Irrigation management in vineyards: Modeling water stress using hyperspectral imaging

Sushma Thapa

doi:10.7273/000004261

Grapes, valued at nearly $6 billion annually, are one of the highest value fruit crops in the United States. They are commonly grown in Mediterranean climatic conditions with dry summers that are dependent upon irrigation for sustained fruit yield and quality. Although in-situ methods such as measurement of leaf water potential ([psi]L) are available for estimation of water stress, these methods are time, labor and cost-intensive. As climate change intensifies reducing the availability of water resources, development of a more robust, reliable water stress estimation technique in the vineyards is essential for irrigation management. In the past, research initiatives have been mostly focused on deriving linear relationships between physiological parameters and reflectance information from thermal and multispectral imagery. The current research focuses on the development of machine learning models for classification of vine water stress into three classes: no to mild water stress ([psi]L > −0.8 MPa), moderate water stress (–0.8 MPa < [psi]L < –1.2 MPa) and severe water stress ([psi]L < –1.2 MPa). First, optimal vegetation indices (VIs) derived from hyperspectral reflectance information and weather variables were identified for modeling plant water stress. Based on the comparison of linear relationships assessed between VIs and [psi]L and relative variable importance computed from random forest classifier (RFC) test models (RFC-Model 1, RFC-Model 2 and RFC-Model 3), Green Normalized Difference Vegetation Index (GNDVI), Photochemical Reflectance Index (PRI) and Anthocyanin Index (ANT) were found to have consistently good relationship with the data. Similarly, air relative humidity, evapotranspiration, minimum soil temperature and solar radiation demonstrated higher significance in model development. These spectral and weather variables were included in the final optimized RFC and artificial neural network (ANN) models. The study revealed that implementation of machine learning models with the inclusion of environmental variables substantially improved the capability of hyperspectral reflectance to detect water stress conditions in vineyards. The RFC and ANN models achieved accuracies of 73% and 70%, respectively, in classifying plant water condition into three classes. These results were promising for water stress prediction and monitoring, and for the development of decision support tools for precision irrigation in the future.

Irrigation management in vineyards: Modeling water stress using hyperspectral imaging

Files and links (1)

Abstract

Metrics

Details