Tropical Cyclone Monitoring Based on Geostationary Satellite Imagery

Abstract

Due to the devastating wind, heavy floods, and coastal inundation from storm surges, tropical cyclones (TC) can cause significant damage to human lives and property. Forecasters and emergency responders both rely on accurate TC location and intensity estimates. In this chapter, two deep convolutional neural networks (CNNs) were designed for locating TC centers (CNN-L) and estimating their intensities (CNN-I) from the brightness temperature data observed by the Himawari-8 geostationary satellite. To train the CNNs, we used 97 TC incidents from the Northwest Pacific from 2015 to 2018. When compared to the Best Track dataset of TC centers and intensities as a reference, the mean location error of CNN-L model is 30 km for TCs in categories H1-H5. The best multi-category CNN classification model provided a pretty good accuracy (84.0%) and low root mean square error (RMSE, 10.19 kt) in TC intensity estimate using a combination of four channels of data as input. We added a focal_loss function to the CNN model to mitigate the negative effect of the severely imbalanced distribution of TC category samples. The accuracy increases to 88.9% if we convert the multi-classification problem to a binary classification problem, and the RMSE is dramatically lowered to 8.99 kt. Our CNN models are robust in assessing TC intensity from geostationary satellite pictures, according to the results.

1 Introduction

Tropical cyclones (TCs) are extreme weather processes developed over the tropical ocean. They are called typhoons in the northwestern Pacific and hurricanes in the eastern Pacific and Atlantic. The Northwest Pacific Ocean is the basin with the largest number of TCs in the world [5, 11], where TCs can be observed throughout the year [48]. TCs can cause huge losses to marine production and transportation, such as sweeping away the fishing nets, destroying the fish ravens and cages for breeding, leading to the death of a large number of fish and shellfish, destroying the breakwaters and offshore oil platforms, overturning ships and aircraft, etc. After landing, they usually cause hazardous disasters such as storm surges, mountain torrents, urban waterlogging, landslides and debris flows owing to the destructive wind and low pressure, which poses a serious threat to coastal areas and brings great damages to human life and property.

The annual number of TCs generated in the Northwest Pacific Ocean accounts for 36% of the total number of TCs in the world [55]. In this basin, TCs occur most frequently from July to September. They can easily develop into super TCs under the action of high temperature and high humidity. Under the influence of marine and atmospheric environment, TCs usually move from the southeast to the northwest. Many countries have listed the TC as one of the major natural disasters affecting national public security, and clearly proposed to strengthen the research on key technologies of TC monitoring and early warning [20, 54].

Locating the TC center and estimating its intensity have been widely considered as an important part of TC monitoring by meteorological forecasting agencies. Under the background of global warming, the intensity of global TCs shows a significant increasing trend [20]. Accurate TC center position and intensity information can better initialize the numerical models and make them more accurate to predict the intensity and movement direction of TCs, especially the rapidly strengthening TCs [20, 50], and thus help people to take precautions in advance and reduce losses.

Traditional TC monitoring platforms include coastal stations, buoys, oil drilling platforms, etc.. But the recorded TC data is very few due to the sparse distribution of these platforms. With the development of airborne remote sensing technology, real-time observation of TCs in a large space and time range is realized by using airborne radars or radiometers, but is still limited by extreme weather conditions. Since 1970, a major breakthrough has been made in satellite remote sensing. Many countries have launched geostationary meteorological satellites in succession, which can realize real-time and continuous observation of the earth. Spaceborne sensors generally have high spatial resolution and have become an important tool for TC monitoring. Although there are more and more observation methods, the monitoring of TCs still relies on the experience of experts to a certain extent. Particularly, there is a lack of objective and effective monitoring methods for weak TCs at the stage of formation or extinction.

