Land-Biosphere-Society

Science Demonstration Case “Land-Biosphere-Society”

Authors: Wanton Li, Gregory Duveiller, Fabian Gans, Jeroen Smits, Guido Kraemer, Dorothea Frank, Miguel Mahecha, Ulrich Weber, Mirco Migliavacca, Andrey Ceglar, Markus Reichstein: Diagnosing syndroms of biosphere-atmosphere-socioeconomic change.

While previous work (Kraemer et al. 2020) to create an index of the Earth System focussed on describing the different spheres (Atmosphere - Biosphere - Society) separately this use case focuses on describing the interaction between these individual spheres. To obtain a first understanding on the main signal that can be extracted from the ESDC we applied canonical correlation analysis (CCA) on annually aggregated time series per country. Applying a linear method first  can be seen as creating a baseline to understand the relationships between variables before applying nonlinear deep-learning methods. The main difference of using CCA in comparison to PCA is that instead of maximizing the explanation of variance in a dataset itself, the CCA tries to explain as much variance as possible for an independent dataset. This method can be applied as a 3-way CCA to sub-datasets of the ESDC from the biosphere and atmosphere as well as to a compiled dataset based on World-Bank socioeconomic indicators. In order to remove confounding spatial patterns, we spatially detrended the input data to concentrate the analysis on the temporal evolution of the country-based data.

Finally, the result of our analysis is a time-dependent interaction index for each country and every pair of variables that encodes the possible interaction between these spheres. This can be interesting from two viewpoints. First one can examine certain known events for single countries and test if there is a signal in multiple of the interaction data streams. This can be an indication that an event had an effect on multiple spheres and hypotheses can be generated about the possible interactions and causal effects.

We use canonical correlation analysis (CCA) to construct interactive socio-biosphere-atmosphere indices and monitor their temporal changes across different countries. The left plot shows a 3d scatter plot of the first component of every sphere for all countries and years. Outliers points are marked in red color. For two of the outliers (Niger and Vanuatu) the time evolution of these indices is shown and the outliers can be related to known events (2017 Niger soil drought and 2015 Cyclone in Vanuatu).

Another approach to investigate the data is to summarize the long-term trajectories different countries take on decadal time scales. For example, it is possible to define clusters of countries with similar co-evolution of different indices based on  trend and standard deviation of their index time series.

The upper figure (a) shows that global countries are distinguished into seven common groups based on clustering on the CCA constructed components. The button figure (b) shows the mean trajectories of CCA constructed socio-biosphere-atmosphere indices across the groups.

We conclude that our results demonstrate the possibilities to explore the interactions of different Earth System components by using dimensionality reduction techniques that aim to summarize the interaction between different data domains. Since these interactions can be very complex and nonlinear there will be future possibilities to explore nonlinear Deep-Learning based extensions of our methods to improve the robustness of our results and provide more capabilities of diagnosing data-driven trends and generating hypotheses on interactions in the Earth System. Work is also foreseen in collaboration with the EU funded Open-Earth-Monitor (OEMC) project, in which the concept is being further developed with a specific use case involving the European Central Bank to diagnose interactions between the financial sector and the natural system.