Deep Learning for Predicting Impacts of Extreme Events and their Drivers

Abstract

Climate change will apparently alter the patterns and frequencies of extreme weather and climate events. Yet, we still struggle to accurately localize and predict where extreme events and their impacts could occur in the future. If we were able to anticipate these events sooner, we could devise more informed decisions and far better adaptation strategies. Recognizing this, scientific efforts are dedicated to identify and forecast extreme events and their impacts using climate data. This dissertation explores applications of artificial neural networks for early warning and forecasting of extreme events and their impacts, such as wildfire, extreme floods and agricultural droughts. We address four main topics. <br/> First, we introduce a novel deep learning approach for the next day wildfire danger forecasting. Previous approaches neglect the intrinsic difference between static and dynamic input variables. In our approach, we propose a 2D/3D two-branch Convolutional Neural Network (CNN) with Location-aware Adaptive Normalization (LOAN) to address this issue. The multi-branch architecture can handle multi-source data and the LOAN layer can modulate dynamic features using the corresponding static features as conditions. Thus, our approach for wildfire forecasting uses a unified 2D/3D model to handle distinct properties of input features. <br/> Second, we address long-term agricultural drought forecasting and introduce Focal-TSMP, a deep learning approach that predicts satellite-derived vegetation indices from a regional climate simulation. We first use the climate simulation for long-term forecasting. Then, the proposed deep learning model translates these forecasts into Normalized Difference Vegetation Index (NDVI) and Brightness Temperature (BT) images that are not produced directly by the simulation. From these predicted images, we then derive key vegetation indices like Vegetation Health Index (VHI) as a proxy for agricultural drought. The proposed methodology is valuable for several downstream applications. Specifically, it allows us to fill gaps in satellite data by estimating NDVI and BT and provides a way to model how vegetation responds to extreme events across different climate change scenarios. <br/> Third, there is a need of deep learning approaches to identify spatio-temporal relations between impacts of extremes and their drivers in high dimensional climate data. There is also a need of standardized benchmarks to tackle this challenging task. To address these issues, we introduce benchmarks and a novel deep learning approach to identify drivers of extreme agricultural droughts from reanalysis data. Our proposed approach is trained end-to-end to simultaneously predict spatio-temporally extreme impacts and their drivers in the input variables. We use a quantization technique to force the network to predict impacts conditioned on the spatio-temporal binary masks of identified drivers. This allows the network to successfully identify drivers that are correlated with the impacts of extremes. <br/> Finally, we introduce RiverMamba for global river discharge and flood forecasting. RiverMamba is a novel deep learning model that was developed to address the limitations of lumped and catchment concept approaches in hydrology. In our proposed approach, we leverage the spatio-temporal connections of bodies of water. RiverMamba achieves this by utilizing space-filling curves and efficient deep learning blocks to capture large-scale river networks. We build RiverMamba as a vision model that can be pretrained with reanalysis data and finetuned on observations to forecast floods on a 0.05 degree grid and up to 7 days lead time globally. <br/> The proposed approaches outperform state-of-the-art and baseline methods in terms of effectiveness and efficiency, which we demonstrate through comprehensive evaluations on diverse benchmarks and datasets.