Research article Special Issues

A time series image prediction method combining a CNN and LSTM and its application in typhoon track prediction


  • Typhoon forecasting has always been a vital function of the meteorological department. Accurate typhoon forecasts can provide a priori information for the relevant meteorological departments and help make more scientific decisions to reduce the losses caused by typhoons. However, current mainstream typhoon forecast methods are very challenging and expensive due to the complexity of typhoon motion and the scarcity of ocean observation stations. In this paper, we propose a typhoon track prediction model, DeepTyphoon, which integrates convolutional neural networks and long short-term memory (LSTM). To establish the relationship between the satellite image and the typhoon center, we mark the typhoon center on the satellite image. Then, we use hybrid dilated convolution to extract the cloud features of the typhoon from satellite images and use LSTM to predict these features. Finally, we detect the location of the typhoon according to the predictive markers in the output image. Experiments are conducted using 13, 400 satellite images of time series of the Northwest Pacific from 1980 to 2020 and 8420 satellite images of time series of the Southwest Pacific released by the Japan Meteorological Agency. From the experimentation, the mean average error of the 6-hour typhoon prediction result is 64.17 km, which shows that the DeepTyphoon prediction model significantly outperforms existing deep learning approaches. It achieves successful typhoon track prediction based on satellite images.

    Citation: Peng Lu, Ao Sun, Mingyu Xu, Zhenhua Wang, Zongsheng Zheng, Yating Xie, Wenjuan Wang. A time series image prediction method combining a CNN and LSTM and its application in typhoon track prediction[J]. Mathematical Biosciences and Engineering, 2022, 19(12): 12260-12278. doi: 10.3934/mbe.2022571

    Related Papers:

    [1] Nicola Bellomo, Francesca Colasuonno, Damián Knopoff, Juan Soler . From a systems theory of sociology to modeling the onset and evolution of criminality. Networks and Heterogeneous Media, 2015, 10(3): 421-441. doi: 10.3934/nhm.2015.10.421
    [2] Yao-Li Chuang, Tom Chou, Maria R. D'Orsogna . A network model of immigration: Enclave formation vs. cultural integration. Networks and Heterogeneous Media, 2019, 14(1): 53-77. doi: 10.3934/nhm.2019004
    [3] Nicola Bellomo, Sarah De Nigris, Damián Knopoff, Matteo Morini, Pietro Terna . Swarms dynamics approach to behavioral economy: Theoretical tools and price sequences. Networks and Heterogeneous Media, 2020, 15(3): 353-368. doi: 10.3934/nhm.2020022
    [4] Fabio Camilli, Italo Capuzzo Dolcetta, Maurizio Falcone . Preface. Networks and Heterogeneous Media, 2012, 7(2): i-ii. doi: 10.3934/nhm.2012.7.2i
    [5] Juan Pablo Cárdenas, Gerardo Vidal, Gastón Olivares . Complexity, selectivity and asymmetry in the conformation of the power phenomenon. Analysis of Chilean society. Networks and Heterogeneous Media, 2015, 10(1): 167-194. doi: 10.3934/nhm.2015.10.167
    [6] Marina Dolfin, Mirosław Lachowicz . Modeling opinion dynamics: How the network enhances consensus. Networks and Heterogeneous Media, 2015, 10(4): 877-896. doi: 10.3934/nhm.2015.10.877
    [7] Emiliano Cristiani, Fabio S. Priuli . A destination-preserving model for simulating Wardrop equilibria in traffic flow on networks. Networks and Heterogeneous Media, 2015, 10(4): 857-876. doi: 10.3934/nhm.2015.10.857
    [8] Rosa M. Benito, Regino Criado, Juan C. Losada, Miguel Romance . Preface: "New trends, models and applications in complex and multiplex networks". Networks and Heterogeneous Media, 2015, 10(1): i-iii. doi: 10.3934/nhm.2015.10.1i
    [9] Martino Bardi . Explicit solutions of some linear-quadratic mean field games. Networks and Heterogeneous Media, 2012, 7(2): 243-261. doi: 10.3934/nhm.2012.7.243
    [10] Pierre Cardaliaguet, Jean-Michel Lasry, Pierre-Louis Lions, Alessio Porretta . Long time average of mean field games. Networks and Heterogeneous Media, 2012, 7(2): 279-301. doi: 10.3934/nhm.2012.7.279
  • Typhoon forecasting has always been a vital function of the meteorological department. Accurate typhoon forecasts can provide a priori information for the relevant meteorological departments and help make more scientific decisions to reduce the losses caused by typhoons. However, current mainstream typhoon forecast methods are very challenging and expensive due to the complexity of typhoon motion and the scarcity of ocean observation stations. In this paper, we propose a typhoon track prediction model, DeepTyphoon, which integrates convolutional neural networks and long short-term memory (LSTM). To establish the relationship between the satellite image and the typhoon center, we mark the typhoon center on the satellite image. Then, we use hybrid dilated convolution to extract the cloud features of the typhoon from satellite images and use LSTM to predict these features. Finally, we detect the location of the typhoon according to the predictive markers in the output image. Experiments are conducted using 13, 400 satellite images of time series of the Northwest Pacific from 1980 to 2020 and 8420 satellite images of time series of the Southwest Pacific released by the Japan Meteorological Agency. From the experimentation, the mean average error of the 6-hour typhoon prediction result is 64.17 km, which shows that the DeepTyphoon prediction model significantly outperforms existing deep learning approaches. It achieves successful typhoon track prediction based on satellite images.



