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.

No abstract

We propose a super-resolution (SR) simulation system that consists of a physics-based meteorological simulation and an SR method based on a deep convolutional neural network (CNN). The CNN is trained using pairs of high-resolution (HR) and low-resolution (LR) images created from meteorological simulation results for different resolutions so that it can map LR simulation images to HR ones. The proposed SR simulation system, which performs LR simulations, can provide HR prediction results in much shorter operating cycles than those required for corresponding HR simulation prediction system. We apply the SR simulation system to urban micrometeorology, which is strongly affected by buildings and human activity. Urban micrometeorology simulations that need to resolve urban buildings are computationally costly and thus cannot be used for operational real-time predictions even when run on supercomputers. We performed HR micrometeorology simulations on a supercomputer to obtain datasets for training the CNN in the SR method. It is shown that the proposed SR method can be used with a spatial scaling factor of 4 and that it outperforms conventional interpolation methods by a large margin. It is also shown that the proposed SR simulation system has the potential to be used for operational urban micrometeorology predictions.

We have proposed a deep convolution neural network (CNN) approach for the accurate estimation of the cloud coverage (CC) from images captured by a consumer camera, i.e., snapshot pictures. This CNN can successfully estimate the CC to within the level of the inherent error in the training dataset. A segmentationbased method using a linear support vector machine (SVM) is shown to be unable to distinguish between water surfaces and the sky, while the present CNN can correctly distinguish between them, possibly because the CNN can understand the positioning of components in the images; the sky is over a water surface. The present CNN can also be applied to photo-realistic computergraphic (CG) images from numerical simulations. Comparisons between the CNN estimates for camera images and for the CG images can provide useful information for data assimilation, and thus contribute to numerical weather forecasting. The CC is a sort of far-field (remote) information. The present CNN has the potential to allow consumer cameras to be used as remote weather sensors.(Citation: Onishi, R., and D. Sugiyama, 2017: Deep convolutional neural network for cloud coverage estimation from snapshot camera images. SOLA, 13, 235−239,

Detecting seismic events, discriminating between different event types, and picking P- and S-wave arrival times are fundamental but laborious tasks in seismology. In response to the ever-increasing volume of seismic observational data, machine learning (ML) methods have been applied to try to resolve these issues. Although it is straightforward to input standard (time-domain) seismic waveforms into ML models, many studies have used time–frequency-domain representations because the frequency components may be effective for discriminating events. However, detailed comparisons of the performances of these two methods are lacking. In this study, we compared the performances of 1D and 2D convolutional neural networks (CNNs) in discriminating events in datasets from two different tectonic settings: tectonic tremor and ordinary earthquakes observed at the Nankai trough, and eruption signals and other volcanic earthquakes at Sakurajima volcano. We found that the 1D and 2D CNNs performed similarly in these applications. Half of the misclassified events were misassigned the same labels in both CNNs, implying that the CNNs learned similar features inherent to the input signals and thus misclassified them similarly. Because the first convolutional layer of a 1D CNN applies a set of finite impulse response (FIR) filters to the input seismograms, these filters are thought to extract signals effective for discriminating events in the first step. Therefore, because our application was the discrimination of signals dominated by low- and high-frequency components, we tested which frequency components were effective for signal discriminations based on the filter responses alone. We found that the FIR filters comprised high-pass and low-pass filters with cut-off frequencies around 7–9 Hz, frequencies at which the magnitude relations of the input signal classes change. This difference in the power of high- and low-frequency components proved essential for correct signal classifications in our dataset. Graphical Abstract