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

https://doi.org/10.1016/j.agrformet.2021.108661Get rights and content

Highlights

  • Nine Machine Learning algorithms were used to estimate leaf colour change 10 days in advance.

  • The kknn approach obtained the best predictive results (RMSE = 3 days and R2 = 0.94).

  • Autumn senescence was relatively constant from 2001 to 2017.

  • The rf method predicted 95CF in 2017 with RMSE = 8 days and R2 = 0.67.

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.

Introduction

The European beech (Fagus sylvatica L.) is one of the most important deciduous forest tree species in Europe, covering up to 12 million ha (Rezaie et al., 2018). International efforts such as GENTREE-h2020 warn about the threats that climate change exerts over many habitats of deciduous tree species in Southern Europe (beech included), which are likely to change by the end of 21st century (Schueler et al., 2014).

According to Lindner et al. (2010), climate regional changes are particularly visible in forests, which can alter tree growth rates and its reproduction (Jump et al., 2006). The generally long-life of trees impedes a rapid adaptation to relatively quick environmental changes (Lindner, 2010). In the Iberian Peninsula, the beech habitat has been mainly restricted to mountainous areas where climatic conditions are more favourable (Blanco et al., 1997), showing sensitivity to climate changes (Dulansurem et al., 2017). For instance, lower growth in mature beech trees was observed at lower elevations (Jump et al., 2006). This fact suggests that warming temperatures and drought periods greatly condition beech distribution. Geßler et al. (2007) proved that beech trees were sensitive to water-logging and flooding against habitat competitors such as oaks, which are more resilient to those circumstances. Rezaie et al. (2018) demonstrated an active response of beech trees to atmospheric CO2 increase, with a positive correlation between CO2 and intrinsic water use efficiency. Thus, the physiological performance, competitive and growth rates of beech trees are closely linked to the environmental conditions of their habitat and therefore, subjected to regional climate changes (Peuke et al., 2002).

Measuring the phenology of colour and fall leaf dynamics may be of great interest in studies of ecology, carbon cycle, and even tourist industry (Zhang and Goldberg, 2011). In addition, long-term records in leaf coloration and senescence can provide useful information concerning global climate changes (Myneni et al., 1997) since autumn phenology is thought to be influenced by factors such as temperature, precipitation or photoperiod (Estrella and Menzel, 2006; Archetti et al., 2013; Keenan and Richardson, 2015). However, the knowledge about the relationship between autumn phenological responses of deciduous trees species and climate change remains somewhat limited (Zhang et al., 2020). The physiological process by which leaves change colour during autumn is well documented by Harvard University (2020). It is merely related to pigment variations before leaves falling from the trees. During the senescence period, the leaf chlorophyll a/b concentration decreases while carotenoids remain relatively constant (resulting in yellowish colours) until the leaves fall (Zhang et al., 2020). In some tree species, leaves may turn red as a result of the active synthesis of anthocyanin pigments (Harvard, 2020).

Some studies aimed to monitor leaf dynamics using remote sensing methods. Richardson et al. (2009) used a sensor with red, green and blue channels to observe how the timing and rate of autumn senescence varied across the canopy of different species, with greater variability in autumn than spring. The temporally-normalized brownness index is another interesting approach developed by Zhang and Goldberg (2011), which detected fall foliage coloration with an overall absolute mean difference of less than 5 days. Other approaches used ground-based data to predict the duration and amount of autumn colours in mixed forests (Archetti et al., 2013; Xie et al., 2018).

Machine learning is a popular set of techniques that build computer programs which automatically improve through experience (Jordan & Mitchell, 2015). As demonstrated by Ameida et al. (2012), these machine learning algorithms can be used to identify or detect some phenological patterns in imagery, relate vegetation status with meteorological data (Czernecki et al., 2018) or analyse photosynthesis data with predictions purposes (Zhang et al., 2020b).

The aim of this study is to develop a model to predict leaf colour change in beech forests in Spain. Thus, we can evaluate the temporal variability of the leaf colour change over 17 years (2001-2017) and assess the impact of regional climate changes over the study area.

Section snippets

Study area

We studied the beech forests contained in the Spanish Forest Dataset -MFE50, 1:50000- (MITECO, 2006). They are mainly located in northern Spain and cover a total extension of 398738 ha (Fig. 1). Based on the elevation data provided by the Shuttle Radar Topographic Mission -SRTM, 90 m- (Jarvis et al., 2008), they are located at a mean elevation of 910 m (standard deviation -std- = 42 m). The highest forest is located in the Pyrenees at 1950 m (0°40`26``E, 42°48´25´´N), and the lowest in the

Peak leaf colour change

For the period 2001 to 2017, the earliest date in which beech forests obtained 75CF was on 281 DOY (∼ 8th October), the median on 317 DOY (∼ 13th November) and the latest on 351 DOY (∼ 17th December). The visual inspection of the time series data presented in Fig. 4 does not evidence any acceleration or deceleration in terms of 75CF over time. The median 75CF values do not follow any discernible trend over time, with scores ranging from 312 to 322 DOY.

