AN ARTIFICIAL INTELLIGENCE APPROACH TO PREDICTION OF EXTREME EVENTS: THE CASE OF STORMS IN WESTERN FRANCE

Storms represent an increased source of risk that affects human life, property, and the environment. Prediction of these events, however, is challenging due to their low frequency of occurrence. This paper proposed an artificial intelligence approach to address this challenge and predict storm characteristics and occurrence using a gated recurrent unit (GRU) neural network and a support vector machine (SVM). Historical weather and marine measurements collected from buoy data, as well as a database of storms containing all the extreme events that occurred in Brittany and Pays de la Loire regions, Western France, since 1996, were used. Firstly, GRU was used to predict the characteristics of storms (wind speed, pressure, humidity, temperature, and wave height). Then, SVM was introduced to identify storm-specific patterns and predict storm occurrence. The approach adopted leads to the prediction of storms and their characteristics, which could be used widely to reduce the awful consequences of these natural disasters by taking preventive measures.


INTRODUCTION
Extreme weather events, particularly storms, are one of the main natural hazards affecting human life, infrastructure, and the environment (Brunkard et al., 2008;Genovese & Przyluski, 2013;Jahn, 2015;Stephenson, 2008;Usbeck et al., 2010). It has been estimated that, on average, 60,000 people die each year due to extreme weather conditions (Franzke & Torelló i Sentelles, 2020). The economic losses from storms have also risen in most countries (Dorland et al., 1999;Kates, 1980;Leckebusch et al., 2007). However, storms are also important, as rain-bearing structures for large parts of the world; the case of West African parts that get 90% of their rainfall from thunderstorms (Kamara, 1986).
Due to the increased risk of these severe events (Haarsma, 2021;Kron et al., 2019), they have attracted significant attention in the last few decades (Ren et al., 2018). Predicting storms is a challenging task, and traditional prediction methods are mainly based on statistical and numerical models, leading to problems of high computational complexity and low prediction accuracy (Goerss, 2000). But machine and deep learning algorithms can rapidly analyze large volumes of data and serve as a costeffective alternative for predicting upcoming events (Sun et al., 2020), such as impending hurricane trajectories and storms (Ghosh & Krishnamurti, 2018).
Recently, the application of machine learning and deep learning models dramatically increased in predicting natural hazards and disasters. For instance, flood prediction using machine learning models (Mosavi et al., 2018) or landslide susceptibility prediction using deep learning (Dao et al., 2020). The use of machine learning and deep learning models becomes more and more popular for the prediction of extreme weather events; deep learning models are currently used to forecast severe convective weather over China (Zhou et al., 2019), for example, and the prediction results prove that machine learning and deep learning solutions outperform traditional methods (Lu et al., 2020;Zhang et al., 2019). (Sun et al., 2020) have also provided an overview of the current applications of these models in disaster management.
Gated recurrent unit (GRU) networks, which was developed as an improved version of recurrent neural networks (RNNs) to solve the problems of vanishing and exploding gradients in standard RNNs when learning long-term dependencies, is among the most advanced deep learning algorithms (Cho et al., 2014). The GRU model has outperformed other benchmark models and has shown to be effective for a wide range of issues, such as wave height prediction during tropical cyclones (Meng et al., 2021), wind power forecasting (Ding et al., 2019), and tropical cyclone track prediction (Lian et al., 2020).
In addition to GRU, support vector machines (SVM) is another robust and efficient machine learning algorithm widely used for time series classification and regression problems (Kampouraki et al., 2009;Sapankevych & Sankar, 2009). This algorithm has outperformed other methods in a wide variety of applications, such as extreme rainfall event prediction (Nayak & Ghosh, 2013), convective storms classification (Jergensen et al., 2020), flood prediction (Mosavi et al., 2018), and tornado Prediction (Trafalis et al., 2004). This paper proposes a new approach for predicting the occurrence and characteristics of storms using buoy data and a storm database containing all storm events that occurred in Brittany and Pays de la Loire regions, Western France, since 1996. The GRU model was introduced to predict storm characteristics (wind speed, pressure, humidity, temperature, and wave height). The SVM model attempted to identify stormspecific patterns and applied an SVM-based classifier for storm occurrence prediction. The remainder of this paper is organized as follows. Section 2 covers the study area, the data used, the architectures of the proposed models, and the methodology adopted. Section 3 presents the results of the experiment. Finally, Section 4 discusses the conclusions and summarizes the paper.

Study Area
This study focuses on a portion of The French Atlantic coasts. It comprises two regions: Brittany and Pays de la Loire with a total area of 66000 km 2 (Figure 1).

