Czekanowsky Hypergraph-Based Deep Learning Classifier for Precision Cyclone Forecasting

Early prediction of cyclones helps reduce deaths and damage to properties worldwide. With the advancement in satellite imaging technology, obtaining atmospheric images and remotely sensed objects such as cyclones is possible using different modalities. Such images are handy for weather prediction, specifically for forecasting cyclonic storms. However, it has a few limitations for achieving higher prediction accuracy with minimal time. A novel technique, proposed by Czekanowsky Dice Hypergraphic Extended Kalman Momentum Filterization based Bivariate Correlative Deep Structure Learning classification (CDHEKMF-BCDSLC), is introduced to achieve better cyclone prediction accuracy. Initially, the multiple satellite images are gathered from cyclone datasets—the proposed technique comprises four processing steps: segmentation, preprocessing, feature extraction, and classification. At first, Czekanowsky dices the Intensity threshold-based Interval Hypergraph (CDIT-IH) model using a Segmentation process. It minimized to perform cyclone prediction time. After that, preprocessing is applied to a novel invariant extended Kalman momentum filter designed to improve image contrast. Next, Bivariate correlative spatiotemporal feature extraction is performed on each image pixel intensity to extract features over time and location. Finally, a multidimensional deep belief network classification model is applied for accurate cyclone prediction. A multidimensional deep belief network classification model is a machine learning technique that consists of multiple layers for learning the given input (i.e., Spatiotemporal features). This process increases the prediction accuracy. Experimental results reveal that the proposed technique noticeably predicts cyclone conditional variants using prediction accuracy, precision, recall, F-measure, and prediction time concerning the number of cyclone images. The Quantitative results show that the proposed technique achieves 6% better accuracy with 17% minimum prediction time, 5% improved precision and f- f-measure, and 4% recall when compared to the state-of-the-art methods.

SECTION I. Introduction

Cyclone intensity evaluation is a significant task in meteorological research and weather forecasting. However, intensity assessment is complex since it requires domain knowledge while extracting the Spatiotemporal features. Two issues related to intensity estimation are the unpredictability of results and the significant pre-processing of images. Hence, building an effective model for cyclone intensity estimation takes time and effort. This study [1] estimates the likelihood of rapid intensification and intensity of cyclones using a deep ensemble approach. The ensemble comprises 20 distinct deep-learning models. This research integrates human-defined environmental parameters and information extracted from satellite images.

Accurate trajectory prediction of tropical cyclones is crucial for effective disaster management [2]. This research employs three deep learning models—multilayer perceptron (MLP), long short-term memory (LSTM), and a hybrid MLP-LSTM—to predict cyclone trajectories in the North Atlantic Ocean. The MLP-LSTM model performs better than its counterparts, achieving prediction errors of 52.73 km, 20.65 km, and 19.54 km for the next three hours.

The prediction of tropical cyclone tracks is treated as a time series forecasting problem in [3]. This problem is addressed by employing a bidirectional gate recurrent unit (BiGRU) network with an attention mechanism to enhance the extraction and utilization of historical track data. Using best-track data from the Joint Typhoon Warning Center (JTWC) for the northwest Pacific from 1988 to 2017, the model’s performance was evaluated for predictions at 6, 12, 24, 48, and 72-hour intervals. Results indicate that the BiGRU with attention mechanism outperforms other advanced deep learning models, including RNN, LSTM, GRU, and BiGRU without attention.

Exploring new techniques to improve the prediction of tropical cyclone formation (TCF) is essential [4]. This study leverages the availability of large volumes of TCF data to demonstrate that convolutional neural networks can offer promising performance in predicting TCF. Popular architectures such as ResNet and UNet successfully process this extensive data, achieving superior performance.

A new tropical cyclone intensity classification and estimation model (TCICENet) has been introduced in [5] using satellite images collected from the northwest Pacific Ocean basin. This model no doubt increases the classification accuracy but fails to extract the temporal features effectively. A Tensor-Based Convolutional Neural Network (TCNN) has been introduced in [6] for intensity classification and regression based on wind speed estimation. However, this model needs to focus on designing efficient network architectures.

