Analysis of spectral response patterns of Kiwifruit orchards using satellite imagery to predict orchard characteristics of commercial value before harvest : : a thesis presented in fulfilment of the requirements for the degree of PhD Prod Tech in the School of Engineering and Advanced Technology at Massey University, Palmerston North, New Zealand

Thumbnail Image
Open Access Location
Journal Title
Journal ISSN
Volume Title
Massey University
The Author
Several characteristics of kiwifruit determine its value to the kiwifruit marketing company, Zespri Ltd, and to the grower. The foremost of these is the dry matter content. Much effort is expended in predicting the final dry matter content of the fruit as early in the season as possible so that the optimal dry matter content can be achieved. Dry matter content is currently measured through a destructive 90-fruit protocol that may be repeated several times in a season on each maturity block. Remote sensing data available from modern satellites can provide four-colour (red, green, blue and near-infrared) data with resolution down to 1-2m, less than the size of one kiwifruit vine. Many indices can be created from these and correlated to the characteristics of plants with indifferent results. This thesis presents the development of an index wherein the four colours are used to create a three-dimensional unit colour vector that is largely independent of light level. This transform was used to allow the direct visualisation of data from a number of satellite images of the Te Puke kiwifruit growing area in New Zealand over five years, for which dry matter content values were available from the 90-fruit protocol. An attenuation model was chosen to correct the top-of-atmosphere light intensities recorded by the satellite cameras to those at ground level. The method of Hall et al., (1991) was found to reduce the variation of fiduciary pixels by the largest amount and was used. The visualisation revealed that there was an axis along which dry matter was ordered by magnitude. A regression line of best fit was applied to this data producing an R2 value of 0.51 with a standard mean-square error of 0.76. This is significantly lower than the average mean-square error of 1.05 for the 90-fruit protocol. Comparison of the predictive power of other indices, based on one image, showed a range of R2 values of 0.008 to 0.49. The method developed in this thesis produced an R2 of 0.70 for the same data.
Research Subject Categories::TECHNOLOGY