A variety of TC center location methods based on infrared images, scatterometer data or synthetic aperture radar images have been developed. These methods can be divided into four categories. The first category is subjective method, which locates the TC center based on the forecaster’s experience judgment on the Central Dense TC Overcast in the satellite images [7,8,9,10, 37]. The second is the threshold method, which is used to segment and identify the TC’s eye area from satellite images and determine the morphological center of the eye area [2, 14, 28, 30, 32, 47]. The third is the spiral curve method. As is well known, the structure of the TC cloud system is not symmetrical, but spiral. The concept of spiral can be expressed by vector distance and the spiral center is the TC center [23, 51]. The fourth category is wind vector or cloud motion wind (CMW) method, which uses the wind field retrieved by scatterometers or time series of infrared satellite images to establish the relationship between the wind vector or CWM and the movement variation of a TC so as to locate the TC center [17, 21, 31, 33, 34, 40, 52, 57].

For TC intensity estimation using satellite images, two traditional methods are Dvorak technology [7, 10, 12, 19, 56] and empirical regression method [3, 16, 24, 26, 36, 38, 42,43,44]. The Dvorak technology was first proposed by Dvorak in 1975 [7] and has been operated for more than 40 years. The method assumes that the rotation and shape of a TC eye area are related to the strengthening and weakening of a TC, and TCs with similar intensities have similar morphological characteristics. In the empirical regression method such as the deviation-angle variance technique (DAVT) [3], the TC intensity is determined by establishing the relationship between the TC intensity and different characteristics. Both methods rely on high levels of artificial features converted from satellite images of TCs. However, it is difficult to obtain general characteristics of TCs and establish empirical regression models for TCs at different development stages and in different regions.

In the past few years, meteorologists and oceanographers have introduced the convolutional neural network (CNN) into TC monitoring. CNN is a kind of feedforward deep neural network, which is trained through back propagation algorithm, and its inspiration comes from the natural visual cognitive mechanism of organisms [22]. A complete CNN model consists of convolutional layers, pooling layers and full connection layers. The convolution layer is used to extract features from the image, the pooling layer filters and reduces the number of features, and the full connection layer learns the relationship between these features and the model output. The CNN model not only avoids the complex image preprocessing, but also does not depend on the priori knowledge of TCs, and thus can meet the requirements of automatic and objective TC center location and intensity estimation.

The CNN has achieved great success in image classification and target recognition [15, 22, 24, 26, 38, 42, 43, 46] and also shows great application potential in TC intensity estimation. Pradhan et al. [36] collected 8138 TC images over the North Pacific Ocean and the Atlantic Ocean from 1998 to 2012, and labelled the images into eight categories based on the Saffir-Simpson hurricane scale and the HURDAT2 Best Track dataset. They used the CNN model to classify the TCs and then the maximum wind speed (MWS) was calculated according to the probability of each TC category. The root mean square error (RMSE) of the derived MWS is 10.19 kt. However, there are green shorelines in the satellite images used in their study, which may affect the accuracy of TC intensity estimation [16]. Combinido et al. [6] used a CNN regression model to estimate TC intensity, but the RMSE of the MWS (13.23 kt) was larger than that of the CNN classification model proposed by [36]. Also with the CNN regression model, Chen et al. calculated the intensity of TCs in the global ocean based on more TC images [36, 44] and the RMSE was reduced to 10.58 kt. Recently, Tian et al. combined the CNN classification and regression models to estimate TC intensity [3], and they obtain a smaller RMSE of 8.91 kt.

The above results show that the CNN classification model or regression model performs well in TC intensity estimation. Compared with the Dvorak technology, the CNN model does not rely on the subjective judgment of the forecaster, which ensures the objectivity of the method. Compared with the empirical regression method, the neural network reduces the requirement of the knowledge level of the person who uses this method. In addition, although the TC center position is necessary in the CNN model, the information is only used to cut the input images. Therefore, compared with the empirical regression method, the CNN model does not require a very high TC positioning accuracy.

In conclusion, the accuracy of TC center location and intensity estimation technology has been greatly improved in recent years, but there are still some limitations. For operational TC monitoring, objective, we need a fast and robust objective method, and the CNN just meets this requirement. In this chapter, a set of CNN models were designed to determine the TC center position and intensity from Himawari-8 geostationary satellite images in the Northwest Pacific Ocean. By discussing the influence of the sensor channel, the number of satellite images and the configuration of the neural network on the model performance, an optimal CNN model was developed for automatic TC monitoring. The data and structure of the CNN model are described in Sect. 2. Sect. 3 and  4 present TC center location and intensity estimation results, respectively. Section 5 is the summary.