Figure 1. Location of the study area.
These two regions have known several storms leading to heavy loss of life and devastating infrastructure damage. Storms Lothar and Martin, for example, which hit the study area from December 26 to 28, 1999, caused the death of 92 people (Dedieu, 2010), in addition to the destructive impact on infrastructure, buildings and the environment (Sacré, 2002).
Due to the extensive damage and the risk represented by these extreme events, several works of literature have investigated the dynamics and paths of past storms affecting the study area; (Athimon & Maanan, 2018) studied storm impacts and vulnerability of past societies based on more than 19691 French historical documents, while other studies attempted to detect damaging storm events in the last 300 and 1000 years by coupling historical documents and sedimentological data (Pouzet et al., 2019;Pouzet & Maanan, 2020). These studies have identified two main categories of storms that affect the study area. The first type of storm has a west-to-east trajectory, and the second type has an SW-NE trajectory (Pouzet & Maanan, 2020).

Data Used
The data used is a combination of a storm database and buoy data. The storm database was obtained from the website Meteo France (http://tempetes.meteo.fr/spip.php?rubrique6), and it represents all the storm events that have occurred in the study area over the past 25 years.
The final time series data used to predict storm occurrence and characteristics represent daily records of wind speed, pressure, humidity, temperature, and wave height, in addition to the exact days that have known the occurrence of a storm.

Gated Recurrent Unit (GRU) neural network:
Recurrent neural networks (RNNs) are a powerful deep learning model known for their effectiveness in time series prediction (Elman, 1990). However, RNN faces problems when learning long-term dependencies due to the vanishing and exploding gradient problems (Bengio et al., 1994). To overcome these issues, Long Short-Term Memory (LSTM) was proposed in 1997 as an improved version of RNN (Hochreiter & Schmidhuber, 1997). Moreover, a gated recurrent unit (GRU) neural network was introduced by (Cho et al., 2014) to simplify the training parameters.
The architecture of GRU implements a restructuring gate mechanism while retaining the advantages of LSTM. The gate mechanism consists of an update gate and a reset gate. The update gate controls how much information from a previous state is brought into the current state. The more information brought into the current state, the higher the value of the update gate. It plays the role of input gate and forget gate in LSTM. On the other hand, the reset gate is used to control how much information from a previous state is ignored. In contrast to the update gate, the lower the reset gate value, the more information is ignored (E et al., 2019).
The processes inside each GRU cell are designed as follows (E et al., 2019): First, the reset gate and update gate are established by the inputs of the current input and the hidden state at previous time −1 . The outputs of the update gate and the reset gate are given as follows: where and = are weight coefficient matrices. = denotes bias vector. = is the logistic sigmoid function. helps to capture the short-term dependencies in temporal sequence, and is able to memorize arbitrary-length information of the input . Second, the current candidate hidden state ̃ is formulated as: where * = denotes the scalar product of two vectors. Finally, applying the current candidate hidden state ̃ and the previous hidden state ̃− 1 to implement the following linear combination that satisfies the sum of weighted coefficient be of a value 1.
where = is the output of the current hidden state.

Support Vector Machines
The SVM algorithm has performed very well on a wide range of problems and has become one of the most powerful methods in machine learning for both classification and regression (Kampouraki et al., 2009;Sapankevych & Sankar, 2009). The reason for using SVM in our approach is that this algorithm is less prevalent in the field of meteorology than other machine The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVI-4/W3-2021 Joint International Conference Geospatial Asia-Europe 2021 and GeoAdvances 2021, 5-6 October 2021, online learning methods, but it has been successfully used to predict tornadoes (Adrianto et al., 2009;Trafalis et al., 2003) and temperature (Radhika & Shashi, 2009).
Linear SVM was introduced by (VAPNIK, 1963) for binary classification; it separates data into two classes by constructing a hyperplane in the predictor space that best separates the two classes (Jergensen et al., 2020).
During training, the linear SVM learns to maximize the margin or the average Euclidean distance between the hyperplane and the correctly classified samples (Jergensen et al., 2020).
The majority of real-world data are not linearly separable. Consequently, nonlinear kernels (Cortes & Vapnik, 1995;Vapnik, 1995) are frequently used to transform the predictor space implicitly.
In this study, the linear kernel, which is defined in equation 5, was used: where and = are predictor vectors for two examples, both of length M.
= is the number of predictors. = is a hyperparameter, which encourages the model to overfit when too small and underfit when too large.