Deep convolutional neural network architecture has been introduced in [7] to estimate the hurricane’s cyclone intensity by performing complex feature extraction. This technique does not help reduce the time taken. Based on adaptive weight, tropical cyclone prediction was performed in [8]. This method increases the improved intensity forecasts, but the minimum time could not be achieved. A global ensemble prediction system was introduced in [9] to improve the quality of cyclone prediction. However, the system could have been more efficient in attaining higher prediction accuracy. NCMRWF Ensemble Prediction System (NEPS) was developed in [10] for cyclone prediction with minimum error. However, higher precision and recall rates were not obtained.

A recurrent neural network was introduced in [11] based on substantial past observation data for intensity prediction. However, the accuracy of intensity prediction still needs to be improved. A deep learning approach was introduced in [12] to find the tropical cyclones (TCs) and their precursors. However, the detection performance with minimum time could not be improved. A new data-driven deep learning approach was introduced in [13] to forecast the tropical cyclone tracks based on spatial and meteorological features. However, classifying different tropical cyclone intensity levels could have been more efficient.

A Hierarchical Generative Adversarial Network (HGAN) was introduced in [14] to predict the typhoon cloud. Though the network minimizes the error, higher precision and recall rate could not be achieved. The -Guided Cyclone Track Forecasting (EGCTF) approach was introduced in [15] for remote tropical cyclone tracking. This approach minimizes the low computational load, but the accuracy still needs to be improved. Two multifactor models, namely the Generalized Linear Model (GLM) and the long short-term memory (LSTM) model, were introduced in [16] for tropical cyclone intensity evaluation. However, high-correlation factors were not considered to improve the accuracy of cyclone intensity estimation.

A novel hybrid approach was introduced in [17] based on convolutional neural networks (CNNs) for cyclone intensity evaluation. While increasing the classification accuracy, the time for cyclone intensity evaluation remained the same. Multivariate linear regression was developed in [18] for cyclone prediction. However, the accuracy of cyclone prediction could not be improved. A novel deep learning algorithm was introduced in [19] to enhance the performance of cyclone detection. However, it could have accurately predicted the cyclone pathway more precisely and quickly. Using a multilayer perceptron model, an image processing-based technique was developed in [20] to estimate the cyclone intensity from satellite images. However, the different intensity levels could not be estimated accurately.

A relevance vector machine was introduced in [21] to construct the cyclone intensity estimation approach at every stage. However, higher accuracy of cyclone intensity estimation still needs to be achieved. A deep-learning-based tropical cyclone intensity estimation system was developed in [22] with less mean squared error. The system failed to evaluate the wind speed for tropical cyclones at a lower intensity. A Multi-Dimensional Convolutional Neural Network (CNN) was developed in [23] to estimate cyclone intensity. However, the time taken could not be minimized. The Center Location Algorithm was designed in [24] for tropical cyclone prediction with satellite infrared images. However, it needed to have constructed a perfect model to complete the difficult task of estimating cyclone intensity. An end-to-end lightweight quantitative modeling framework based on ensemble convolutional neural networks (ECNNs) was proposed in [25] to eradicate the required dimensionality reduction of the raw spectrum and other pre-processing operations. However, the accuracy of the prediction did not improve the ECNN method.

A. Novel Contributions of the Paper

The novel contribution of proposed CDHEKMF-BCDSLC method is described in below:

• To improve the prediction accuracy of cyclone intensity level, a Czekanowsky Dice Hypergraphic Extended Kalman Momentum Filterization based Bivariate Correlative Deep Structure Learning classification (CDHEKMF-BCDSLC) is designed with different processes namely segmentation, preprocessing, feature extraction and classification.

• To minimizing the prediction time of cyclone intensity, CDHEKMF-BCDSLC uses Czekanowsky dice Intensity threshold-based Interval Hypergraph model-based image segmentation to partition the input images into different parts.

