Classification of defoliation of Scots Pine using SVM algorithm and Landsat imagery

Temperatures are expected to rise 1.5ºC for 2100 due to climate change caused by anthropogenic emissions. These risen of temperatures are causing an increase in frequency and intensity of drought events that are expected to get even more frequent and more extreme in the following decades. Most predictions suggests that drought stress will cause a large-scale tree mortality, species distribution range contractions and a general productivity loss throughout this century. Defoliation represents an early stage of dieback, at which some silvicultural practices can be done to reduce competence or to improve water retention. Therefore, identifying and mapping these areas is of crucial importance to forest managers in the context of global change. 

The aim of this study is to develop a classification model to differentiate die-off and healthy Scots Pine plots using field data and Landsat time series and to extrapolate the model results to other areas of the species’ distribution. 

We have 51 plots (17 m ~ Landsat pixel) in four sites with healthy (<30% defoliated) and die-off (>= 30% defoliated) areas for which we collected dasometric information along with defoliation and mortality. We downloaded, decomposed, and modelled Landsat time series (1985-2023) of Tasseled Cap components and fitted linear models since the last severe drought (2017) that affected Pinus sylvestris populations. A set of candidate explanatory variables were built including the slope of the linear model between 2017 and 2024, mean trend values from 2022 and mean amplitude from 2022.  

Field data were randomly divided in two independent groups (train and test plots). This process was repeated 100 times and a support vector machines (SVM) model was calibrated for every possible combination of explanatory variables. The best candidate model was chosen according to the model performance metrics of the 100 calibrated models. 

The best model had an AUC > 96.97%, an OAA>88.23% and a die-off prediction rate over 83.3% for the 75 out of 100 repetitions. Applying model ensemble to all the plots an AUC = 97.5%, a die-off prediction rate of 95% was obtained. With this work we conclude that Landsat info from 2017 can successfully classify plots into healthy and non-healthy ones. The results of the model are good enough to extrapolate to areas of similar conditions to our study plots and will be, hence, applicable to most of the Pinus sylvestris distribution.