Estimating grapevine water status using hyperspectral and multispectral remote sensing and machine learning algorithms : a thesis presented in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Agriculture and Horticulture at Massey University, Manawatū, New Zealand

Thumbnail Image
Open Access Location
Journal Title
Journal ISSN
Volume Title
Massey University
The Author
Moderate water deficit is desirable for achieving the optimal grape composition which determines the values of wine, especially for red cultivars. To attain consistency in grape quality in vineyards, it is critical to manage grapevine water status (GWS) to the target range, but to avoid severe dehydration, between fruit set and veraison. Together with the foreseeable climate changes and stricter environmental regulations, there is a need for viticulture to estimate GWS variability across fields along growing seasons before irrigating to eliminate the uncertainty of controlling hydration status and to produce consistent grapes with premium quality. Precision viticulture (PV) recognizes that not all areas within a vineyard are uniform in terms of their soil, climate, and other environmental conditions. Therefore, it tailors viticultural management to the unique needs of different vineyard zones by focusing on applying site-specific or time-specific management practices. PV aims to enhance grape quality and yield while minimizing resource usage and environmental impacts. As the final product (wine) for viticulture has the potential of high additional values, it is worth considering the application of PV in decision-making according to the information on spatio-temporal variability across the fields. The advancement and availability of remotely sensed spectral information, geospatial technologies, and machine learning models have opened a new chapter for spatio-temporal GWS monitoring. However, there are technical shortcomings that need to be addressed before extensive application and adoption of these techniques in viticulture. These include a lack of understanding of GWS-related spectral data analysis methods, a lack of data interoperability for GWS estimation between data sourced from various devices with different formats, and a lack of availability of high-coverage images with high spatial and temporal resolution. Therefore, this study tackled the technical bottlenecks, related to the application of proximal sensing and remote sensing (RS) in GWS estimation, from three perspectives: (i) the exploration of relevant spectral regions over the electromagnetic spectrum, (ii) the complementation from differently sourced datasets other than RS information, (iii) the provision of large-scale GWS prediction. This study was undertaken in two Pinot Noir vineyards trained with vertical shoot positioning for two growing seasons in Martinborough, New Zealand. The investigation window, corresponding to the critical periods of GWS management, between fruit set and veraison in each growing season, includes November, December, January, and February. Stem water potential (Ψstem), serving as a proxy for GWS, is measured on 85 and 63 canopies in the first and second growing seasons. Each sampled grapevine is recorded for its location with a global navigation satellite system with real-time kinematic correction. Five times field data collection, including measuring hyperspectral point data using ASD FieldSpec 4 Spectroradiometer (proximal sensing data) and taking multispectral images using DJI Phantom 4 UAV (remote sensing data), were carried out in each growing season. An electromagnetic induction survey was implemented by using EM38-MK2 to acquire apparent electrical conductivity (ECa) maps (complementary data). Several satellite images collected by PlanetScope (remote sensing data) during the study periods and the LiDAR-based digital elevation model (complementary data) were downloaded and added to the analysis datasets. An on-site weather station continuously records and provides meteorological information, including air temperature (°C), relative humidity (%), rainfall (mm), wind speed (km/h), and irradiance (W/m2) (complementary data). The identification of the relationships between spectral information and Ψstem is an essential step for the robust application. By analyzing hyperspectral spectra, it shows that the statistically relevant wavelengths disperse across visible, near-infrared, and shortwave infrared (SWIR) spectral bands. They are specifically located around blue, red, and red edge bands, two weak water absorption bands at 970 and 1200 nm, two strong absorption bands at 1400 and 1940 nm, and some dry matter-related bands. When analyzing multispectral images taken by UAV, it shows Transformed Chlorophyll Absorption Reflectance Index and Excess Green Index are the multispectral indices mostly correlated (R2 = 0.35 and 0.3, respectively) with the changes in Ψstem. It implies that the variation in leaf pigments, especially chlorophylls, is better for describing the Ψstem variation of Pinot Noir than the alteration in canopy structure. When applying the Ψstem-sensitive spectral bands through airborne or spaceborne platforms, the missing SWIR for most commercial multispectral sensors and the presence of vapor in the air obstruct the usefulness of the Ψstem-sensitive spectral bands. Therefore, this study assesses the complementary effects provided by other environmental aspects, including soil/ terrain, vegetation, temporal, and weather variables, to improve the GWS estimating capabilities of aerial multispectral sensors. The results prove the complementary effects by displaying that the detection accuracy is improved from RMSE of 213 to 146 kPa and RMSE of 221 kPa to 138 kPa. To monitor the fields at a large-scale using multispectral satellites, it is common to encounter several technical issues: coarse resolution pixels that contain background information, weather dependence, and delay in the image delivery. This study addresses the limitation of coarse spatial resolution using two-stage calibration to scale information provided by ground measurement up to satellite images, along with removing interference from the inter-row components. It demonstrates that satellite images can approximate the collected Ψstem with high accuracy (RMSE = 59 kPa). To deal with the contamination by weather and delivering delay of image products, a prediction model is established based on the calibrated satellite images and various environmental variables (day of the year, rainfall, potential evapotranspiration, irrigation, fertigation, plucking, trimming, normalized difference vegetation index, ECa, elevation, and slope). The developed model is able to predict Ψstem trend in an independent growing season with high consistency when compared with the reference (r = 0.89 and 0.87 for the two vineyards, respectively).
Grapes, Irrigation, New Zealand, Machine learning, Hyperspectral imaging, Industrial applications