2 Data and Methodology

2.1 Data

The Japan Meteorological Agency (JMA) launched the Himawari-8 (H-8) geostationary meteorological satellite in October 2014. The Advanced Himawari Imager (AHI) onboard H-8 provides observations of different regions at different modes: Full Disk (global scope), Japan Area (scope of two Japanese regions), Specific Area (scope of two regions), and Landmark Area (scope of two regions). The Full Disk and Japan area are fixed, while the other two specific areas and landmark areas can be adjusted flexuously. The scanning range is shown in Fig. 1 ($60\,^\circ$S-$60\,^\circ$N, $80\,^\circ$E-$160\,^\circ$W). The on orbit working life of H-8 is 8 years [4]. After calibration and calibration test, the data were provided since July 2015 http://www.eorc.jaxa.jp/ptree.

Fig. 1

Himawari-8 satellite scanning range

AHI’s 16 channels cover the whole Northwest Pacific: three visible, three near-infrared, and ten thermal-infrared channels. Full-disk observations are taken every 10 minutes. The brightness temperature data from five infrared (IR, Channels 7, 8, 13, 14, and 15, see Table 1 for details from http://www.eorc.jaxa.jp/ptree.) channels was obtained.

Table 1 Himawari-8 channel setup [4]

Based on AHI data, we collected 6,690 satellite images of 97 TCs over the Northwest Pacific with a time interval of 3 hours during the whole life cycle of the TCs from 2015 to 2018, and the image resolution is 5 km. As shown in Fig. 2, the brightness temperature data of 5 channels (7, 8, 13, 14, 15) with higher transmittance near the large window (center wavelength are 3.9, 6.2, 10.4, 11.2, 12.4 $\mu$m, respectively) were selected to locate TC center and estimate TC intensity. Channel 7 is mainly used for observing cloud and natural disasters in the lower layer, channel 8 for observing water vapor content in the upper and middle layer, channel 13 for observing cloud images and cloud top conditions, and channels 14 and 15 are mainly used for observing cloud images and sea surface temperature [4].

Fig. 2

Bright temperature(unit: K) image of different channels of “Soudelor” obtained by AHI on Himawari-8 satellite at 18:00 (UTC) on August 15, 2015 a channel 7, b channel 8, c channel 13, d channel 14, e channel 15. The image space range is 1255 km$\times$1255 km

We use the TC Best Track dataset provided by the tropical cyclone information center of China Meteorological Administration (CMA) http://tcdata.TC.org.cn as ground truth data. The data set was compiled by Shanghai Typhoon Institute. The TC Best Track dataset over the Northwest Pacific from 1949 to 2019 includes the TC number and name, time, longitude and latitude of the TC center, the MSW (2 minute average wind speed), the minimum air pressure, etc.. The time interval is 6 hours before 2017 and has been encrypted to 3 hours for landing TCs since 2017. Starting from 2018, the Best Track data has provided 3-h TC information 24 hours before the landing activities [1, 45]. Consistent with H-8 satellite observations, we downloaded the Best Track data of 97 TCs over the Northwest Pacific during 2015-2018.

2.2 Data Pre—Processing

We selected H-8 satellite images synchronized with the 6-h Best Track dataset to design the CNN-based TC center location (CNN-L) model. There are 3298 images in total, among which 1971 are used for model training, 657 for validation and 670 for testing. For each training or validation image, we re-extracted a sub-image covering an area of 1500 km $\times$ 1500 km (pixel size 301 $\times$ 301), and then cut the pixel size from 301 $\times$ 301 to 151$\times$151 three times by moving it up, down, left or right randomly. Finally, we labeled the processed image with the number and direction of moving pixels. For example, if a TC center is shifted left by 5 pixels, and up by 10 pixels, the image is labeled (−5, 10). The test images were only cropped. Finally, we obtained 5913 training images, 1971 validation images, and 670 test images with a reduced size of 151$\times$151.