• To improve the quality of images, invariant extended Kalman momentum filter is applied and extract the features based on Bivariate correlative Spatio-temporal Feature extraction on each pixel intensity of the image.

• To improve the accurate cyclone intensity prediction based on RuzickaIndexive Regression Function, multidimensional deep belief network classification model is designed.

• Extensive and comparative experiments were conducted to evaluate the quantitative and qualitative analysis of the proposed method with the different performance metrics.

B. Organization of the Paper

The rest of the paper is organized into different sections. Section II describes the proposed model with a different process for cyclone intensity estimation. Section III illustrates the experimental mode, followed by its analysis. Charts and tables are presented in Section IV. Section V concludes the paper.

SECTION II. Proposal Methodology

Cyclone intensity prediction is a challenging task since it involves extracting significant features, pre-processing, and various sets of parameters for analysis. Different processes in our model are shown in Figure 1.

FIGURE 1.

Architecture diagram of CDHEKMF-BCDSLC technique.

Figure 1 illustrates the architecture of the proposed technique consisting of four major processes for improving cyclone prediction accuracy. Images are taken from the North Pacific Cyclone dataset. Then the segmentation is carried out to partition the input cyclone image into several segments using the Czekanowsky dice Intensity threshold-based Interval Hypergraph (CDIT-IH) model. The preprocessing of segmented images is performed using an invariant extended Kalman momentum filter. Feature extraction process is carried out using machine learning technique. Feature selection process is performed by applying Bivariate correlative Spatio-temporal Feature extraction. Finally, deep structure learning is applied for classifying cyclone intensity as Depression, Deep Depression, Cyclonic storm, severe cyclonic storm, very severe cyclonic storm, extremely severe cyclonic storm, super cyclonic storm with higher accuracy. These four different processes of the proposed technique are explained in the following subsections.

A. Czekanowsky Dice Intensity Threshold-Based Interval Hypergraph Model-Based Image Segmentation

Multiple satellite images are first collected from the cyclone datasets. After collecting the input images, image segmentation is carried out by applying the Czekanowsky dice Intensity threshold-based Interval Hypergraph (CDIT-IH) model.

Suppose the cyclone images $I_{1},I_{\mathrm {2,}}I_{3}\mathrm {,\ldots .}I_{n}$ are mapped in two-dimensional spaces. The input image is segmented using Czekanowsky dice Intensity threshold-based Interval Hypergraph model. The hypergraph ‘$G=\left ({{ V,E }}\right)$ ’, where ‘V’ denotes set of vertices i.e. pixels $\alpha _{1},\alpha _{2},\alpha _{3}\mathrm {,\ldots .}\alpha _{n}$ in the input images and ‘edges E(I) denotes the connection (i.e. relationship) between the vertices.

Figure 2 illustrates Czekanowsky dice Intensity threshold-based Interval Hypergraph model for image segmentation. As shown in figure 2, the circle with red, blue, and green colors indicates the similar vertices connected each other (i.e., pixels). Circle denotes a segmented region of an image. The connections (i.e., relationships) between the vertices are measured using Czekanowsky dice similarity, which is given below,

$$\begin{equation*} \varphi =1-2\left ({{ \frac {\alpha _{i}\cap \alpha _{j}}{\alpha _{i}\cup \alpha _{j}} }}\right ) \tag {1}\end{equation*}$$

FIGURE 2.

Czekanowsky dice intensity threshold-based interval Hypergraph.

In the above equation, $\varphi$ indicates similarity coefficient, $\alpha _{i}$ denotes a pixel, where as $\alpha _{j}$ is the neighboring pixels. In (2), the intersection symbol $\cap$ indicates mutual dependence between the pixels, whereas the union symbol $\mathrm {'\cup '}$ signifies mutual independence between the pixels. The similarity coefficient provides values in the range from 0 to 1. Based on similarity values, the images are segmented as $p_{1},p_{2}\mathrm {,\ldots .}p_{k}$ vertices obtain different parts of the image. The threshold is the set to a segmented region.

$$\begin{align*} Y={\begin{cases} p_{k}\lt th;regionsuppressed \\ p_{k}\gt th;regionselected \\ \end{cases}} \tag {2}\end{align*}$$ Here ‘Y’ denotes segmentation output, ‘th’ indicates the threshold for the segmented region. If the segmented region below the threshold is suppressed and important segmented parts are selected for further processing. The segmentation process reduces the cyclone prediction time.

