Abstract

The spatio-temporal relations of extreme events impacts and their drivers in climate data are not fully understood and there is a need of machine learning approaches to identify such spatio-temporal relations from data. The task, however, is very challenging since there are time delays between extremes and their drivers, and the spatial response of such drivers is inhomogeneous. In this work, we propose a first approach and benchmarks to tackle this challenge. Our approach is trained end-to-end to predict spatio-temporally extremes and spatio-temporally drivers in the physical input variables jointly. We assume that there exist precursor drivers, primarily as anomalies in assimilated land surface and atmospheric data, for every observable impact of extremes. By enforcing the network to predict extremes from spatio-temporal binary masks of identified drivers, the network successfully identifies drivers that are correlated with extremes. We evaluate our approach on three newly created synthetic benchmarks where two of them are based on remote sensing or reanalysis climate data and on two real-world reanalysis datasets.

Overeview

Overview of the task

Overview of the objective of this work. We are interested in identifying spatio-temporal relations between the measurable impacts of extreme events like the vegetation health index and their drivers. As drivers, we focus on anomalies in state variables of the land-atmosphere and hydrological cycle. The task is very challenging since the drivers can occur at a different region than the extreme event and earlier in time.

Model architecture

model design

An overview of the proposed model to identify the spatio-temporal relations between extreme agricultural droughts and drivers. The input variables are first encoded into features. In a subsequent step, a lockup free quantization layer (LFQ) takes the extracted features and classifies the variables into a binary representation of normal or anomalous events. Finally, a classifier is used to predict extreme events from the identified anomalies.

Comparison to the baselines

Comparison to the baselines 1
Comparison to the baselines 2

Results on real-world reanalysis data

Results on synthetic data

Qualitative results on the synthetic CERRA reanalysis

Qualitative results on the synthetic CERRA reanalysis from the test set at time step 2160. Shown are the prediction, the ground truth, and the false positive. Albedo and relative humidity are not correlated with extremes, meaning that they do not have target anomalies but only random ones.

Results on real-world reanalysis data

The averaged spatial distribution of anomalies

The averaged spatial distribution of drivers/anomalies related to Portugal in Europe. For this experiment, we use prediction on EUR-11 from ERA5-Land and select frames (weeks) within the period 2018-2024 where there were extreme drought of at least 25% of the pixels in Portugal. Then we normalize the identified anomalies by the total number of frames to obtain the final map.

The averaged spatial distribution of anomalies

The averaged spatial distribution of anomalies related to a specific place in Europe (North Rhine-Westphalia). For this experiment, we use prediction on EUR-11 from ERA5-Land and select frames (weeks) within the period 2018-2024 where there were extreme drought of at least 25% of the pixels in the North Rhine-Westphalia. Then we normalize the identified anomalies by the total number of frames to obtain the final map.

Real-world reanalysis dataset

The definition of the domains used in the study

The definition of the domains

We conducted the experiments on two real-world reanalysis (ERA5-Land and CERRA) including data from five continents. ERA5-Land reanalysis is mapped onto the CORDEX domains. CERRA has its own domain definition.

CORDEX Domains

CORDEX Domains

Synthetic dataset

Perceptual examples of the synthetic CERRA reanalysis data

Perceptual examples of the synthetic CERRA reanalysis data. The target anomalies are visualized under each variable directly. Here, albedo and relative humidity are not correlated with the extremes.

Perceptual examples of the synthetic data

Synthetic artificial data

Synthetic artificial data

The target anomalies are visualized under each variable directly. Here, variables 01 and 05 are not correlated with the extremes.

Synthetic CERRA reanalysis data

Synthetic CERRA reanalysis data

The target anomalies are visualized under each variable directly. Here, albedo and relative humidity are not correlated with the extremes.

Visualization of the generated signals Φ

Poster

Poster

BibTeX


	@inproceedings{eddin2024identifying,
					  title={Identifying Spatio-Temporal Drivers of Extreme Events},
					  author={Mohamad Hakam Shams Eddin and Juergen Gall},
					  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
					  year={2024},
					  url={https://openreview.net/forum?id=DdKdr4kqxh}
					  }
				  

Dataset download

-->