    In any country, there are people who are going through a difficult life situation. This difficulty can have been caused by an epidemic [1], a natural disaster [2], a war [3]; however, one of the primary reasons for a hard life is poverty [4,5]. The scarcity of financial resources to get safe shelter, clean water, nutritious food, adequate sanitation, and medical relief can result in a life-threatening condition. In 2020, about 800 million people in the world faced hunger [6]. The recent COVID-19 pandemic pushed tens of millions of people into extreme poverty, by disrupting their livelihoods [7].

    Fortunately, part of the poor population is supported by charitable foundations (such as non-governmental organizations and religious institutions) and philanthropists [8,9]. In fact, people help each other to overcome a critical moment by donating money, clothes, food, expertise, time. As a consequence, the resource differences among wealthy and underprivileged persons are decreased, at least slightly. In evolutionary terms, the practice of charity reduces the fitness differences of unrelated individuals [10,11].

    Game theory has been a framework commonly used to study the emergence and persistence of collective cooperative behavior [12,13,14,15,16,17] (for reviews, see [18,19,20]). This theory assumes that the players are rational; that is, they play to maximize their own payoffs [21]. Therefore, they must be selfish and self-interested. A paradigmatic game usually employed in these studies is the prisoner's dilemma [22,23,24,25]. In its original single-trial version, mutual cooperation yields the highest collective payoff, but the highest individual payoff can only be obtained by defecting, which is considered the rational choice for both players [18,19,20,21,22,23,24,25]. Mutual defection, however, yields the second worst payoff for both players. Since real people tend to be more cooperative than theoretically predicted [12,26,27,28,29,30], modifications have been proposed in order to promote cooperation in the prisoner's dilemma. In the iterated version of this game, cooperation can be established via reciprocity [10,18,19,20,31]. In the spatial version, cooperation can be sustained by the topological structure of the social contact network [10,19,20,32]. Other variations take into account punishment for defection [33], memory of previous encounters [34], learning mechanisms [35], scale-free complex networks [36].

    Donating is a particular type of cooperative behavior found in humans and other animals, such as bats and chimpanzees [10]. Studies on this behavior have also interested the business world. For instance, there are analyzes addressing the link between corporate financial performance and corporate philanthropy [37,38,39]. From the company perspective, the aims of the strategic philanthropy are concomitantly to generate social benefits and to promote the brand reputation, which can be converted into financial returns. Thus, altruism and capitalism can find win-win solutions.