Algorithm 1 Czekanowsky dice Intensity threshold-based Interval Hypergraph

Input:

Cyclone dataset, number of images $I_{1},I_{2,}I_{3},\ldots .I_{n}$ ,

Output:

Image segmentation Begin

Step 1:

For each image $I_{i}$

Step 2:

Extract the number of pixels $\alpha _{1},\alpha _{2},\alpha _{3},\ldots .\alpha _{n}$

Step 3:

For each pixel ‘$\alpha _{i}$ ’ and neighboring pixel ‘$\alpha _{j}$ ’ S tep 4: Measure the similarity ‘$\varphi$ ’

Step 5:

Segment the images into different regions $p_{1},p_{2},\ldots .p_{k}$ in the hypergraph

Step 6:

End for

Step 7:

if ($p_{k}\lt th$ ) then

Step 8:

Region is suppressed

Step 9:

else

Step 10:

Select the region

Step 11:

End if

Step 12:

End for

End

Above Algorithm describes step-by-step process of image segmentation using Czekanowsky dice Intensity threshold-based Interval Hypergraph. Initially, images are collected from the dataset. Then the pixels are extracted from each image. Czekanowsky dice similarity is measured between any two pixels. Based on the similarity values, different segmented regions are obtained. Finally, the threshold is set to the segmented region. The selected regions are then used for further processing

B. Invariant Extended Kalman Momentum Filter-Based Preprocessing

Preprocessing is carried out using an invariant extended Kalman momentum filter to enhance the image contrast, which will enable us to extract the features from the segmented region accurately instead of extracting the features from the whole image. The invariant extended Kalman momentum filter is a nonlinear filter that linearizes about an estimation of the current mean and covariance between the pixels. The preprocessing step of the CDHEKMF-BCDSLC technique reducing the computational burden involved in accurate cyclone prediction.

Figure 3. Illustrates the block diagram of the invariant extended Kalman momentum filter to obtain the quality- enhanced image. It is a non-linear digital filtering technique that helps to remove the noise from an image. Such noise removal is a typical pre-processing step to improve the results of classification.

FIGURE 3.

Block diagram of the invariant extended Kalman momentum filter.

Invariant extended Kalman momentum filtering is very useful in digital image processing since it preserves edges while removing noise. By applying the filtering technique, the noise is processed as given below,

$$\begin{equation*} Q_{t+1}=fQ_{t}+\varphi _{t}x_{t}+I_{t}\theta _{t} \tag {3}\end{equation*}$$

Here, $Q_{t+1}$ denotes updated filtering results, f denotes state transition function, $Q_{t}$ current state, $\varphi _{t}$ denotes control input function, $x_{t}$ indicates control input, $I_{t}$ denotes noise input image, $\theta _{t}$ denotes process noise. This noise is removed from the input image for increasing the image quality.

Consider the segmented images ${SI}_{i}={SI}_{1},{SI}_{2},..{SI}_{k}$ and pixels in the input image which is represented as $\alpha _{1},\alpha _{2},\alpha _{3},\ldots .\alpha _{n}$ . The input image pixels are positioned in the form of a matrix (i.e. rows and columns) in the filtering window to find the noisy pixels and remove them from the image.

Figure 4 illustrates the filtering window which consists of different pixels. The denoising process is expressed as follows,

$$\begin{equation*} Z_{\left ({{ \alpha _{i}\thinspace \vert \thinspace \alpha _{i-1}}}\right ) }=\frac {1}{V\sqrt {2\pi }}\exp \left [{{ -\frac {1}{2}\left ({{ \frac {\alpha _{i}-m}{v} }}\right )^{2} }}\right ] \tag {4}\end{equation*}$$

FIGURE 4.

