Understanding the dynamics of terrestrial ecosystems in a changing environment is critical because of their fundamental role in the global carbon (C) cycle. Climate extremes, ecological disturbances, and anthropogenic activities are currently altering the functioning of terrestrial ecosystems. As a result, there is a need to improve the monitoring of the terrestrial ecosystem’s and the role of extreme events (i.e. natural and human-induced disturbances) in the biogeochemical cycles for better quantifying regional and global C dynamics. In recent years, there has been an intensive global effort to measure and model carbon dioxide (CO2) exchanges between the terrestrial biosphere and the atmosphere. The integration of multiple modeling methods, remote sensing data, climate data, and a global network of eddy-covariance (EC) flux towers has provided unprecedented insights in understanding the mechanisms controlling CO2 fluxes from ecosystem to regional scales. However, current bottom-up approaches do not explicitly account for the effects of site history (e.g. forest age) and the ecological memory effects of both vegetation and climate dynamics on CO2 fluxes.
Although most scientists agree on the importance of forest age and ecological memory effects in controlling the CO2 flux variability, there is still a debate about the quantitative role of forest age and ecological memory effects in estimating CO2 fluxes. In my thesis, I explored ways both to integrate forest age as well as ecological memory effects and to quantify their relevance when estimating the spatiotemporal variability of the CO2 fluxes. This was approached from two directions. First, a statistical method based on a combination of climate, ancillary, and EC data was developed to quantify the role of forest age towards the terrestrial net CO2 fluxes. Second, the application of a deep learning (DL) method was explored for understanding the contribution of vegetation and climate’s ecological memory effects on CO2 fluxes.
In Chapter 2, I carried-out an observational synthesis to determine to what extent environmental conditions and site history (i.e. forest age) influence the spatiotemporal variability of forest annual net ecosystem production (NEP) across a set of forest EC flux sites globally. The proposed empirical model yielded a substantial capacity for reproducing the spatiotemporal (Nash-Sutcliffe model efficiency (NSE) of 0.62) and across-site variability (NSE of 0.71) of annual forest NEP. By investigating the model structure, I found that forest age was the main driver of NEP spatiotemporal variability in both space and time (decrease in NSE of 0.42 and 0.50 for spatiotemporal and across-site variability, respectively). These results confirmed the importance of forest age in quantifying spatiotemporal variation in NEP using data-driven approaches and paved the way towards further developments in upscaling EC data. Based on the findings of Chapter 2, I provided new global estimates of forest C balance by accounting for both forest age and climate spatial variations (chapter 3). Gridded estimates of forest NEP inferred from a new forest age map and environmental gridded global products (i.e. air temperature, gross primary production, and nitrogen deposition) at 0.5 spatial resolution for the period 2000-2013 were produced. This approach estimated the global forest NEP as a sink of around +50.2 PgC yr-1 and the net biome production (NBP) of forests of around +30.3 PgC yr-1. Forest NBP estimates matched results of independent forest inventories globally, while discrepancies were found at biome level (i.e. temperate, boreal, and tropical regions). Overall, this first attempt to include forest age for estimating the forest C balance globally provided new insights on both the location and the magnitude of the global land C sink.
Furthermore, I investigated the relevance of capturing the vegetation and climate temporal properties, the so-called ecological memory effects, for predicting net ecosystem exchange (NEE) at 185 forest and woodland FLUXNET sites (Chapter 4). To answer this question, I used a data-driven DL model that translates the response of net CO2 fluxes to past climate and vegetation fluctuations: Long-Short-Term Memory (LSTM) model. The findings of the experiments were two folds: (1) an LSTM approach with embedded climate and vegetation ecological memory effects outperforms a non-dynamic statistical model (i.e. Random Forest) and (2) the vegetation mean seasonal cycle embeds most of the information content to realistically explain the spatial and seasonal variations in NEE. To further explore the contribution of vegetation and climate’s temporal dynamic properties to CO2 fluxes globally (Chapter 5), I expanded the approach developed in Chapter 4. A bottom-up approach and a series of experiments provided evidence with respect to the geographical distribution and magnitude of vegetation and climate’s ecological memory effects on NEE for the 2001-2018 period. The spatial patterns of ecological memory effects as well as the controls of the ecosystem properties and climatic conditions on the observed ecological memory effects’ spatial patterns were explored. The results depicted widespread and substantial ecological memory effects across the globe, confirming the importance of explicitly capturing the vegetation and climate temporal properties to accurately reproduce CO2 flux spatiotemporal patterns. Finally, I explored to what degree vegetation and climate’s ecological effects control the biosphere-atmosphere CO2 responses to a specific climate extreme event (i.e. 2018 European heatwave).
Chapter 6 summarized the main findings of the thesis and provided additional reflections as well as outlooks for future research. Overall, this thesis strengthened the role of ecosystem history in understanding the biosphere-atmosphere CO2 exchange. Methodologically, my works have demonstrated the potential of new modeling approaches in the Earth system science, such as DL. Yet, the accommodation of new data streams in the presented modeling schemes and the development of new model frameworks are of relevance. Thereby, the potential of integrating new datasets (e.g. biomass time-series, soil moisture) for investigating the control of C stocks and soil moisture stress on CO2 fluxes for providing more reliable global CO2 fluxes products were briefly presented in Chapter 6. In addition, the potential application of a data-driven method (i.e. transfer learning) for overcoming the problem of extrapolation when modeling CO2 fluxes from site to globe was introduced. This chapter also pointed at the design of modeling schemes that are not only data-adaptive but also embed physical ecosystem properties, the so-called hybrid models. Finally, a brief reflection on the feasibility of implementing operational systems for CO2 flux monitoring was provided.