[Paper Review] Semi-Supervised Detection of Extreme Weather Events in Large Climate Datasets

Evan Racah, Christopher Beckham+3 more|arXiv (Cornell University)|Apr 24, 2017

Meteorological Phenomena and SimulationsEarth and Planetary Sciences17 references19 citations

TL;DR

This paper proposes a multichannel spatiotemporal encoder-decoder CNN that leverages unlabeled climate data and temporal dynamics to improve semi-supervised detection of extreme weather events via a 3D convolutional autoencoder with a modified one-pass bounding box regression loss. The method enhances localization accuracy across diverse event types like hurricanes and blocking highs without requiring extensive labeled data.

ABSTRACT

The detection and identification of extreme weather events in large scale climate simulations is an important problem for risk management, informing governmental policy decisions and advancing our basic understanding of the climate system. Recent work has shown that fully supervised convolutional neural networks (CNNs) can yield acceptable accuracy for classifying well-known types of extreme weather events when large amounts of labeled data are available. However, there are many different types of spatially localized climate patterns of interest (including hurricanes, extra-tropical cyclones, weather fronts, blocking events, etc.) found in simulation data for which labeled data is not available at large scale for all simulations of interest. We present a multichannel spatiotemporal encoder-decoder CNN architecture for semi-supervised bounding box prediction and exploratory data analysis. This architecture is designed to fully model multi-channel simulation data, temporal dynamics and unlabelled data within a reconstruction and prediction framework so as to improve the detection of a wide range of extreme weather events. Our architecture can be viewed as a 3D convolutional autoencoder with an additional modified one-pass bounding box regression loss. We demonstrate that our approach is able to leverage temporal information and unlabelled data to improve localization of extreme weather events. Further, we explore the representations learned by our model in order to better understand this important data, and facilitate further work in understanding and mitigating the effects of climate change.

Motivation & Objective

• To address the scarcity of labeled data for diverse extreme weather events in large-scale climate simulations.
• To improve detection and localization of spatially localized climate patterns such as hurricanes, fronts, and blocking events.
• To develop a method that effectively utilizes unlabelled simulation data and temporal dynamics for improved generalization.
• To create a framework that supports exploratory data analysis by learning meaningful representations of extreme weather patterns.
• To enable scalable detection of extreme weather events across multiple climate models and simulations.

Proposed method

• The model employs a 3D convolutional autoencoder architecture to reconstruct multi-channel climate simulation data, preserving spatial and temporal structure.
• It integrates a modified one-pass bounding box regression loss to predict event locations directly during training.
• The architecture processes multi-channel input (e.g., surface pressure, wind, temperature) to model complex spatiotemporal patterns.
• Unlabeled data is used to pre-train the encoder and refine representations through reconstruction, improving downstream detection.
• Temporal dynamics are modeled through 3D convolutions that capture evolution across time steps in climate simulations.
• The framework enables end-to-end training that jointly optimizes reconstruction and localization objectives.

Experimental results

Research questions

• RQ1 Can a semi-supervised deep learning framework improve detection of extreme weather events when labeled data is limited?
• RQ2 How effectively can temporal dynamics and multi-channel climate data be leveraged to enhance event localization?
• RQ3 To what extent does incorporating unlabelled data improve detection performance across diverse event types?
• RQ4 What kind of representations does the model learn, and how do they support interpretability and exploratory analysis?
• RQ5 Can the method generalize across different climate models and simulation datasets?

Key findings

• The model improves localization of extreme weather events by leveraging unlabelled data and temporal dependencies, even with limited annotations.
• The learned representations capture salient features of diverse events such as hurricanes, fronts, and blocking highs, enabling effective detection.
• The integration of reconstruction and bounding box regression losses leads to more robust and accurate predictions than supervised baselines with sparse labels.
• The method demonstrates strong performance on unseen climate simulations, indicating generalization capability across models.
• Exploratory analysis of learned features reveals meaningful spatial and temporal patterns aligned with known meteorological phenomena.
• The approach reduces reliance on large-scale manual labeling, making detection scalable across large climate datasets.