    There are several experimental [40,41,42] and theoretical [43,44,45] studies on charitable activities. For instance, there are studies about the social network supporting charity in New York City [46], about the optimal location of charity boxes for maximizing public donations [47], and about gamification of donation-based crowdfunding campaigns to improve the engagement of users [48]. Recently, investigations on charity were motivated by the COVID-19 pandemic [49,50,51]. A usual theoretical approach is to analyze a donor's utility function looking for an optimal solution [43,44,45]. In this article, the donation behavior is modeled by using game theory.

    The prisoner's dilemma is not convenient for modeling altruistic interactions, even when such interactions are mainly motivated by self-interest. Hence, the spread of charitable practice is here investigated by using a new game-theoretical model. In the proposed model, the players are potential donors with two possible donation behaviors and the payoffs express the (tangible) psychological benefits of making (or not) a donation. This approach was not found in the literature. Our main goal is to determine how the percentage of donors is affected by the values of the model parameters.

    The remainder of this article is organized as follows. In Section 2, the proposed model is described. In addition, the differences between this model and the prisoner's dilemma are stressed. In Section 3, the results obtained from computer simulations are presented. In Section 4, the practical implications of these results are discussed. These implications are especially important in times of crisis.

    Consider a square lattice composed of n×n cells, in which each cell represents an individual. To prevent edge effects, the left and right edges are connected and the top and bottom edges are connected too; thus, a three-dimensional torus is formed from a two-dimensional lattice. Consider also that each individual contacts the eight surrounding neighbors (which is known as Moore neighborhood of unitary radius [52]). Since all N=n2 individuals in this social network are equivalent from a geographical point of view, their spatial locations are irrelevant. Despite its simplicity, this kind of connectivity pattern has been used in several theoretical works on epidemiology [53], game theory [54], and opinion dynamics [55], for instance.

    Assume that each individual is a potential donor. In this society, there are two types of donors, denoted by I (from intrinsic) and E (from extrinsic). The percentage of I-individuals is p; hence, the percentage of E-individuals is 1p. At each time step t, each potential benefactor can give (donate) or keep (not donate). These actions are denoted by G and K, respectively. In this study, the charitable cause (for instance, children's support, food distribution, homeless services) is not specified. The donated amount is also not stipulated; what matters here is only the practice of charity. The payoffs received by the individuals reflect their personal satisfaction as a consequence of their acts. In addition, the received payoffs are neither accumulated nor shared from one time step to the next.

    In this model, I-individuals represent those who help others for pure altruism. They are intrinsically engaged in charitable activities in the sense that they are not influenced by the neighbors' actions. Their motivations come from internalized beliefs and values. Here, at each time step, each I-individual makes a donation with probability q. In other words, an I-individual plays G with probability q and K with probability 1q, regardless the payoffs received by this individual or by the neighbors in previous time steps. If the donation is made, the I-individual receives the payoff Y; if not, the received payoff is X. It is assumed that Y>X, since an I-individual is happier when a donation is completed.

    In this model, E-individuals make charitable donations to acquire social acceptance and prestige. They are extrinsically motivated in the sense that their actions are driven by peer pressure. Here, at each time step, each E-individual plays either G or K and becomes aware of the action chosen by each one of the eight neighbors. There are four possible combinations: (G,G), (G,K), (K,K), and (K,G). In each combination, the first letter denotes the action of this E-individual and the second letter denotes the action of a neighbor (which can be an E or an I-individual). Table 1 shows the payoffs of this charity game. As shown in this table, an E-individual receives R for (G,G), S for (G,K), P for (K,K), and T for (K,G). Since there are eight neighbors, each E-individual receives eight payoffs per time step. These eight payoffs are summed up. Then, at the time step t+1, each E-individual will play the same action of the neighbor who received the highest payoff in the time step t. This strategy is called copycat or "best takes all" rule. It was already used in other studies [32,56] and it is suitable in this model for E-individuals. In fact, as they are motivated by social pressure, it is natural to consider that they will imitate the better performing neighbor in the hope of increasing their prestige or avoiding social sanctions. The purpose of imitating successful behaviors is also supported by the economics literature [57,58].

