Cross-disciplinary big data in environmental sciences: A blueprint for efficient utilisation of new options

Environmental research is facing a drastic increase of available high-quality data, not the least due to the eLTER activities. Here simultaneous time series of numerous observables from the atmosphere, soil, streams, lakes and groundwater, etc., and comprising both abiotic and biotic variables will be made available from hundreds of sites. On the one hand quality control of these large data sets becomes a major challenge. On the other hand, though, it opens up completely new options for science as long as some key problems are solved: · How to differentiate between different effects? · How to deal with the filter effects of environmental systems? · How to identify unexpected relationships that a model would not depict? However, environmental sciences still lack a toolbox of approved integrated exploratory data analysis approaches to tackle these challenges in a systematic way. Here we suggest a combination of different methods that proved very efficient both in terms of data quality control and of exploratory data analysis for large sets of time series. Examples will be presented from the AgroScapeLab Quillow (LTER site DE-07-UM, Germany) and the Hurdal ICOS and ICP Forest Level II site (Norway). The Hurdal site is planned to be established as an elTER site as well. Any change of boundary conditions, of input fluxes, emerging invasive species etc. (termed “signal propagation” for short) in environmental systems is subject to filtering effects. A key feature thereof is low-pass filtering. Here we suggest the new Cumulative Periodogram Convexity (CPC) index to quantify the effect size for comparison of various time series. Principal Component Analysis of time series (termed Empirical Orthogonal Function approach in climatology) is suggested as another decisive step. Loadings on single components can be used for assessing the size of single effects on observed time series. Visualization of the communalities and of similarities between different observables and sites in a combination of Self-Organizing Maps and Sammon Mapping allows a rapid survey of some tens to hundreds of time series at a glance, e.g., for quality control. Additional consideration of the CPC index proved a powerful tool for identification of the respective key drivers and of the pathways of signal propagation through environmental systems, comprising both biotic and abiotic observables. Applying machine learning approaches to principal components rather than to the raw data facilitates developing a better understanding of complex interactions in environmental systems. To conclude, we see great potential in a systematic combination of existing approaches deserving to be explored further.