[논문 리뷰] Application of Deep Convolutional Neural Networks for Detecting Extreme Weather in Climate Datasets

Yunjie Liu, Evan Racah외 7명|arXiv (Cornell University)|2016. 05. 04.

Traffic Prediction and Management Techniques참고 문헌 30인용 수 315

한 줄 요약

paper translates to Korean: The paper presents a CNN-based approach to detect extreme climate events (tropical cyclones, atmospheric rivers, weather fronts) in climate datasets, achieving high accuracy and avoiding subjective thresholds.

ABSTRACT

Detecting extreme events in large datasets is a major challenge in climate science research. Current algorithms for extreme event detection are build upon human expertise in defining events based on subjective thresholds of relevant physical variables. Often, multiple competing methods produce vastly different results on the same dataset. Accurate characterization of extreme events in climate simulations and observational data archives is critical for understanding the trends and potential impacts of such events in a climate change content. This study presents the first application of Deep Learning techniques as alternative methodology for climate extreme events detection. Deep neural networks are able to learn high-level representations of a broad class of patterns from labeled data. In this work, we developed deep Convolutional Neural Network (CNN) classification system and demonstrated the usefulness of Deep Learning technique for tackling climate pattern detection problems. Coupled with Bayesian based hyper-parameter optimization scheme, our deep CNN system achieves 89\%-99\% of accuracy in detecting extreme events (Tropical Cyclones, Atmospheric Rivers and Weather Fronts

연구 동기 및 목표

• Motivate the need for objective, data-driven extreme weather pattern detection in large climate datasets.
• Develop a deep CNN architecture capable of learning high-level climate pattern representations from labeled multivariate data.
• Reduce reliance on subjective threshold-based event definitions in climate analytics.
• Demonstrate Bayesian hyper-parameter optimization to improve model performance.
• Evaluate CNN performance across multiple event types using real climate datasets.

제안 방법

• Formulate extreme weather detection as a visual pattern recognition task using CNNs.
• Construct a shallow 4-layer CNN (2 conv, 2 fully connected) with ReLU activations and max pooling.
• Stack climate variables into image patches centered on events for model input.
• Apply Bayesian optimization (Spearmint) to tune hyper-parameters for training.
• Train and evaluate separate CNNs for Tropical Cyclones, Atmospheric Rivers, and Weather Fronts.
• Use a logistic activation on the final layer to produce class probabilities.

실험 결과

연구 질문

• RQ1 Can deep CNNs learn distinguishing representations of tropical cyclones, atmospheric rivers, and weather fronts from multi-variable climate data?
• RQ2 What classification accuracy can be achieved for each event type using a shallow CNN with Bayesian-optimized hyper-parameters?
• RQ3 Is it feasible to avoid subjective thresholding in extreme event detection by learning from labeled examples?
• RQ4 How do training time and performance vary across event types given the data constraints?

주요 결과

• Tropical cyclone classification achieved 99% train accuracy and 99% test accuracy with ≈30 min training time.
• Atmospheric river classification achieved 90.5% train accuracy and 90% test accuracy with 6–7 hour training time.
• Weather front classification achieved 88.7% train accuracy and 89.4% test accuracy with ≈30 min training time.
• The model shows no over-fitting due to small shallow architecture and weight decay regularization.
• CNNs can learn climate patterns from labeled data without relying on subjective threshold criteria.