Drought Effects

Linear estimation of drought effects on productivity

Authors: Chaonan Ji

The rationale for the demonstration case on climate extremes is the following: especially excess drought and heatwave events (DHEs) are increasingly co-occurring, severely reducing the productivity of both semi-natural vegetation and crops and, therefore, carbon sequestration. Although the impacts of individual extreme events are partly very well investigated, we still lack a general understanding of the impacts of compound event years on ecosystems. Specifically, it remains unclear, firstly, whether DHEs systematically reduce vegetation’s productivity potential of the entire growing season, if temporal compensations undo these effects, and secondly, how such responses differ among vegetation types. In a first exploratory phase (Ji et al. in prep – figures below), we analysed European climate conditions from 2001 to 2022 and used a rank-based method to identify the hottest and driest compound event per year for the entire continent. We used vegetation greenness data per vegetation classes to assess the integral growing season greenness anomalies across events and climate zones. Our results clearly show the large-scale signatures of DHEs on the annual growing conditions for vegetation, which lead to clear impacts on vegetation growing conditions that cannot be compensated. However, the effect is much more pronounced for grasslands than forests, which seem to have much higher seasonal resilience to DHE events. In the case of subsequent DHE years, the effects on forests are, however, much more pronounced indicating the risks of clustered extremes as we expect them in the near future. Given that this study investigated growing-season integrals, it corroborates earlier findings that individual extreme events have the potential to affect the inter-annual variability of the terrestrial carbon cycle. Future land-management strategies should consider such effects in landscape planning for buffering the impacts of climate extremes, reducing the volatility of the carbon sequestration potential of ecosystems, and regulating regional climate feedback.

Ranking results of CHD years for Europe identifies regions experiencing sweltering summers, defined by the top three temperature years in the ranking, shown in reddish tones and highlighted with green edges. Similar analyses were done for all relevant climate variales.

kNDVI value reductions by temperature but in different vegetation types. Clear reductions are seen in grasslands for very high temperature regimes, showing the susceptibility of these ecosystems to DHS.

Vegetation responses in climate space.

Figures 4-6 aim to investigate the responses of different vegetation types (e.g., grasslands, conifers) in hot and dry years worldwide. Specifically, we would like to address (1) the rapid detection of extreme years, (2) the spatial trend of vegetation responses, and (3) the different responses of different vegetation types. We used ERA5 T, P, and soil moisture data, spectral indices, and land cover maps to investigate these points. Based on these preliminary results, we embarked with developing deep learning frameworks that can effectively predict the responses especially in ecosystems that do not show obvious results i.e., forest ecosystems. One first methodological idea towards deep learning was using Echo State Networks, an advanced variant of recurrent neural networks. The results by Martinuzzi et al. (accepted) has not been as conclusive in the sense of gaining significant improvements over existing RNNs as expected. In Echo State Networks (ESNs) only the last layer is trained through linear regression. The absence of derivatives guarantees no vanishing or exploding gradients, offering an alternative solution to gating. To ensure a comprehensive comparison, we also investigate the performance of other RNN architectures. The comparison of these models has a strong focus on the extreme responses of vegetation indices to climate drivers. The primary focus of this model comparison lies in understanding vegetation’s extreme responses to climate drivers. We conducted a comprehensive analysis of recurrent neural networks in the context of modeling biosphere dynamics in response to climate factors. By using daily data, we assessed the effectiveness of these network architectures in capturing extreme events within vegetation dynamics. To discern variations in performance across different scenarios, we employed various metrics. Echo State Networks (ESNs) slightly outperformed other RNNs, but the improvements are relatively minor, despite multiple theoretical arguments in favour of ESN. This is why we still started to explore other methodological avenues before going to a continent-wide deep analysis of extremes in DeepESDL.