More training images may also help to improve the CNN-based TC intensity estimate (CNN-I) model performance. Hence, before the estimation of TC intensity, we interpolated the MSW provided by the Best Track dataset every 3 hours to increase the data samples. The total of 6690 3-h images with a size of 251 $\times$ 251 was labeled correctly according to eight categories using the Saffir–Simpson hurricane wind scale (H1 to H5) along with intensity categorization for the tropical storm (TS) and tropical depression (TD) as TC intensity categories (Table 2). The total 6690 images are divided into the training, validation and testing images with 4014, 1338 and 1338 images, respectively. Studies show that the rotation of images in a CNN model can reduce the sensitivity of orientation and does not affect the classification accuracy. Therefore, the normalized training and validation images were artificially rotated $90\,^\circ$, $180\,^\circ$ and $270\,^\circ$ clockwise. In this way, the number of images is increased by three times, and finally, we obtained 16,056 training images and 5352 validation images for the construction of the CNN based TC intensity estimation model.

Table 2 Saffir-Simpson hurricane wind scale and related classifications (The Saffir–Simpson Team, 2012). MSW is ten-minute averaged maximum sustained wind speed

2.3 Methodology

The convolutional neural network (CNN) model consists of convolutional layers (C), pooling layer (P) and fully connected layer (FC). The convolutional layer is used to extract image features. In order to reduce the number of features and speed up the calculation, the pool layer uses the filtering methods such as maximum, average and minimum values to process these features. The nonlinear relationship between the filtered features and the output results is learnt by the full connection layer. To prevent overfitting, a dropout layer is usually added before the full connection layer [26], which invalidates part of the connection between the upper and lower layers. CNN is usually designed as a feed forward network that can be trained with a backpropagation algorithm. When the error propagates back in the model, the network is optimized by updating the weights and deviations to minimize the value of the loss function [22].

Fig. 3

Framework of the CNN based TC center location model

The CNN model architecture for TC monitoring is shown in Fig. 3. It consists of one input layer, four convolutional layers, four pooling layers, one dropout layer, two fully connected layers, and one output layer. Table 3 lists the parameter setting of the CNN model as shown in Fig. 3. The parameters include the convolutional kernel shape, output shape and number of parameters, as well as the convolution kernel shape used in each convolution layer, the filter shape used in each pooling layer and the step size. Taking the CNN model shown in Fig. 3 as an example, the satellite image with a size of 151$\times$151 was input into the CNN model. The first convolution operation is performed on the satellite image. The size of each convolution kernel was 10$\times$10, and the step size was 2. In this process, the model needs to learn 3232 parameters; In the first pooling layer, 32 feature images with the size of 3$\times$3 and step size of 2 are filtered, and 32 feature images with the size of 38$\times$38 are obtained by using the maximum pooling method. After four convolutional layers and four pooling layers, 256 abstract feature images of the size of 3$\times$3 are finally generated. These feature images are expanded into 1-dimensional data of the size of 2304, input into the dropout layer (the dropout rate is 0.5), and then connected with the first fully connected layer. The model needs to learn 2360320 parameters. Finally, the model output is obtained in the output layer. In the whole training process, the model needs to learn a total of 2915298 parameters.

Table 3 The MLE of CNN based TC center location model

For TC center location and intensity estimation from multi-channel satellite images, the input channel, the image resolution and model parameters would affect the training efficiency and accuracy of the CNN model. By Setting up a group of sensitivity experiments to investigate the influence of different factors on the model performance, we aim to develop an optimal CNN model for TC monitoring.

3 TC Center Location

The CNN model for TC center location, i.e., the CNN-L model, as shown in Figure 3.3 is established in this section. We use the mean location error (MLE) to evaluate the performance of the model:

$$\begin{aligned} MLE = \frac{\sum \limits _{i=1}^n\sqrt{(x_i-x'_i)^2+(y_i-y'_i)^2}}{n} \end{aligned}$$

(1)

where n is the number of test images; (x, y) and (x’, y’) are TC center positions from the Best Track data and CNN-L model outputs, respectively.

Fig. 4

Himawari-8 Channel 7 images of TCs: a Halola (9:00 UTC on 13 July 2015), b Soudelor (18:00 UTC on 30 July 2015), c Atsani (12:00 UTC on 16 August 2015), d Meranti (12:00 UTC on 14 September 2016)

As shown in Fig. 4, there are sometimes some noises in the image of channel 7 or 8 due to the sensor calibration problem, which would seriously affect the TC location accuracy. Therefore, we only use the images of channels 13, 14 and 15 as the input of the CNN-L model.

