A mask R-CNN model for reidentifying extratropical cyclones based on quasi-supervised thought

The applications of machine learning/deep learning (ML/DL) methods in meteorology have developed considerably in recent years. Massive amounts of meteorological data are conducive to improving the training effect and model performance of ML/DL, but the establishment of training datasets is often time consuming, especially in the context of supervised learning. In this paper, to identify the two-dimensional (2D) structures of extratropical cyclones in the Northern Hemisphere, a quasi-supervised reidentification method for extratropical cyclones is proposed. This method first uses a traditional automatic cyclone identification method to construct a trainable labeled dataset and then reidentifies extratropical cyclones in a quasi-supervised fashion by using a (pre-trained) Mask region-based convolutional neural network (Mask R-CNN) model. In comparison, the new method increases the number of identified cyclones by 8.29%, effectively supplementing the traditional method. The newly recognized cyclones are mainly shallow or moderately deep subsynoptic-scale cyclones. However, a considerable portion of the new cyclones along the coastlines of the oceans are accompanied by strong winds. In addition, the Mask R-CNN model also shows good performance in identifying the horizontal structures of tropical cyclones. The quasi-supervised concept proposed in this paper may shed some light on accurate target identification in other research fields.

Scientific RepoRtS | (2020) 10:15011 | https://doi.org/10.1038/s41598-020-71831-z www.nature.com/scientificreports/ spatial profile) based on the Faster R-CNN model (usually used to classify and locate objects) and outperforms existing object pixel-level detection models. Zhang et al. 31 applied the Mask R-CNN model to the classification and identification of Arctic ice-wedge polygons because of its simple concept and superior capability. The Mask R-CNN model has been shown a good performance in identifying the object-shape in the field of computer vision as well as the geophysics research. This inspires us to apply the model to the identification of the extratropical cyclone, considering the variety and complexity of the structure of extratropical cyclone. Although DL algorithms have achieved outstanding results (some of which even surpass the human level) on many supervised learning and object identification problems, constructing a large-scale labeled database to train a DL model (such as the Mask R-CNN model) is still a challenging task 32 . Therefore, this paper proposes a method to construct a more reliable labeled dataset for the training of DL models (i.e., the Mask R-CNN model) by using the traditional automatic cyclone identification method instead of a manual labeling scheme. On the other hand, the identification of cyclone extent can effectively reduce the complexity of tracking cyclone-merging or splitting events (e.g., Hewson 33 ; Hanley and Caballero 34 ). In this paper, first, the CAA proposed by Lu 9 is used to identify the 2D structure of an extratropical cyclone (north of 20° N), and a 2D extratropical cyclone dataset (quasi-ground-truth) is constructed for training. The Mask R-CNN model is then applied to further identify 2D extratropical cyclones; we refer to this algorithm as an extratropical cyclone quasi-supervised reidentification method. Several previous studies have shown that DL models can provide reliable results when processing complex multi-dimensional meteorological data [18][19][20] . Our method solves the problem of constructing a large-scale labeled database for DL models by using traditional identifying algorithms, that may efficiently improve the practical efficiency of DL models. Therefore, the performance of the proposed method on extratropical cyclones and its applications in the identification of tropical cyclones are discussed.

Data and methods
Data. The data used in this study consist of the ERA-Interim dataset 35  Identification process. The general workflow of the extratropical cyclone quasi-supervised reidentification method based on the Mask R-CNN model includes two main steps (Fig. 1). (1) The CAA is used to initially generate the cyclone's regime (mask); this algorithm simultaneously outputs the boundary of each cyclone, detected by the outermost enclosed contour of the cyclone, while searching for the cyclone center by expanding outward from the center until the outermost enclosed contour is found 9 . (2) Both the cyclone masks output by the CAA and the 850 hPa geopotential height field data are processed as grayscale images to construct a training  www.nature.com/scientificreports/ North Atlantic, the Mediterranean as well as Northeast China. The high consistency of results from our model with the ones from Wernli and Schwierz 7 indicates the robustness of our methodology in this paper. However, our values are comparatively higher than the ones from Wernli and Schwierz 7 . This is because we regard the multi-center cyclone system as one individual cyclone, while a multi-center cyclone system is separated into several cyclones with only one local minimum center in each cyclone in Wernli and Schwierz 7 . Therefore, the values are generally higher in our model. To illustrate the overall motion characteristics of a newly added cyclone, the ratio of the number of grid points with a positive relative vorticity in the cyclone's regime to the number of grid points in the entire cyclone mask is defined as the positive vorticity ratio (PVR). Here, the PVR can be expressed as Although high-resolution data of a relative vorticity field can be very noisy 36 , a higher PVR value of a single cyclone generally denotes a stronger counterclockwise rotation of air flow within the cyclone's area. Therefore, a high PVR represents strong cyclonic motion. Table 1 shows the distribution of newly added cyclones in different PVR ranges. Most of the cyclones reidentified by the Mask R-CNN model are characterized mainly by counterclockwise rotational motion; among them, 50,416 cyclones (~ 86.54%) are in the range of PVR ≥ 70%. This means that the Mask R-CNN model is good at learning/describing the horizontal structure of an extratropical cyclone and can be used as an effective complement to the CAA. It should be noted that both the CAA and the Mask R-CNN model have a certain proportion of cyclone results with PVR < 50% (0.64% and 1.79%, respectively). As mentioned above, the high resolution of the relative vorticity field could introduce chaotic signals into the cyclone's regime. Additionally, local terrain or uneven heating could also result in small-scale negative vorticity inside a low pressure system.
The newly added cyclones are located mainly in the western and central regions of Eurasia (WCE, 20° W-80° E, 20°-60° N), accounting for 35.6% of all newly identified extratropical cyclones. As shown in Fig. 3a, areas with large proportions of newly added cyclones are found in the mountains of WCE (the Armenian Plateau, Zagros Mountains, and Hindu Kush Mountains) and the Atlas Mountains in northwestern Africa, while the areas with the second-highest proportions are distributed mainly along the oceanic coastlines (for example, along the coastlines of Western Europe and the Mediterranean Sea). On the other hand, the supplements along the two major storm tracks in the North Pacific and the North Atlantic are inconspicuous (figure not shown). This is probably because these storm track regions are located mainly on the ocean surface, resulting in a relatively symmetrical cyclone shape. Both the Mask R-CNN model and CAA can identify these cyclones well, and the supplementary effect of reidentification is not notable. We also applied the Mask R-CNN to the cyclone identification with NCEP I and JRA55. Consistently, the cyclone frequency in NCEP I and JRA 55 agree well with ERA-interim. WCE is still the most newly cyclone-added regions (43.7%, 39.2% for NCEP I and JRA55 respectively). Furthermore, the spatial distributions of newly added cyclones with PVR ≥ 70% in NCEP I and JRA 55 are also identical to ERA-interim (Fig. 3b,c). Hence, the following analysis of these newly identified cyclones will focus on WCE based on ERA-interim reanalysis dataset.
Compared to the identified cyclones in the CAA results, the newly added cyclones in WCE are relatively weak. As shown in Fig. 4a, a large proportion (84.60%) of the minimum geopotential heights of cyclones (denoting   Fig. 4b, both the Mask R-CNN model and the CAA focus mainly on the detection of cyclones with a subsynoptic scale (300-1,000 km) and a synoptic scale (1,000-2,000 km). However, the proportion of newly added cyclones at the subsynoptic scale identified by the Mask R-CNN model is notably higher than that identified by the CAA. Therefore, the new cyclones identified by the Mask R-CNN model are mostly located at the shallow or moderate subsynoptic scale. Note that, most of the newly added cyclones occurred in the Iranian Plateau and west of Tibet Plateau, and the cyclones identified over high terrain are generally shallow, which satisfies common sense. Figure 3 shows that most newly added cyclones occur predominantly over mountainous areas, which may be due to the filtering out of terrain at high elevations (> 1,500 m) and the Tibetan Plateau region (20°-45° N, 65°-110° E) in the CAA 9 . In fact, different traditional automatic cyclone identification algorithms suffer from great uncertainties in identifying cyclones over mountainous areas because they deal with mountains using different strategies 11 . To alleviate the influences of local artificial lows and reduce the complexity of the algorithm, some algorithms directly filter out mountainous terrain with different thresholds. However, we found that 14.28% of newly identified cyclones with PVR ≥ 70% in WCE are located in high-elevation mountainous areas (> 1,500 m). It is worth noting that simply filtering out high-elevation terrain may ignore some high-impact extratropical cyclones over such mountainous areas. For example, Fig. 5a,b show two of the strongest newly added cyclones identified over high terrain (> 1,500 m). These cyclones are accompanied by clearly cyclonic winds and distinct local precipitation. Therefore, the Mask R-CNN model could be an effective way to objectively detect extratropical cyclones over mountains.
In addition to mountainous cyclones, many new cyclones are identified near the coastlines of the oceans. These coastal cyclones could cause severe winds and other disasters (e.g., Pinto and Silva 37 ). For example, the new cyclones in the coastal region of Western Europe (20°-8° W, 32°-42° N, red box in Fig. 3) are accompanied by relatively high wind speeds (Fig. 6). In particular, although these cyclones are mostly subsynoptic-scale cyclones (approximately 89.02%), the maximum 6-h-mean 1,000 hPa wind speeds of 44.9% of all coastal cyclones are over 10.8 m/s (above the level of a strong breeze). Since the meso-or subsynoptic-scale cyclones are usually with short-lived life cycles, they are generally filtered to omit some of the local heat lows. However, because of relatively high wind speeds related to the coastal cyclones, the spatiotemporal variation of newly added cyclones in the coastal region of Western Europe deserved further investigation.
The nearest-neighbor method is widely applied to track cyclones. However, it is difficult for the nearestneighbor method to detect cyclones when more than two points in a particular time frame become merged into a single point in the following time frame (or vice versa). Therefore, some methods for the identification of multicenter cyclones (MCCs) have been proposed to detect the merging and splitting of cyclones (e.g., Inastu 8 ; Hanley and Caballero 34 ; Lu 9 ). Among the newly detected cyclones with PVR ≥ 70% in WCE, 42.21% of them are MCCs. Figure 7a,b display the two strongest new cyclones with the largest horizontal scale; both cyclones clearly show multicenter structures. Furthermore, both MCCs have a uniform overall cyclonic circulation accompanied by obvious local precipitation.
Tropical cyclone (TC) identification. Since the Mask R-CNN model also has great portability, we further apply this model to identify the 2D structures of tropical cyclones (TCs); the model is extended in comparison with Hong et al. 19 , who employed CNNs to track the eyes of typhoons. According to the typhoon track dataset  Table 2, the matching rate of tropical depressions (TDs) is 78.79%. As the intensity of the TC becomes stronger, the matching rate exceeds 90%. Taking the lowest pressure in the identified TC area as the center of the TC, Fig. 8 shows the tracks and intensities of typhoon events No. 24 and No. 25 in 2017 based on the Mask R-CNN model. In these two cases, the tracks of these two TCs and their 2D structures are highly consistent with the manually identified typhoon tracks based on the SLP dataset and are also in good consistency with the CMA typhoon track records. Based on the above results, the transfer learning capability of the Mask R-CNN model could be effectively applied to objectively identifying the 2D structures of TCs. Note that the TC extend is an arguably realistic and reliable quantification of the system strength. This characteristic can be used to study the local physical relationship between cyclones and extreme precipitation 38,39 .     www.nature.com/scientificreports/ labeled dataset and then reidentifies the 2D structures of extratropical cyclones in a quasi-supervised fashion by using a (pre-trained) Mask R-CNN model. We found that our quasi-supervised reidentification method for extratropical cyclones based on the Mask R-CNN model adds 58,260 new cyclones from 1979 to 2013. As measured by their PVR values, most of these new cyclones display obvious counterclockwise rotational motion characteristics, with 86.54% of these cyclones having PVR ≥ 70%. These new cyclones are located mainly (~ 35.6%) in the mountainous areas of WCE, including the Atlas Mountains in northwestern Africa and along the coastlines of Western Europe and the Mediterranean Sea. Approximately 81.8% of all newly added cyclones are subsynoptic-scale cyclones, most of which are shallow or moderately deep, and 14.28% of the new cyclones are situated above high-elevation (> 1,500 m) mountainous areas. In addition, we found that the new cyclones affecting the coastal areas of Western Europe mostly occur at the subsynoptic scale but with relatively strong wind speeds and potentially high impacts for these areas.
The quasi-supervised method based on the Mask R-CNN model also has a good transfer learning ability for identifying TCs. In particular, the Mask R-CNN model can effectively capture the track and 2D structure of a TC (in comparison with manual identification approaches) with a matching rate above 90%. Therefore, the quasisupervised concept proposed in this paper may shed light on target recognition tasks in other research fields by using classic automatic algorithms for the construction of training datasets to improve the practical efficiency of ML/DL and the reliability of object recognition.