-vineyard field measurements conducted in different areas at different times with Normalised Difference Vegetation Index (NDVI) image pixel values extracted from measured vines. Pruning mass is evaluated temporally and spatially against NDVI image pixel values.

Results from three studies in the Stellenbosch wine growing region are presented where high resolution multispectral images were collected over different growing seasons from 2005 to 2007, along with field measurements at selected mini-plots consisting of 10 to 12 vines each. After creation of NDVI images, pixel values were extracted for these plots, coinciding with differing vigour levels in the vineyard, and compared to pruning and yield data, as well as other selected parameters for the same plots. A simple mask-scan method of image pixel value extraction was used from open source software (Image J, NIH, Bethesda, Maryland, USA) (Figure 1). Note that the NDVI pixel values correspond to ratios of a maximum pixel value of 255 in the 8-bit greyscale NDVI image.

multispectral_thumbnail
FIGURE 1. Steps to extract plot mean pixel values for field data comparison. FIGURE 2. Correlation between plot average pruning mass (kg) and plot average NDVI over two consecutive years, 2005 R2 = 0.74 and 2006 R2 = 0.58. FIGURE 3. Mean pruning mass per plot (left), as well as plot mean NDVI values (right), for different vigour level plots over two consecutive seasons. FIGURE 4. Correlation between 2005 plot average NDVI pixel values and predicted 2006 pruning mass (kg) (R2 = 0.67). FIGURE 5. Multispectral NDVI image classified to three pruning mass levels determined from pruning mass values predicted from NDVI:pruning mass linear regression equations. FIGURE 6. Relationship between mean pruning mass per grapevine in the mini-plots (kg) and the mean NDVI values for the mini-plots in the 2006 growing season for a Stellenbosch vineyard. FIGURE 7. Multispectral NDVI image classified to three pruning mass levels determined from pruning mass values predicted from NDVI:pruning mass linear regression equations.

The correlation between the mean plot pruning mass and mean NDVI pixel values over two consecutive seasons showed a higher R2 value (strength of correlation) in the first season (Figure 2), which shows that a season (or vineyard) where a smaller range of pruning mass values occur, would lead to more difficulty for the NDVI index to differentiate pruning mass differences. Another challenge is shown in the data, namely that for these two years the vineyard’s main vigour classes’ pruning mass was relatively similar, but the NDVI values were different over years for the same plots (Figure 3). This highlights the need to calibrate NDVI values over years with vineyard reaction.

The offset in NDVI values between the two years were corrected by normalising the NDVI values to the 2005 season (correcting offset) and this was used to predict the 2006 mean pruning mass from the normalised NDVI values from 2005 using the linear equation, as an example showing that with these corrections predictions can even be performed over seasons (Figure 4). Creating predictions of pruning mass from NDVI values can yield very well classified images “tuned to the vineyard” and this approach is recommended over “artificial” classification systems which yields arbitrary classes. A viticulturist can respond much better to statements like: “The low vigour area of vineyard block A has a pruning mass of approximately 0.6 to 0.8 kg per grapevine” than: “The low vigour area has a mean NDVI value of 0.6 to 0.8”. Pruning mass values can be related to indices, such as the Ravaz index (yield to pruning mass ratio), which can then be used to make spatial bud load adaptations in these zones.

In another example from a more variable vineyard, with less suspected cover crop/weed influence on the pixel classification, the correlation between plot pruning mass and NDVI was even better (R2 = 0.99) (Figure 6). This yielded an excellent improvement over the raw NDVI classification of the image, with the result shown in Figure 7.

Conclusion

Updated and automated image calibration software can be developed to improve image classification and to ensure stability of image observations between different seasons and between different vineyards. It was clear in this investigation that a better approach is to calibrate images using field data, but in order for this to happen, producers need to establish plots where pruning mass and/or other parameters can be measured consistently over different growing seasons. Furthermore remote sensing service providers (independent of the platform) need to ensure that their software has the capability to incorporate grapevine measurements into a calibration algorithm for the data to be useful and “tangible” to the producer. Further work was also done to evaluate the possibility to extract single-vine pixel information in a similar approach to the mini-plot extraction, and to automate the method using software algorithms (Smit et al., 2010).

Acknowledgements

Conrad Schutte, Zelmari Coetzee, Dr Victoria Carey, Dr Julian Smit, THRIP and Winetech for funding of the projects.

– For more information, contact Albert Strever at aestr@sun.ac.za.

References

Coetzee, Z.A., Carey, V.A. & Strever, A.E., 2006. Quantification of vigour: A correlation between remote sensing and field measurements (poster). Presented at the 3rd International SASEV Conference, Somerset West, South Africa.

Deloire, A., Lopez, F. & Carbonneau, A., 2002. Réponses de la vigne et terroir. Eléments pour une méthode d’étude. Progrѐs Agricole et Viticole 119, 78 – 86.

Ravaz, L., 1903. Sur la brunissure de la vigne. Les Comptes Rendus de l’Academie des Sciences 136, 1276 – 1278.

Schutte, J.C., Coetzee, Z.A., Strever, A.E. & Carey, V.A., 2006. Investigating multispectral image calibration using field measurements of vine vigour for the purpose of spatial and temporal comparisons within vineyards (poster). Presented at the 3rd International SASEV Conference, Somerset West, South Africa.

Smit, J.L., Sithole, G. & Strever, A.E., 2010. Vine signal extraction: An application of remote sensing in precision viticulture. South African Journal of Enology & Viticulture 31, 65 – 74.

Article Archives

Author

Search for more articles

Generic selectors
Exact matches only
Search in title
Search in content

Stay current with our monthly editions

Shopping Cart
There are no products in the cart!
Continue Shopping
0