Hyperspectral-based predictive modelling of grapevine water status in the Portuguese Douro wine region

• Type: Article in International Scientific Journal
• Date: 2017

Authors

I. Poças (Author) FCUP; Goncalves, J (Author) Other; Costa, PM (Author) Other; Goncalves, I (Author) Other; Pereira, LS (Author) Other; Mario Cunha (Author) FCUP

Journal

Title: International Journal of Applied Earth Observation and Geoinformation

Vol. 58

Pages: 177-190

ISSN: 0303-2434

Publisher: International Institute for Aerial Survey and Earth Sciences

Indexing

ISI Web of Knowledge - 46 Citations

Scopus - 46 Citations

Other information

Authenticus ID: P-00M-NZ1

DOI: 10.1016/j.jag.2017.02.013

Abstract (EN): In this study, hyperspectral reflectance (HySR) data derived from a handheld spectroradiometer were used to assess the water status of three grapevine cultivars in two sub-regions of Douro wine region during two consecutive years. A large set of potential predictors derived from the HySR data were considered for modelling/predicting the predawn leaf water potential (Psi(pd)) through different statistical and machine learning techniques. Three HySR vegetation indices were selected as final predictors for the computation of the models and the in-season time trend was removed from data by using a time predictor. The vegetation indices selected were the Normalized Reflectance Index for the wavelengths 554 nm and 561 nm (NRI554;561), the water index (WI) for the wavelengths 900 nm and 970 nm, and the D1 index which is associated with the rate of reflectance increase in the wavelengths of 706 nm and 730 nm. These vegetation indices covered the green, red edge and the near infrared domains of the electromagnetic spectrum. A large set of state-of-the-art analysis and statistical and machine-learning modelling techniques were tested. Predictive Modelling techniques based on generalized boosted model (GBM), bagged multivariate adaptive regression splines (B-MARS), generalized additive model (GAM), and Bayesian regularized neural networks (BRNN) showed the best performance for predicting Psi(pd), with an average determination coefficient (R-2) ranging between 0.78 and 0.80 and RMSE varying between 0.11 and 0.12 MPa. When cultivar Touriga Nacional was used for training the models and the cultivars Touriga Franca and Tinta Barroca for testing (independent validation), the models performance was good, particularly for GBM (R-2 = 0.85; RMSE = 0.09 MPa). Additionally, the comparison of Psi(pd) observed and predicted showed an equitable dispersion of data from the various cultivars. The results achieved show a good potential of these predictive models based on vegetation indices to support irrigation scheduling in vineyard.

Language: English

Type (Professor's evaluation): Scientific

No. of pages: 14