Estimating defoliation of Scots pine stands using machine learning methods and vegetation indices of Sentinel-2

Abstract

In the presented study, the Sentinel-2 vegetation indices (VIs) were evaluated in context of estimating defoliation of Scots pine stands in western Poland. Regression and classification models were built based on reference data from 50 field plots and Sentinel-2 satellite images from three acquisition dates. Three machine-learning (ML) methods were tested: k-nearest neighbors (kNN), random forest (RF), and support vector machines (SVM). Regression models predicted stands defoliation with moderate accuracy. R2 values for regression models amounted to 0.53, 0.57, 0.57 for kNN, RF and SVM, accordingly. Analogically, the following values of normalized root mean squared error were obtained: 12.2%, 11.9% and 11.6%. Overall accuracies for two-class classification models were 78%, 75%, 78% for kNN, RF and SVM methods. The Green Normalized Difference Vegetation Index and MERIS Terrestrial Chlorophyll Index VIs were found to be most robust defoliation predictors regardless of the ML method. We conclude that Sentinel-2 satellite images provide useful information about forest defoliation and may contribute to forest monitoring systems.