Machine learning approach to predict leaf colour change in Fagus sylvatica L. (Spain)

Abstract The European beech (Fagus sylvatica L.) is one of the most important deciduous tree species in Europe. In this study, we analyse the autumn senescence dynamics of Spanish beech forests using time series satellite data between 2001 and 2017. In addition, we used nine Machine Learning (ML) algorithms to predict the day of the year (DOY) in which their leaf colour change is at 75% (75CF), 10 days ahead, using precipitation and temperature data. The used algorithms were generalized linear model (glm), ridge regression (ridge), least absolute shrinkage and selection operator (lasso), bayesian generalized linear models (bayesglm), partial least squares (pls),weighted k-nearest neighbours (kknn), extreme gradient boosting (xgbTree), support vector machine radial (svmRadial) and random forest (rf). The time series analysis of the actual 75CF did not show any acceleration or deceleration over time. Nevertheless, we noticed a decreasing negative trend between the mean elevation above the mean sea level of the forests and their actual 75CF in terms of Pearson correlation coefficient (r). The best performing ML model was kknn with RMSE = 3 days and R2 = 0.94. To further explore the predictive capacity of the models in a realistic scenario, we held out the last year of the time series (2017) and trained the models with data from 2001 to 2016. The test results proved that rf was the best method in this hypothetical scenario with RMSE = 8 days and R2 = 0.67. This study provides a cost-effective method to predict leaf colour change in beech forests, reducing the shortcomings of previous approaches with a similar goal. It can be used with management purposes for local or regional authorities, as well as being of interest to further investigate climate change impacts on tree species.