We used the SRTM (90 m) elevation map to

Discussion

Leaf colour dynamics can provide very relevant information about beech forests because of their sensitivity to subtle changes in climate. However, studies based on autumn plant phenology are scarce due to the limited or unappropriate amount of ground data (Lang et al., 2019).

In our study, we used a time-series satellite-based method (Zhang and Goldberg, 2011) to infer leaf colour change in beech trees. We applied an EDA step to preprocess the data before fitting the models. Firstly, we explore

Conclusions

In this study, we developed a methodology to estimate leaf colour change in beech forests, 10 days in advance, using climate data.

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 of the forests and their actual 75CF in terms of Pearson correlation coefficient.

The performance of nine ML algorithms were compared to evaluate their predictive capacity. We used an EDA

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

We acknowledge the E-OBS dataset from the EU-FP6 project UERRA (https://www.uerra.eu) and the Copernicus Climate Change Service, and the data providers in the ECA&D project (https://eca.knmi.nl).

References (72)

  • X. Zhang et al.

    Monitoring fall foliage coloration dynamics using time-series satellite data

    Remote Sens. Environ.

    (2011)
  • S. Zhang et al.

    Responses of Autumn phenology to climate change and the correlations of plant Hormone Regulation

    Sci. Rep.

    (2020)
  • T. Adole et al.

    Photoperiod controls vegetation phenology across Africa

    Commun. biol.

    (2019)
  • D.L. Alexander et al.

    Beware of R 2: simple, unambiguous assessment of the prediction accuracy of QSAR and QSPR models

    J. Chem. Inf. Model.

    (2015)
  • J. Almeida et al.

    Remote phenology: Applying machine learning to detect phenological patterns in a cerrado savanna

  • M. Archetti et al.

    Predicting climate change impacts on the amount and duration of autumn colors in a New England forest

    PLoS One

    (2013)
  • P.S.A. Beck et al.

    Displaying remotely sensed vegetation dynamics along natural gradients for ecological studies

    Int. J. Remote Sens.

    (2008)
  • BDN (2001). Mapa forestal de España. MFE50 Escala, 1, 50000....
  • G. Biau et al.

    A random forest guided tour

    Test

    (2016)
  • E. Blanco et al.

    Los bosques ibéricos. Una interpretación geobotánica

    (1997)
  • L. Breiman

    Random forests

    Machine learning

    (2001)
  • C. Cammarota et al.

    Variable selection and importance in presence of high collinearity: an application to the prediction of lean body mass from multi-frequency bioelectrical impedance

    J. Appl. Statist.

    (2020)
  • Chen, T., He, T., Benesty, M., Khotilovich, V., & Tang, Y. (2015). Xgboost: extreme gradient boosting. R package...
  • B. Czernecki et al.

    Machine learning modeling of plant phenology based on coupling satellite and gridded meteorological dataset

    Int. J. Biometeorol.

    (2018)
  • R. Cornes et al.

    An Ensemble Version of the E-OBS Temperature and Precipitation Datasets

    J. Geophys. Res. Atmos.

    (2018)
  • A. Donnelly et al.

    Autumn leaf phenology: discrepancies between in situ observations and satellite data at urban and rural sites

    Int. J. Remote Sens.

    (2018)
  • IPOPP 2.3 User's Guide

    (2014)
  • C. Dulamsuren et al.

    European beech responds to climate change with growth decline at lower, and growth increase at higher elevations in the center of its distribution range (SW Germany)

    Trees

    (2017)
  • E-OBS, 2020....
  • P. Estiarte

    Alteration of the phenology of leaf senescence and fall in winter deciduous species by climate change: effects on nutrient proficiency

    Global change biology

    (2015)
  • N. Estrella et al.

    Responses of leaf colouring in four deciduous tree species to climate and weather in Germany

    Climate Res.

    (2006)
  • S.E. Fick et al.

    WorldClim 2: new 1-km spatial resolution climate surfaces for global land areas

    Int. J. Climatol.

    (2017)
  • Y. Fracheboud et al.

    The control of autumn senescence in European aspen

    Plant Physiol.

    (2009)
  • A. Geßler et al.

    Potential risks for European beech (Fagus sylvatica L.) in a changing climate

    Trees

    (2007)
  • Gelman, A., Su, Y. S., Yajima, M., Hill, J., Pittau, M. G., Kerman, J., & Su, M. Y. S. (2020). Package ‘arm’....
  • Harvard-Forest, 2020. https://harvardforest.fas.harvard.edu/leaves/process (Accessed at...
  • Cited by (2)

    • Machine learning-based risk prediction model for cardiovascular disease using a hybrid dataset

      2022, Data and Knowledge Engineering
      Citation Excerpt :

      It is regarded as a component of artificial intelligence. It has a wide range of applications in the fields of electrical [8], health care [9], agriculture [10], meteorology [11], and so on. The HD risk prediction model was developed by Shah et al. [12] using 14 essential attributes.

    View full text