    Table 1.  The payoff matrix of the charity game. In this game, G means give (donate) and K keep (not donate). An E-individual can receive the payoff P, R, S, or T. The elements of this matrix are composed by a pair of numbers. The first number is the payoff of the E-individual; the second number, the payoff of the neighbor, which can be an E-individual or an I-individual. Recall that I-individuals receive Y if they play G and X if they play K.
    E or I-individual
    G K
    E-individual G R,R or Y S,T or X
    K T,S or Y P,P or X

     | Show Table
    DownLoad: CSV

    For E-individuals, a positive reputation is built by doing charitable acts; hence, it is reasonable to assume that S>R>P>T. Thus, the highest payoff S is obtained when this E-individual donates and the neighbor does not; therefore, the worst payoff T refers to the opposite situation. The second highest payoff R is received when the E-individual and the neighbor simultaneously donate. The second lowest payoff P is received when both simultaneously do not donate. For simplicity, assume that Y=R and X=T; thus, the satisfaction scores of both types of donors are related.

    The average number of donors at t=1 (the initial time step) is Npq, because only I-individuals play at t=1 (since E-individuals copy the most successful neighbor, they start to play at t=2, after I-individuals playing once). Therefore, p=0 or q=0 implies no donors for t2. In fact, if there is no I-individual (p=0) or all I-individuals never donate (q=0), then E-individuals will not donate too as time passes (recall that E-individuals are driven by peer pressure).

    In this manuscript, numerical simulations are performed by varying N (the population size), p (the proportion of the two types of benefactors), q (the probability of I-individuals playing G), and Y (the highest payoff of an I-individual, which is equal to the payoff of an E-individual when such an E-individual and the neighbor donate). Throughout each numerical simulation, the actions of all individuals are simultaneously updated. The results obtained in these simulations are presented in the next section.

    Before continuing, it is relevant to point out the differences between the charity game and the prisoner's dilemma. Observe that the inequality concerning the payoffs for E-individuals in the charity game shown in Table 1 is different from that found in the prisoner's dilemma [10,18,19,20,21,22,23,24,25,26,31,32,33,34,35,36,56]. In this dilemma, two individuals are arrested and accused of jointly committing a crime. The individuals are separately interrogated and each one can either remain silent or admit to committing the crime. The first action is usually denoted by C (from cooperation) and the second action by D (from defection). When the individual chooses C, the received payoff is either R or S, if the partner plays either C or D, respectively. When the individual chooses D, the received payoff is either T or P, if the partner plays either C or D, respectively. In this classic game, T>R>P>S (and 2R>T+S) [10,18,19,20,21,22,23,24,25,26,31,32,33,34,35,36,56]. Assume that the two suspects involved in the prisoner's dilemma can be replaced by two E-individuals in the charity game. In addition, assume that the actions C and D in the prisoner's dilemma respectively correspond to actions G and K in the charity game. Notice that, by comparing these two games side by side, they are not equivalent, because the highest and the lowest payoffs are switched; that is, for the proposed model S>R>P>T.

    Computer simulations were performed by taking X=T=1, P=0, and S=6. They were run for 200 time steps. The goal is to investigate how the number of donors (individuals that play G) is affected by the values of N, p, q, and Y=R. In these simulations, N=1024 or 4096, p[0,1], q[0,1], and Y[0.5,5.0].

    Figure 1 exhibits the time evolution of the percentage of donors δ for N=4096 (n=64), p=0.5, q=0.5, and Y=2.5. After a few time steps, this percentage fluctuates around 0.51±0.02. The average value of the percentage of donors, denoted by ˉδ, and the corresponding standard deviation, denoted by Δ, were computed by taking into account the last 150 time steps. Observe that, in the first time step, ˉδ=pq=0.25 (as mentioned in the previous section).

    Figure 1.  Time evolution of the percentage of donors δ for N=4096, X=1, P=0, S=6, p=0.5, q=0.5, and Y=2.5. In the long term, this percentage tends to 0.51±0.02.

