European vegetation response to climate drivers in the last decade: using 1 km MERIS data for modelling changes in land surface phenology

Phenological events, such as onset on greenness and senescence, occur at a specific time depending upon the local climatic conditions. Given this dependency between phenology and climate, the former has emerged as an important focus for scientific research because phenological events are regarded as an indicator of global warming. On the other hand, phenology also affects climate, playing an important role in many feedbacks of the climate system by influencing albedo, and fluxes of water, energy and CO2. Thus, a better understanding of the drivers of phenology is of paramount importance, especially for the senescence phenophases, to which the controlling factors are not well documented.

Temperature is one of the key parameters to regulate vegetation growing states in high latitude regions such as Europe, changes in air temperature will lead to changes in vegetation growth. Numerous studies have been conducted to evaluate the sensitivity of spring phenology to warming using plant phenological records. Additionally, others studies have used time series of satellite sensor derived vegetation indexes to up-scale phenology (Land Surface Phenology; LSP) and study the influence of climate at global or continental scales. These studies performed linear regression between phenology trends or anomalies and temperature values. However, the relation between phenology and climatological drivers is complex, and it is not necessarily linear. Therefore, there is a need for the application of new generation computational tools to assist in extracting as much information as possible from the rapidly growing volumes of digital data. This is the case of the present research, related to a considerably large phenological and climatological dataset retrieved for the whole Pan-European Continent in the last decade.

Regression Trees (RT), a machine learning technique, appears as an alternative to traditional regression (global single predictive models), allowing for multiple regressions using recursive partitioning. When the database has many variables which interact in complicated, nonlinear ways, assembling a single global model can be very difficult and hopelessly confusing. An alternative approach to nonlinear regression is to sub-divide, or partition, the space into smaller regions, where the interactions are more manageable. The application of machine learning techniques has different advantages: i) ability to learn complex patterns, considering nonlinear relationships between explanatory and dependent variables; ii) generalisation ability, hence applicable to incomplete or noisy databases; iii) integration of different types of data in the analysis due to the absence of assumptions about the data used (e.g. normality); and iv) interpretability of results, since RT allows obtaining patterns for a better explanation of a given phenomenon, showing the most important variables and their threshold values.

This contribution reports the application of RT to model the differences in phenology for the natural vegetation of Europe in the last decade using temperature and precipitation data. Multi-temporal Medium Resolution Imaging Spectrometer (MERIS) Terrestrial Chlorophyll Index (MTCI) data at 1 km spatial resolution were used to derive key phenological metrics (onset on greenness and end of senescence) for a 10-year time series data from 2002 to 2012. Differences in phenology were computed as the difference from the decadal median. Surface air temperature data and precipitation were acquired from the European Climatic Assessment Dataset and interpolated at the satellite data spatial resolution from an original of 0.25°. We used the daily mean temperature and precipitation and computed monthly and trimestral averages, as well as growing degree days and chilling requirements for every year. All these variables were used as input to the Regression Tree model. This approach is, to the knowledge of the authors, attempted here for the first time. The goal is to gain access to novel information regarding relationships and potential interactions between differences in phenology (synergy between different climatological drivers and threshold values in temperature, growing degree days, etc), not directly or easily provided by more traditional statistical methods. Apart from focusing on the present case, this research aims to encourage other researchers dealing with complex and interacting systems or processes to further contribute with new insights to this novel line of research.