Using Artificial Intelligence to Reconstruct Missing Climate Data In Extreme Events Datasets

Evaluating the trends of extreme indices (EI) is crucial to detect and attribute extreme events (EE) and establish adaptation and mitigation strategies to the current and future climate conditions. However, the observational climate data used for the calculation of these indices often contains many missing values and leads to incomplete and inaccurate EI. This problem is even greater as we go back in time due to the scarcity of the older measurements.

To tackle this problem, interpolation techniques such as the kriging method are often used to fill in the gaps. However, it has been shown that such techniques are inadequate to reconstruct specific climatic patterns [1]. Deep-learning based technologies give the possibility to surpass standard statistical methods by learning complex patterns and features in climate data.

In this work, we are using an inpainting technique based on a U-Net neural network made of partial convolutional layers and a loss function designed to produce semantically meaningful predictions [1]. Models are trained using vast amounts of climate model data and can be used to reconstruct large and irregular regions of missing data with few computational resources.

The efficiency of the method is well demonstrated through its application to the HadEX3 dataset [2]. This dataset contains gridded land surface EI, among which the TX90p index that measures the monthly (or annual) frequency of warm days (defined as a percentage of days where daily maximum temperature is above the 90th percentile). As for other EI, there is a lack of TX90p values in many regions of the world, even in recent years. It is particularly true when looking at an intermediate product of HadEX3 where the station-based indices have been combined without interpolation. This is illustrated by the left map of the figure where the gray pixels correspond to missing values. By training our model using data from the CMIP6 archive, we have been able to reconstruct the missing TX90p values for all the time steps of HadEX3 (see right map in the figure) and detect EE that were not included in the original dataset. The reconstructed dataset is being prepared for the community in the framework of the H2020 CLINT project [3] for further detection and attribution studies.

[1] Kadow C. et al., Nat. Geosci., 13, 408-413 (2020)
[2] Dunn R.J.H. et al., J. Geophys. Res. Atmos., 125, 1 (2020)
[3] https://climateintelligence.eu/