    Figures 2 and 3 present the average percentage of donors ˉδ in function of p and Y for q=0.1 and q=0.9, respectively. In both figures, ˉδ is not significantly affected by Y; however, in Figure 2, ˉδ tends to decrease with p and, in Figureg 3, ˉδ tends to increase with p. In these plots, the standard deviation Δ (not shown in the Figures) decreases with p; for instance, in both plots, Δ=7% for p=0.1 and Δ=1% for p=0.9,

    Figure 2.  Average percentage of donors ˉδ in function of p and Y for q=0.1. This average percentage was calculated by considering the last 150 time steps of each simulation. The other parameter values are the same as those used in Figure 1.
    Figure 3.  Average percentage of donors ˉδ in function of p and Y for q=0.9. The other parameter values are the same as those used in Figure 1.

    Figure 4 presents ˉδ in function of q for Y for p=0.5. This plot confirms that ˉδ does not significantly vary with Y and it increases with q. Also, Δ is not affected by Y and q. In this Figure, Δ=2% for all values of ˉδ.

    Figure 4.  Average percentage of donors ˉδ in function of q and Y for p=0.5. The other parameter values are the same as those used in Figure 1.

    Figure 5 exhibits ˉδ in function of p for q for Y=2.5. In order to better understand this plot, two other figures were made: Figure 6 shows how ˉδ depends on p for Y=2.5 and q=0.1 (blue line), q=0.5 (green line), and q=0.9 (red line); and Figure 7 shows how ˉδ depends on q for Y=2.5 and p=0.1 (blue line), p=0.5 (green line), and p=0.9 (red line). Figures 5-7 confirm that ˉδ increases with q, but it can decrease or increase with p, depending on the value of q, as observed in Figures 2 and 3.

    Figure 5.  Average percentage of donors ˉδ in function of p and q for Y=2.5. The other parameter values are the same as those used in Figure 1.
    Figure 6.  Average percentage of donors ˉδ in function of p for Y=2.5 and q=0.1 (blue line), q=0.5 (green line), and q=0.9 (red line). The other parameter values are the same as those used in Figure 1.
    Figure 7.  Average percentage of donors ˉδ in function of q for Y=2.5 and p=0.1 (blue line), p=0.5 (green line), and p=0.9 (red line). The other parameter values are the same as those used in Figure 1.

    As already stated above, ˉδ does not vary with Y. For instance, by taking N=4096 and the other parameter values used in Figure 1, then ˉδ=0.51±0.02 for Y[0.5,5.0]. The simulations also show that values of ˉδ and Δ are not notably affected by changing the population size N. For instance, by taking N=1024 (instead of 4096) and the other parameter values used in Figure 1, the long-term percentage of donors is 0.52±0.02 (instead of 0.51±0.02).

    These numerical results suggest that ˉδ can be analytically written in terms of p and q as:

    ˉδ(p,q)pq+(1p)pqH(p,q) (3.1)

    In this expression, the first term is the average percentage of I-donors and the second term is an approximation for the average percentage of E-donors. Notice that the probability of an E-individual playing G is supposed to be given by pqH(p,q). Thus, this probability is null for p=0 or q=0, as mentioned in Section 2. The fitting function H(p,q) is here written as H(p,q)=Uu=0Vv=0αuvpuqv, in which U and V determine the order of this polynomial function and αuv are constants. The values of the constants αuv can be computed by using a least-squares method [59]. For instance, take U=2 and V=1. From the three plots shown in Figure 6, then H(p,q)=6051q136p+66pq+76p2. For p=0.5 and q=0.5, then H(0.5,0.5)=2 and, from Eq. (3.1), ˉδ(0.5,0.5)0.50, which is close to the average number of donors obtained in Figure 1.

    Common terms employed in studies on charitable attitudes are altruism, generosity, solidarity. Despite their differences, these terms are related to people making decisions that benefit strangers [10,28,60].

    In the model proposed here, there is a fraction p of I-individuals that make heartfelt charity and a fraction 1p of E-individuals that make self-interested charity. Also, I-individuals act according to a probabilistic rule and E-individuals act according to a deterministic rule. Hence, the higher the value of p, the lower the value of the standard deviation Δ.