We take TC “Maria” occurring during 4 July to 11 July, 2018 as an example. Figure 5 shows the TC center location results using the network configuration listed in Table 3. The mean location errors between different CNN-L model outputs and observations are similar, suggesting that the image channel has little effect on the accuracy of the TC center location model.

Fig. 5

CNN based TC center location model results for TC “Maria” occurring from 4 July to 11 July, 2018 at different stages: a Tropical depression, b Tropical storm, c Category 1, d Category 2, e Category 3, f Category 4. The red, yellow and purple dots represent the TC center positions determined by the CNN-L model with brightness temperature of Channels 13, 14 and 15 as the input, respectively. The blue dot shows the TC center position from the Best Track dataset provided by CMA

Using Channel 15 data as the input of the CNN-L 3 model, a variety of network configurations listed in Table 4 were further tested. CNN-L 3, CNN-L 4, and CNN-L 5 models all consist of three or four convolutional layers, three or four pooling layers, and two FC layers. However, each model has a different number of kernels in the convolutional layers. The stride and zero-padding, the shape of the FC layer, and the dropout in each model are listed in Table 4. One can see that CNN-L 3 produces the lowest MLE, indicating that too many or few convolutional layers do not help to improve the CNN-L model.

Table 4 Results of the CNN based TC center location models with different parameters(For example, C1 means the first convolutional layer and P1 means the first pooling layer, C1(32@3 $\times$ 3) denotes 32 kernels in the first convolutional layer with a size of 3 $\times$ 3)

The results of CNN-L 3 model for different categories of TCs are shown in Figs. 5, 6 and Table 5. Strong TCs generally demonstrate a more distinct and stable structure and more obvious TC eye area than weak TCs. As a result, the mean location error of the CNN-L 3 model also decreases rapidly with the increase of TC intensity. The average MLEs of H1-H5 and H4-H5 TCs are 30 km and less than 25 km, respectively. As shown in Table 5, the accuracy of our CNN based TC center location model is comparable to that of some techniques that also locate TCs from IR images with spatial resolutions of 2.5–12.5 km (Table 6).

Fig. 6

Mean location error (km) of the CNN-L 3 model for different categories of TCs. Numbers in brackets represent the number of samples in the test group for this category

Table 5 Mean location error (km) of the CNN-L 3 model

Table 6 Performance of the CNN model in this study and other methods for TC center location

4 TC Intensity Estimation

Similar to Sec.3, the influence of the selection of input data and model parameters on the performance of the CNN based TC intensity estimation model, i.e., the CNN-I model shown in Fig. 7, is investigated in this section. The possible solution of the side effects of imbalanced dataset is also discussed.

Fig. 7

Framework of the CNN based TC intensity estimation model

We evaluate the performance of the CNN-I model from three aspects:

(1) Accuracy

The number of exact-hits, which refers to the correct classification of a TC with the highest confidence, is the accuracy metric.

(2) Root mean square error (RMSE) and mean average error (MAE) of TC intensity

For categories TD through H4, we define the estimated intensity or MSW of a TC as the weighted average of the two highest categories for their probabilities. Otherwise, we use the mean speed of the category that has the highest confidence. The MSW (W) is evaluated as [35]:

$$\begin{aligned} W = U _1 \times P _1 + U _2 \times P _2 \end{aligned}$$

(2)

where P₁ and P₂ are the probabilities which output by the model of the TC categories with the highest and second-highest confidence, respectively; U₁ and U₂ are the mean wind speed of the corresponding category.

(3) Confusion matrix and classification report

As shown in Table 9, The confusion matrix depicts a model’s overall classification performance. The number along the diagonal line in a confusion matrix represents the number of correctly identified images for any category. A CNN model’s precision (P), recall (R, or confidence of detection), and f1-score (F1) are described in the classification report. The ratios of real positive class values to total positive classifications and the number of positive class values in the test data, respectively, are P and R. F1 = 2P $\times$ R/(P + R) is the harmonic mean of recall and precision.

