We propose a deep learning approach for identifying tropical cyclones (TCs) and their precursors. Twenty year simulated outgoing longwave radiation (OLR) calculated using a cloud-resolving global atmospheric simulation is used for training two-dimensional deep convolutional neural networks (CNNs). The CNNs are trained with 50,000 TCs and their precursors and 500,000 non-TC data for binary classification. Ensemble CNN classifiers are applied to 10 year independent global OLR data for detecting precursors and TCs. The performance of the CNNs is investigated for various basins, seasons, and lead times. The CNN model successfully detects TCs and their precursors in the western North Pacific in the period from July to November with a probability of detection (POD) of 79.9–89.1% and a false alarm ratio (FAR) of 32.8–53.4%. Detection results include 91.2%, 77.8%, and 74.8% of precursors 2, 5, and 7 days before their formation, respectively, in the western North Pacific. Furthermore, although the detection performance is correlated with the amount of training data and TC lifetimes, it is possible to achieve high detectability with a POD exceeding 70% and a FAR below 50% during TC season for several ocean basins, such as the North Atlantic, with a limited sample size and short lifetime.
Gravity waves play an essential role in driving and maintaining global circulation. To understand their contribution in the atmosphere, the accurate reproduction of their distribution is important. Thus, a deep learning approach for the estimation of gravity wave momentum fluxes was proposed, and its performance at 100 hPa was tested using data from low‐resolution zonal and meridional winds, temperature, and specific humidity at 300, 700, and 850 hPa in the Hokkaido region (Japan). To this end, a deep convolutional neural network was trained on 29‐year reanalysis data sets (JRA‐55 and DSJRA‐55), and the final 5‐year data were reserved for evaluation. The results showed that the fine‐scale momentum flux distribution of the gravity waves could be estimated at a reasonable computational cost. Particularly, in winter, when gravity waves are stronger, the median root means square errors (RMSEs) of the maximum momentum flux and the characteristic zonal wavenumber were 0.06–0.13 mPa and 1.0 × 10−5, respectively.
This dataset consists of 3500 images of beach litter and 3500 corresponding pixel-wise labelled images. Although performing such pixel-by-pixel semantic masking is expensive, it allows us to build machine-learning models that can perform more sophisticated automated visual processing. We believe this dataset may be of significance to the scientific communities concerned with marine pollution and computer vision, as this dataset can be used for benchmarking in the tasks involving the evaluation of marine pollution with various machine learning models. The beach litter images were obtained from coastal environment surveys conducted between 2011 and 2019 by the Yamagata Prefectural Government, Japan. These images were originally obtained owing to the reporting guidelines concerning regular coastal-environmental-cleanup and beach-litter-monitoring surveys. Based on these images, the Japan Agency for Marine-Earth Science and Technology created 3500 images comprising eight classes of semantic masks for beach litter detection [1].
15 December 2021 in Progress in Earth and Planetary Science
Image data classification using machine learning is an effective method for detecting atmospheric phenomena. However, extreme weather events with a small number of cases cause a decrease in classification prediction accuracy owing to the imbalance in data between the target class and the other classes. To build a highly accurate classification model, I held a data analysis competition to determine the best classification performance for two classes of cloud image data, specifically tropical cyclones including precursors and other classes. For the top models in the competition, minority data oversampling, majority data undersampling, ensemble learning, deep layer neural networks, and cost-effective loss functions were used to improve the classification performance of the imbalanced data. In particular, the best model of 209 submissions succeeded in improving the classification capability by 65.4% over similar conventional methods in a measure of the low false alarm ratio.
In this study, a stationary front is automatically detected from weather data using a U-Net deep convolutional neural network. The U-Net trained the transformation process from single/multiple physical quantities of weather data to detect stationary fronts using a 10-year data set. As a result of applying the trained U-Net to a 1-year untrained data set, the proposed approach succeeded in detecting the approximate shape of seasonal fronts with the exception of typhoons. In addition, the wind velocity (zonal and meridional components), wind direction, horizontal temperature gradient at 1000 hPa, relative humidity at 925 hPa, and water vapor at 850 hPa yielded high detection performance. Because the shape of the front extracted from each physical quantity is occasionally different, it is important to comprehensively analyze the results to make a final determination.