    Figures 2, 3, 5, and 6 show that for high values of q (that is, q>0.5), ˉδ increases with p (that is, with the proportion of I-individuals); for low values of q (that is, q<0.5), ˉδ decreases with p. Therefore, despite E-individuals making charitable donations for promoting themselves, their presence can increase ˉδ if the probability q of spontaneous donation for I-individuals is low. Thus, "impure" altruism can be as beneficial as "pure" altruism. Individuals who donate only for self-promotion can motivate others to making contributions for the same reason and, because of that, the number of donors increases.

    Figures 4, 5, and 7 show that ˉδ increases with the probability q of I-individuals making donations. Thus, the pure altruism of I-individuals increases the number of donors, by stimulating E-individuals to engaging in charitable solicitations. Evidently, q=1 leads to the (optimal) scenario in which ˉδ is maximum (that is, ˉδ=1). In a real-world society, the value of q can be influenced by the current socioeconomic scenario (which impacts the financial situation of all individuals).

    Figures 24 reveal that ˉδ is not significantly affected by the value of Y. The rationale behind this result is the following. I-individuals receive Y for donating; however, they do not donate for obtaining this payoff; E-individuals receive Y for simultaneous donations (by considering a neighbor); however, this is not the highest payoff available for them. Therefore, the payoff Y is either unimportant (for I-individuals) or unappealing (for E-individuals).

    In short, despite E-individuals being motivated by self-interest, their behavior can positively affect the contributions to fundraising campaigns, especially when the chance of I-individuals making donations declines (which can occur, for instance, in times of economic crisis). By publicly acknowledging the donations of E-individuals, charitable foundations concurrently improve the reputation of these individuals and, more importantly, attract new donors.

    Observe that Eq. (3.1) can be used to estimate the average number of donors Nˉδ in this simulated society from the values of p and q. The validity of this equation (and of the model itself) can be tested by using actual data. These data should also contain realistic values of the psychological payoffs, in order to derive an analytical expression for ˉδ(p,q) suitable for a real-world society. Such an expression could be used by charitable foundations to plan their activities.

    The model proposed here is based on the assumptions that I-individuals randomize their choices and E-individuals inevitably copy the best neighbor. Variations of this model could consider that I-individuals follow a deterministic donation schedule and/or E-individuals are ruled by a probabilistic imitation dynamics. In fact, different profiles of donors can be taken into account in future studies.

    The data used to support the findings of this study are available from the first author upon request.

    LHAM is partially supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) under the grant #302946/2022-5. This study was financed in part by Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) - finance code 001.

    The authors declare that there are no conflicts of interest regarding the publication of this article.