Methodology Adopted
The methodology used in this study consists of two major parts ( Figure 2): The first one aims to predict the different characteristics of storms (wind speed, pressure, humidity, temperature, and wave height), which represents a univariate time-series prediction, and the second part focuses on predicting the occurrence of storms based on their characteristics. To achieve these goals; the following steps were taken: First, data collection: a storm database containing all storm events that have occurred in the study area over the last 25 years was used, in conjunction with buoy data that represents hourly records of wind speed, pressure, humidity, temperature, and wave height. Daily values for each characteristic were calculated from these records and then combined with the storm database to create a final timeseries data that represents a day-by-day record of weather and marine data, in addition to the exact days that have known the occurrence of a storm. The final time series data created can be used to predict storm occurrence and characteristics.
Then, data pre-processing, which aims to prepare the data for the implementation of the algorithms. This step involves dealing with missing values and then scaling the features to normalize all independent variables. Next, the data was divided into training and testing sets; 80% of data was used for the training, and the last 20% was used to test the effectiveness of the developed models.
The third step is data processing. Five GRU models were developed to predict storm characteristics (wind speed, pressure, humidity, temperature, and wave height). Each model was specified for the prediction of a particular feature (a univariate time series prediction). Then, an SVM-based classifier was developed and used to predict storm occurrence by identifying and extracting storm-specific patterns.
Finally, performance evaluation is the last step of the study. The GRU models used to predict storm characteristics were evaluated using three evaluation metrics: mean absolute error, mean square error, and root mean square error, and the SVM classifier used to predict storm occurrence was evaluated using the confusion matrix and ROC (receiver operating characteristics) curve.

EXPERIMENTAL RESULTS
In this study, GRU and SVM were introduced to predict storm occurrence and characteristics. The GRU model has first used for predicting the different characteristics of storms, and then an SVM classifier was used to predict storm occurrence based on their characteristics. The results of the developed GRU and SVM models are presented separately below.

Prediction of Storm Characteristics
Brittany buoy data, which contains historical records of wind speed, pressure, humidity, temperature, and wave height from 1 January 1996 to 31 May 2021, was used to predict storm characteristics. Since the task is a univariate time series prediction, the values for each storm characteristic should be predicted separately. Therefore, a unique model was created for each of them.
Taking wind speed prediction as an example, a GRU model was created based on the univariate dataset of past wind speed, and all other characteristics were predicted in the same manner. Finally, five GRU models were developed, each designed to predict a particular characteristic (wind speed, pressure, humidity, temperature, and wave height).
The data was split into training and testing sets, with a ratio of 80 to 20. 80% was used to train the model and find the best hyperparameters, then the last 20% of data was used to test the robustness of the final model developed. The prediction results for other characteristics (pressure, humidity, temperature, and wave height) will be presented later in this section. Figure 3 depicts the GRU prediction results. A light blue curve represents the actual wind speed, while a light brown curve represents the predicted wind speed. As can be observed, the GRU model was successful in following the pattern and accurately predicting the general variability of wind speed. The GRU model doesn't yield the exact values of actual storm wind speeds. However, the results obtained are still promising as the model was able to predict higher wind speed values than the usual values on other days, which can help indicate the occurrence of a storm. For example, the wind speed of storm Zeus, which hit the study area on 6 March 2017, with a value of 19.72 m/s, was predicted by the GRU model to be 13.07 m/s.
where = is the number of samples. = represents the real value. = is the predicted value.
The evaluation of error metrics shows that the GRU model has few prediction errors. It has generated MAE, MSE, and RMSE values of 1.829, 5.309, and 2.304, respectively.
The prediction results for other storm characteristics (wave height, temperature, pressure, and humidity) are shown in Figure  4. The GRU model has successfully predicted the general variability and all the characteristics of storms, as shown in the figure. Therefore, it appears that the GRU model can be used as an effective tool for the prediction of storm characteristics. Figure 4. Prediction of storm characteristics; wave height, temperature, pressure, and humidity using the GRU model.

Prediction of storms
In this second part, a binary SVM classifier was developed to predict storm occurrence based on their characteristics. To achieve this goal, multivariate time series data was created by combining the storm database and the buoy data to predict storm characteristics. The multivariate time series data represent historical measurements of wind speed, temperature, pressure, humidity, and wave height from 1 January 1996 to 31 May 2021, in addition to the target variable to be predicted, which is in the range of 0 and 1, where 1 represents the occurrence of a storm and 0 indicates that there is no storm. This time series data was divided into training and testing at the ratio of 80 to 20%; 80% for training and 20% for testing the model's performance. In this study, there were two classes; storms and no storms. The storm class represents the positive class with 68 samples, and the no storm class represents the negative class with 9215 samples. The prediction outcomes for the SVM model are shown in Figure  5. The predicted storms are represented with points in light brown color, while the actual storms are represented with points in light blue color.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVI-4/W3-2021 Joint International Conference Geospatial Asia-Europe 2021 and GeoAdvances 2021, 5-6 October 2021, online To assess the performance of the SVM classifier developed, five classification metrics were used: specificity, sensitivity, falsepositive rate (FPR), false-negative rate (FNR), and the area under the receiver operating characteristic curve (AUC) metric. These evaluation metrics are not sensitive to changes in data distribution and can be used effectively with imbalanced data, as is the case in this study (Tharwat, 2018).
These assessment metrics can be calculated using the following equations: where = true positives where the model predicted samples correctly as positives. In this case, the storms were classified as storms.
= true negatives where the model predicted samples correctly as negatives, no storms predicted as no storms.
= false positives where the model mispredicted samples as positives, no storms predicted as storms.
= false negatives where the model mispredicted samples as negatives, storms predicted as no storms.  The ROC curve is another evaluation method used in this study to assess the capability of the binary SVM classifier. It is a twodimensional graph in which the FPR (False positive rate) represents the x-axis and the TPR (True positive rate) represents the y-axis. It can be generated by changing the confidence score threshold (Tharwat, 2018). The AUC value is a metric extracted from the ROC curve and used to evaluate the model's performance. The AUC metric ranges from 0 to 1, and a bigger value indicates a better model. Finally, the SVM binary classifier has proven its efficiency for predicting storm occurrence since it has an AUC value of 1, as shown in Figure 7.

