Modelling Mediterranean oak tree dieback using the Plant Phenology (PPI) vegetation index

Gergely Söptei

Master's thesis

Abstract:

Vegetation indices have long been used in land use and land cover change identification. The Plant Phenology Index (PPI), one of the many vegetation indices, was introduced in 2014 to overcome the limitations of other indices when used for phenology, for example, with evergreen vegetation or with vegetation affected by snow cover. It is a physically based vegetation index; it is derived from the radiative transfer equation and has a linear connection to the Leaf Area Index (LAI). Tree mortality changes land cover, and has been affecting the montado, a silvo-pastoral agroforestry system, of mainland Portugal for decades. Dead tree data was collected in Companhia das Lezírias, a state-owned property with high proportion of montado (dominated by Quercus suber - cork oak), from 2014 to 2019. This dataset, in conjunction with the High-Resolution Vegetation Phenology and Productivity (HR-VPP) remote sensing product with thirteen yearly phenology parameters, was used in this thesis to determine a relationship between phenological patterns (PPI) and tree mortality. The modelling of the relationship between tree mortality and tree phenology was done with one-class Support Vector Machine (SVM) models, because of the nature of the field dataset. SVMs are supervised learning models that are frequently used for classification tasks. However, the choice for using a one-class classifier was made, which is an unsupervised classification algorithm for outlier detection, because the dataset contains only dead tree observations. The initial dead tree dataset was divided on a yearly basis, including only the ones that overlapped with the HR-VPP product, 2017 and onwards. To validate our models, a dataset was created through visual observations using Google Earth Pro imagery in Quantum GIS. Before creating our final models, experimentation was done using different tools for hyperparameter search and for feature selection. After this experimentation process, to reduce complexity, pairs of phenology parameters were created, and models were trained and tested with them using different sets of hyperparameters. We found that certain combination of hyperparameters and phenology parameters as features, were able to perform classification of dead trees and outliers with convincing accuracy for one-class classification.

Keywords

Support Vector Machines, Sentinel 2, HR-VPP, Oak Decline, Montado, Remote Sensing, One-Class Classification, Plant Phenology

Language used: English

Date on which the thesis was submitted / produced: 28. 4. 2024

Thesis defence

• Date of defence: 5. 6. 2024
• Supervisor: doc. Ing. Radim Matula, Ph.D.
• Reader: Peter Surový, doc. Ing.

Czech University of Life Sciences Prague

Faculty of Forestry and Wood Sciences

Master programme:
Tropical Forestry and Agroforestry