    [1] L. S. Chen, Y. H. Ding, Introduction to Typhoons in The Western Pacific, Science Press, 1979.
    [2] Y. Sun, Z. Zhong, T. Li, L. Yi, Y. J. Hu, H. C. Wan, et al., Impact of ocean warming on tropical cyclone size and its destructiveness, Sci. Rep., 7 (2017), 1–10. https://doi.org/10.1038/s41598-017-08533-6 doi: 10.1038/s41598-016-0028-x
    [3] J. P. Kossin, A global slowdown of tropical–cyclone translation speed, Nature, 558 (2018), 104–107. https://doi.org/10.1038/s41586-018-0158-3 doi: 10.1038/s41586-018-0158-3
    [4] H. Kim, U. Yun, B. Vo, J. C. W. Lin, W. Pedrycz, Periodicity-oriented data analytics on time-series data for intelligence system, IEEE Syst. J., 15 (2021), 4958–4969. https://doi.org/10.1109/JSYST.2020.3022640 doi: 10.1109/JSYST.2020.3022640
    [5] T. T. Tang, Q. L. Li, G. X. Li, Y. L. Peng, Research on statistical model of typhoon intensity prediction based on meteorological big data, Integr. Technol., 2 (2016), 73–84. https://doi.org/10.12146/j.issn.2095-3135.201602006 doi: 10.12146/j.issn.2095-3135.201602006
    [6] T. Song, Y. Li, F. Meng, P. F. Xie, D. Y. Xu, A novel deep learning model by BiGRU with attention mechanism for tropical cyclone track prediction in the northwest pacific, J. Appl. Meteorol. Climatol., 61 (2022), 3–12. https://doi.org/10.1175/JAMC-D-20-0291.1 doi: 10.1175/JAMC-D-20-0291.1
    [7] J. Lian, P. P. Dong, Y. P. Zhang, J. G. Pan, K. H. Liu, A novel data–driven tropical cyclone track prediction model based on CNN and GRU with multi-dimensional feature selection, IEEE Access, 8 (2020), 97114–97128. https://doi.org/10.1109/ACCESS.2020.2992083 doi: 10.1109/ACCESS.2020.2992083
    [8] S. Gao, P. Zhao, B. Pan, Y. R. Li, M. Zhou, J. L. Xu, et al., A nowcasting model for the prediction of typhoon tracks based on a long short term memory neural network, Acta Oceanol. Sin., 37 (2018), 8–12. https://doi.org/10.1007/s13131-018-1219-z doi: 10.1007/s13131-018-1219-z
    [9] R. S. T. Lee, J. N. K. Liu, Tropical cyclone identification and tracking system using integrated neural oscillatory elastic graph matching and hybrid RBF network track mining techniques, IEEE Trans. Neural Network, 11 (2000), 680–689. https://doi.org/10.1109/72.846739 doi: 10.1109/72.846739
    [10] R. Kovordányi, C. Roy, Cyclone track forecasting based on satellite images using artificial neural networks, ISPRS J. Photogramm., 64 (2009), 513–521. https://doi.org/10.1016/j.isprsjprs.2009.03.002 doi: 10.1016/j.isprsjprs.2009.03.002
    [11] M. Rüttgers, S. Lee, S. Jeon, D. You, Prediction of a typhoon track using a generative adversarial network and satellite images, Sci. Rep., 9 (2019), 1–15. https://doi.org/10.1038/s41598-019-42339-y doi: 10.1038/s41598-018-37186-2
    [12] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, et al., Generative adversarial networks, Commun. ACM, 63 (2020), 139–144. https://doi.org/10.1145/3422622 doi: 10.1145/3422622
    [13] K. Q. Liu, J. Y. Yuan, W. Z. Chen, Y. P. Chen, Verification of typhoon storm surge forecast model based on adcirc hydrodynamic model, Water Resour. Hydropower Express, 40 (2019), 27–34. https://doi.org/10.15974/j.cnki.slsdkb.2019.04.005 doi: 10.15974/j.cnki.slsdkb.2019.04.005
    [14] L. G. Wu, Z. P. Ni, J. J. Duan, H. J. Zong, Sudden tropical cyclone track changes over the western north pacific: A composite study, Mon. Weather Rev., 141 (2013), 2597–2610. https://doi.org/10.1175/MWR-D-12-00224.1 doi: 10.1175/MWR-D-12-00224.1
    [15] Z. Xu, J. Du, J. J. Wang, C. X. Jiang, Y. Ren, Satellite image prediction relying on GAN and LSTM neural networks, in IEEE International Conference on Communications, (2019), 1–6. https://doi.org/10.1109/ICC.2019.8761462
    [16] S. Kim, J. S. Kang, M. H. Lee, S. Song, Deeptc: Convlstm network for trajectory prediction of tropical cyclone using spatiotemporal atmospheric simulation data, in Conference and Workshop on Neural Information Processing Systems (NIPS), (2018). https://openreview.net/pdf?id=HJlCVoPAF7