SUMMARY AND DISCUSSION
Artificial intelligence, specifically machine learning and deep learning methods, has recently exploded in popularity in many fields, including meteorology (McGovern et al., 2019). The remarkable results obtained by these methods have prompted experts to introduce this concept into their various research directions in meteorology. For example, a deep learning approach is currently being used to forecast severe convective weather in China (Zhou et al., 2019).
The high performance of machine learning and deep learning methods in meteorology is also supported by findings from (Lu The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLVI-4/W3-2021Joint International Conference Geospatial Asia-Europe 2021and GeoAdvances 2021, 5-6 October 2021, online et al., 2020, who showed a Mask region-based convolutional neural network (Mask R-CNN) model accurately identified extratropical cyclones with strong winds along the Northern Hemisphere Ocean coasts. (Lu et al., 2020) also found that the number of identified cyclones increased by 8.29% compared to the results obtained by the traditional method. Consequently, we also examine the efficiency of an artificial intelligence approach in the context of storm prediction. Therefore, this study targeted the building of time series prediction models to predict storm occurrence and characteristics, which could be used globally to mitigate the disastrous consequences of these extreme events by taking the necessary precautions.
The paper's main contribution is applying a new artificial intelligence approach using two of the most advanced machine learning and deep learning algorithms, Support Vector Machine (SVM) and Gated Recurrent Unit (GRU) Neural Network.
Experiments were carried out using a database of storm events in the western regions of France over the past 25 years and buoy data representing historical measurements of wave height, wind speed, temperature, pressure, and humidity since 1996.
As a first step, the GRU model was used to predict the different characteristics of storms (wind speed, pressure, humidity, temperature, and wave height). Each storm characteristic was predicted separately, and a GRU model was created for each of them based on past data. (Khosravi et al., 2018) also developed a time-series prediction model using machine learning algorithms to predict wind speed over Brazil based on past measured wind speed data values. The results demonstrated that the adopted method could successfully predict wind speed with high accuracy.
Wind speed prediction is one of the most critical aspects of storm events in recent years. (Yang et al., 2017(Yang et al., , 2019 have recently proposed the Bayesian linear regression to predict wind speed for storms that impacted the northeast United States based on a database composed of historical storms. The lower values of root mean square error indicate the models' high performance for storm wind speed prediction. Our analyses extend these recent studies by considering the prediction of other storm features in addition to wind speed. The method used will result in the prediction of different storms characteristics (wind speed, wave height, temperature, pressure, and humidity) and the exact days of their occurrence, which can help reduce and anticipate the impacts of upcoming events. Several evaluation criteria are applied to evaluate the performance of the developed GRU models in terms of prediction accuracy. Mean absolute error, mean square error, and root mean square error between actual data and predicted ones were used. The developed models were found to make few prediction errors, which proves the efficiency of these models for predicting all storms characteristics.
In the second step of this study, an SVM-based classifier was applied to identify storm-specific patterns and predict storm occurrence. The same hypothesis was applied by (Nayak & Ghosh, 2013) to predict extreme rainfall events using a support vector machine to identify those specific patterns for extreme rainfall events and then apply an SVM-based classifier for extreme rainfall classification and prediction.
The SVM-based classifier developed in this study effectively predicted storm occurrence and achieved high performance in terms of all the evaluation metrics adopted (specificity, sensitivity, and AUC score).
The findings revealed that the models mentioned above (GRU and SVM) could serve as an effective tool for predicting storm occurrence and characteristics, as both models have performed very well and produced good generalization ability with unseen data. Although future studies will be needed to address the lack of skill in making a long-term prediction of storms characteristics, the GRU model can only generate short-term predictions. Also, the GRU model predicted lower values than the actual values of storms characteristics. Therefore, further investigations into which models other than GRU and SVM can improve the prediction are necessary.
Finally, considering the results obtained, our results appear promising overall. The adopted methodology can predict future storms characteristics and occurrence, which can help avoid the severe damage of these extreme events.