Winetech Technical Yearbook 2022

FIGURE 1. Boruta-selected bands (grey bars). A mean spectral signature is also shown for reference.

TABLE 2. Existing spectral indices related to water stress that were used in this study. Reflectance index Acronym Equation Disease Water Stress Index 4 DWSI 4 R550 / R680 2 Normalised Difference Vegetation Index NDVI (R800 - R670) / (R800 + R670) 3 Photochemical Reflectance Index 1 PRI 1 (R570 - R531) / (R570 + R531) 4 Photochemical Reflectance Index 2 PRI 2 (R531 - R570) / (R531 + R570) 5 Red Edge RE R750 / R710 6 Red Edge Vegetation Stress Index RVSI 0.5 (R722 + R763) - R733 7 Water Index WI R900 / R970 7

CONCLUSION This project presented a novel hyper spectral-machine learning framework for the non-destructive identification of wa ter-stressed vines. The results indicated the viability of narrow wavebands to model vineyard water stress and established the utility of tree-based machine learning al gorithms within the domain of viticulture. Furthermore, the study demonstrated the feasibility of feature selection methods for the development of optimised spec

TABLE 3. The overall accuracy for all the developed optimised indices using the Boruta-selected bands. R525 / R__ 825 745 822 751 627 582 407

579

576

563

KHAT

0.77

0.7

0.57

0.5

0.27

0.23

0.2

0.2

0.17

0.1

OA (%)

88

85

78

75

63

62

60

60

58

52

TABLE 4. The overall accuracy (%) for all the existing indices and the optimised index for all four datasets.

Existing indices

Optimised index

Dataset

Red Edge NDVI

RVSI

PRI 1

PRI 2

WI

DWSI

R525 / R825

T1

77

65

65

60

56

47

53

88

T2

79

76

66

45

45

50

48

82

T3

68

77

56

50

50

53

47

83

T5

67

88

57

53

53

51

33

87

Rj Ri

1), indicating a 74% reduction in data dimensionality. After comparing different iterations of two-band ratios (table 3), an optimised index was derived comprising the 525 nm and 825 nmwavebands (R525/ R825). The optimised index illustrated a clear improvement in classification accu racy when compared to the existing indices (table 4). A training accuracy of 88% (kappa = 0.77) was obtained with the optimised index, with test accuracies ranging between 82% - 87% (kappa ranged from 0.65 - 0.74) across the three remaining datasets. Over all, the study showed the potential of a two-band vegetation index to model water stress in a laboratory environment. These results illustrate the operational potential of the developed two-band index for water stress detection in Shiraz vineyards.

tral indices. The study provides a point of departure for the operationalisation of future machine learning-remote sensing frameworks for water stress monitoring. ABSTRACT The research aimed to derive a two-band optimised vegetation index to model wa ter stress in Shiraz vineyards. The opti mised index was developed using visible near-infrared (VNIR) hyperspectral data and a Boruta-derived subset of important wavebands. The optimised index (R525/ R825) was compared to existing water stress indices noted in the literature. The results demonstrated a clear improvement in classification accuracy, with the opti mised index producing overall accuracies ranging between 82% - 88%.

Optimised index (j.i) =

Eq 1

where Rj and Ri are the reflectance values from the relevant Boruta-select ed wavebands. Rj remained consistent throughout all the indices and was equal to the waveband with the highest Boruta importance score. The optimised index was compared to existing water stress indices (table 2) to determine its performance. The indices were compared based on confusion matrix and KHAT statistic values. RESULTS The Boruta wrapper identified 47 wave bands as important for the detection of water-stressed Shiraz vines (see figure

REFERENCES https://www.wineland.co.za/hyperspectral-the-answer-to-water-stress-part-2/

For more information, contact Kyle Loggenberg at kyleloggenberg@sun.ac.za.

28

WINETECH TECHNICAL YEARBOOK 2022