Table 7 shows the results of the CNN based TC intensity estimation model based on 1338 test images with the model parameters shown in Fig. 7.

Table 7 Results of the CNN based TC intensity estimation model with images of different channels as the input

As shown in Table 7. for CNN-I models with single-channel input, Channels 13-15 can achieve a higher accuracy. Channels 14 and 15 correspond to the thermal-IR bands, which are often used to observe the cloud and calculate surface temperature (SST). From a different perspective, the data support the conclusion of [41] that TC intensity is connected with the temperature deficit of cloud top against sea surface. We first combined Channel 15 with the other channels in CNN-I 6 through CNN-I 9 models to examine the use of multi-channel combination in TC intensity estimate. Comparing CNN-I 9 with CNN-I 5, we can see that the combination of Channels 15 with 7, 8 or 13 improves the model performance, while Channel 14 has little contribution to the improvement of the model because its wavelength is close to Channel 15. These two thermal infrared channels may provide redundant information for TC intensity estimation, resulting in data redundancy.

With the input of more (3 or 4) channels of data, the 4-channel model CNN-I 12 produces the best result with an accuracy of 82.9% and a RMSE of 10.64 kt. This is also consistent with some theoretical studies, indicating that the information of water vapor, cloud characteristics, the brightness temperature difference of cloud top and sea surface provided by the combination of channels 7, 8, 13 and 15, plays an important role in TC intensity estimation [29, 49, 53]. Through the combined input of multi-channel satellite images, the CNN-I model can learn the complex nonlinear relationship between various elements and TC intensity.

Table 8 Results of the CNN based TC intensity estimation model with different parameters

A variety of network configurations were tested further using the CNN-I 12 model, and the results are showed in Table 8. Four convolutional layers, four pooling layers, and two fully connected layers make up CNN-I 13-16 models. For CNN-I 12 to CNN-I 15, the number of convolution kernel in the convolutional layer increases gradually. The model’s accuracy improves slightly as the number of convolution kernels increases (CNN-I 13), but it decreases dramatically beyond a certain range of kernel numbers (CNN-I 15). Although more kernels allow for the extraction of more feature maps, these maps may not always have a favorable impact on the model’s improvement. In addition, the decrease of the model performance with the increasing number of convolution kernel may also be associated with the number of training samples. More complex CNN models need to learn more parameters. It is difficult to obtain a high accuracy when the difference between the training sample size and the number of model parameters is too large. Compared with the CNN-I 12 model, CNN-I 14 changes the step size of the convolution kernel operation, but the accuracy is reduced.

Fig. 8

Schematic diagram of spatial and channel attention layers in the CNN based TC intensity estimation model a, channel attention b, and spatial attention c. “+” and “$\times$” are plus and multiply signs, respectively

Recently, Woo et al proposed the spatial attention and channel attention mechanism, which is based on the study of human vision [54]. As show in Fig. 8, after adding the spatial and channel attention layers, the CNN-I 17 model gives the highest accuracy (86.0%) and the lowest RMSE (10.06 kt) of the MWS. The CNN based TC intensity estimation model can focus on the key factors revealed by the attention mechanism, which helps to improve the accuracy of the model.

In general, the CNN-I 16 multi-category classification model does a good job at estimating TC intensity. However, as demonstrated in Tables 9, 10, the classification results for TC categories with little training samples are not particularly satisfactory. For example, there are 62 samples in the H3 category, but only 48 have been identified correctly. Only 77.0% of the H3 category is accurate. The degradation is caused by an imbalance in the training data across different types of TC datasets. Taking H3 as an example, the H3 category accounts for only 2.3% of total training numbers. Even if most of the H3 images are misclassified during the training step of the CNN-based TC intensity estimation model, the loss will only increase somewhat.Because the CNN adjusts the weight value of each layer in response to the loss, the network will struggle to learn the features of a category if there are few samples. As a result, the accuracy of the CNN-I model for H2 category is lower.