    [17] C. Tan, Tclnet: Learning to locate typhoon center using deep neural network, in International Geoscience and Remote Sensing Symposium (IGARSS), (2021), 4600–4603. https://doi.org/10.1109/IGARSS47720.2021.9554524
    [18] R. Pradhan, R. S. Aygun, M. Maskey, R. Ramachandran, D. J. Cecil, Tropical cyclone intensity estimation using a deep convolutional neural network, IEEE Trans. Image Process., 27 (2018), 692–702. https://doi.org/10.1109/TIP.2017.2766358 doi: 10.1109/TIP.2017.2766358
    [19] S. Woo, J. Park, J. Y. Lee, I. S. Kweon, CBAM: Convolutional block attention module, in Proceedings of the European conference on computer vision (ECCV), (2018), 3–19. https://doi.org/10.1007/978-3-030-01234-2-1
    [20] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput., 9 (1997), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735 doi: 10.1162/neco.1997.9.8.1735
    [21] A. Graves, Supervised sequence labelling with recurrent neural networks, in Studies in Computational Intelligence, 385 (2008). https://doi.org/10.1007/978-3-642-24797-2
    [22] F. Yu, V. Koltun, T. Funkhouser, Dilated residual networks, in Proceedings-30th IEEE Conference on Computer Vision and Pattern Recognition, (2017), 636–644. https://doi.org/10.1109/CVPR.2017.75
    [23] P. Q. Wang, P. F. Chen, Y. Yuan, D. Liu, Z. H. Huang, X. D. Hou, et al., Understanding Convolution for Semantic Segmentation, in Proceedings-2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018, (2018), 1451–1460. https://doi.org/10.1109/WACV.2018.00163
    [24] J. B. Yu, The Research on the Recognition of the Classification of Cloud, Typhoon Segmentation and Locating of the Typhoon's Centre Based on the Meteorological Satellite, Master's thesis, Wuhan University of Technology, 2008. https://doi.org/10.7666/d.y1365665
    [25] H. Mahmoud, N. Akkari, Shortest path calculation: A comparative study for location-based recommender system, in Proceedings-2016 World Symposium on Computer Applications and Research, (2016), 1–5. https://doi.org/10.1109/WSCAR.2016.16
    [26] S. Giffard-Roisin, M. Yang, G. Charpiat, C. K. Bonfanti, B. Kégland, C. Monteleoni, Tropical cyclone track forecasting using fused deep learning from aligned reanalysis data, Front. Big Data, 3 (2020), 1. https://doi.org/10.3389/fdata.2020.00001 doi: 10.3389/fdata.2020.00001
    [27] S. Alemany, J. Beltran, A. Perez, S. Ganzfried, Predicting hurricane trajectories using a recurrent neural network, in Proceedings of the AAAI Conference on Artificial Intelligence, (2019), 468–475. https://doi.org/10.1609/aaai.v33i01.3301468
    [28] J. Lian, P. P. Dong, Y. P. Zhang, J. G. Pan, A novel deep learning approach for tropical cyclone track prediction based on auto–encoder and gated recurrent unit networks, Appl. Sci., 10 (2020), 3965. https://doi.org/10.3390/app10113965 doi: 10.3390/app10113965
    [29] S. F. Tekin, F. Ilhan, Cs559 project report forecasting of tropical cyclone trajectories with deep learning, preprint, arXiv: 1910.10566v2. https://doi.org/10.48550/arXiv.1910.10566
    [30] W. Qin, J. Tang, S. Y. Lao, Deepfr: A trajectory prediction model based on deep feature representation, Inf. Sci., 604 (2022), 226–248. https://doi.org/10.1016/j.ins.2022.05.019 doi: 10.1016/j.ins.2022.05.019
    [31] W. Qin, J. Tang, C. Lu, S. Lao, A typhoon trajectory prediction model based on multimodal and multitask learning, Appl. Soft Comput., 122 (2022), 108804. https://doi.org/10.1016/j.asoc.2022.108804 doi: 10.1016/j.asoc.2022.108804
  • This article has been cited by:

    1. D. Y. Charcon, L. H. A. Monteiro, On Playing with Emotion: A Spatial Evolutionary Variation of the Ultimatum Game, 2024, 26, 1099-4300, 204, 10.3390/e26030204
  • Reader Comments
  • © 2022 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(4453) PDF downloads(529) Cited by(6)

Figures and Tables

Figures(10)  /  Tables(9)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog