Towards ML-based detection of terrestrial vegetation responses to seasonality and extreme weather events

The response of terrestrial vegetation to seasonal or extreme weather events is complex and dynamic. In recent decades, the increased frequency and intensity of extreme events, driven by global warming, have led to adaptive processes in the biosphere. These processes can also place additional stress on ecosystems, limiting their functionality.

Possible consequences of extreme weather events include shifts in the timing of seasonal vegetation activity, as well as changes in the strength of ecosystem functions such as carbon dioxide assimilation. These temporal characteristics include the start, peak, and end of the vegetation growing phase. Such shifts challenge the accuracy of traditional monitoring and modelling of ecosystem dynamics based on climatic thresholds or phenology, which have become less accurate over the past few decades. The existing methods also overlook the irregular vegetation responses under stress conditions caused by short-term impacts. New indices, parameters and methods are needed to better capture evolving vegetation responses, especially in the context of overall ecosystem functioning.

We propose that anomalies in seasonal photosynthetic activity, measured through near-real-time fluctuations in aboveground atmospheric CO₂ concentrations, could be used to qualitatively assess the impacts of extreme events on terrestrial ecosystems. When interpreted in conjunction with meteorological and remote sensing data, CO₂-based metrics could enhance our understanding of ecosystem functioning. We show preliminary results obtained with this approach, in combination with methods of correlation analysis of CO₂ trends and net ecosystem exchange index, We find good sensitivity and an adaptive response, which could be promising to advance ecological monitoring.

We expect that limitations of our approach, such as generalisation and behaviour-averaging, could be overcome with machine learning approaches. These could focus on the detection of vegetation functional periods, as well as in the qualitative assessment of functioning.

Our research results, based on a high-level carbon balance model, statistical methods, and time-series analysis, provide a preliminary non-phenological detection of vegetation activity periods and CO₂ uptake strength. We expect that our method can be applied in conjunction with existing approaches to aid identification of vegetation activity and ecosystem functioning, or as a standalone tool for their preliminary evaluation in near-real-time.