Table 9 Confusion matrix of the CNN-I 16 multi-classification model

Table 10 Classification report of the CNN-I 16 binary classification model

We use Focal_loss function to replace the original loss function in the CNN TC intensity estimation model. This function aids a model’s learning of features by raising the category’s weight with fewer data in loss, and has demonstrated excellence in the field of target recognition. In this way, In the event of limited samples, the model can better learn the TC category’s relevant attributes. The Focal_loss function’s definition is as follows: [27]:

$$\begin{aligned} FL (p _{t}) = -a _{t}(1-p _{t})^r log(p _{t}) \end{aligned}$$

(3)

where p_t is the output of model for NC to H5 category, a_t is the weight coefficient which is determined by the proportion of the number of NC to H5 category to the total data, and r is the empirical parameter. The values of a_t and r for each TC category used in this study are shown in Table 11. The accuracy of the TC intensity estimation model (CNN-I 17) using the Focal_loss function is improved to 86.6%, and the RMSE is reduced by 2.1%.

In many domains, multi classification can be transformed into target recognition or binarization problem [18]. In this study, we further used eight binary models to replace the multi-classification model, and each binary model can learn the TC characteristics corresponding to each intensity category. Eight CNN based TC intensity estimation binary models were constructed to identify NC to H5 category, respectively.

Table 11 Values of a_t and r in FOCAL_LOSS function

The configuration of each model in Table 12 is the same as that of CNN-I 16. The Focal_loss with values of a_t and r listed in Table 11 was also used. We changed the classification label to “1” or “0”, which represents the intensity of whether a TC corresponds to a particular category. If the maximum sustained wind speed of a TC sample is 52.0 kt, which belongs to the TS category, the corresponding label of this image is “1” in the binary model which is responsible for judging the TS category, and "0" in the other binary models.

Table 12 Results of the CNN based TC intensity estimation multi models with/without Focal_loss function and binary model

As shown in Table 12, the CNN-I 18 binary model has a much higher performance than that of the multi-classification model CNN-I 16. Compared with CNN-I 17, the accuracy of CNN-I 18 is improved to 88.9%, and the RMSE is reduced to 8.99 kt. The results show that the introduction of the Focal_Loss function and the transformation of the multi-classification model into eight binary classification models helps to reduce the side effects caused by the imbalanced dataset.

Table 13 Confusion matrix of the CNN-I 18 binary classification model

The CNN-I 18 binary classification model’s confusion matrix and classification report are shown in Tables 13 and 14. The number of exact-hits for NC, TD, TS, H1, H2, H3, and H4 all rose when compared to the results from the multi-classification model (Tables 9 and 10). The precisions of H1-H4 classification have improved by 4.9%, 7.4%, 9.1%, and 3.8%, respectively.

Table 14 Classification report of the CNN-I 18 binary classification model

Table15 compares the performance of the CNN model and other TC intensity estimation methods. The RMSE of the maximum wind speed estimated by the CNN based TC intensity estimation model proposed in this study is smaller than that of the DAVT technique or most CNN regression or classification models, which proves the potential of the CNN method in TC monitoring.

Table 15 Comparison of our model and other methods

5 Summary

Accurately locating the TC center and estimating its intensity is an essential step for forecasters and emergency responders to make disaster warnings. In this chapter, a set of CNN-based model has been developed to automatically identify TC’s center (CNN-L model) and intensity (CNN-I model) from H-8 geostationary satellite IR imagery, which can provide a reliable technical and information support for TC prediction and early warning systems.

Results show that the selection of satellite image channels has a significant impact on the performance of the TC intensity estimation model but hardly affects the TC center location model. Network parameters play an essential role in both models. The mean distance between the TC centers identified by the CNN-L model and by the Best Track dataset is 30 km for TCs in categories H1–H5. The accuracy of our CNN-L model is comparable to some techniques that locate a TC center based on its morphological features in IR images. Using four-channel (Channels 7, 8, 13, and 15) IR imagery, we found that the CNN-I 16 model has the best performance among the multi-classification models.

For TC categories with smaller training datasets, due to the unbalanced distributions of TC categories, the multi-classification model cannot produce a very good result. By introducing the Focal_loss function in the CNN model and adopting eight binary classification networks, the side-effect of the unbalanced training data is reduced. In TC intensity estimate, the binary classification model CNN-I 18 gives a substantially lower RMSE (8.99 kt) of the maximum wind speed than the multi-classification model.