Filtering window.

Here $Z_{\left ({{ \alpha _{i}\thinspace \vert \thinspace \alpha _{i-1}}}\right) }$ denotes filtering output, where $\alpha _{i}$ denotes a pixel at the current state, and $\alpha _{i-1}$ denotes a pixel in the previous state where ‘m ’and v’ denotes mean and variance respectively.

Noisy pixels are any pixels that differ from the average. The segmented region is cleared of these noisy pixels. As a result, this helps to enhance the image quality.

Algorithm 2 Invariant Extended Kalman Momentum Filter

Input:

Segmented image

Output:

Preprocessed Image

Begin

Step 1:

For each segmented image ${SI}_{i}$

Step 2:

Arrange the pixels $\alpha _{1},\alpha _{2},\alpha _{3},\ldots .\alpha _{n}$ into filtering window

Step 3:

For each pixel ‘$\alpha _{i}$ ’ and neighboring pixel ‘$\alpha _{i-1}$ ’

Step 4:

Apply filtering technique

Step 5:

Measure relationship between pixels $Z_{\left ({{ \alpha _{i}\thinspace \vert \thinspace \alpha _{i-1}}}\right)}$ and mean ‘m’

Step 6:

If (pixel intensities are closer to mean) then

Step 7:

Pixels are said to be normal pixels

Step 8:

else

Step 9:

Pixels are said to be noisy pixels

Step 10:

end if

Step 11:

Remove the noisy pixels from the filtering window

Step 12:

Obtain quality improved image

Step 13:

end for

Step 14:

End for

End

The algorithm above outlines the detailed steps involved in the invariant extended Kalman momentum filter picture preprocessing procedure. The objective is to obtain the quality enhanced image. The pixels are arranged into rows and columns of the matrix. Then the relationship between the pixels in the filtering window and the mean value is estimated using the Gaussian covariance function. Based on the covariance estimation, normal and noisy pixels are identified. These noisy pixels are removed, and quality enhanced image is obtained. As a result, the mean square error is decreased and the peak signal to noise ratio is improved.

C. Bivariate Correlative Spatio-Temporal Feature Extraction

After preprocessing, the Bivariate correlative Spatio-temporal Feature extraction is performed for extracting the features over time and location. This technique is employed on each pixel intensity of the cyclone image, to extract the necessary features. A characteristic feature of cyclones is an area of the eye, a central region, the area of the cyclone, wind speed, and so on. Based on the correlation of each pixel’s intensity for two different states, both spatial and temporal features are extracted for accurate classification. To enhance the level of prediction accuracy of satellite imagery, it is required to reduce the noise. In our work a singular value decomposition methodology is adopted, which helps to provide clarity in the image. By applying SVD, it is also noted that the decomposition of the original image is done correctly by discarding the values that are not required. This methodology also reduces the image size and when it is quite useful in the reconstruction of the original image.

Bivariate correlation is applied for measuring the spatial representation of pixel intensities in a cyclone image and is estimated as follows,

$$\begin{equation*} \rho =\frac {\sum {\left ({{ \alpha _{i}-\mu _{i} }}\right )\left ({{ \alpha _{j}-\mu _{j} }}\right )}}{\sqrt { \sum {\left ({{ \alpha _{i}-\mu _{i} }}\right )^{2}\sum \left ({{ \alpha _{j}-\mu _{j} }}\right )^{2}}} } \tag {5}\end{equation*}$$

In the above equation $\rho$ denotes correlation coefficient, $\alpha _{i}$ , $\alpha _{j}$ denote pixel intensites, $\mu _{i}$ and $\mu _{j}$ indicate means of the pixels. The correlation coefficient gives a value between −1 and +1. The Bivariate correlation coefficient of ‘+1’ indicates positive correlation between the pixel intensity and mean whereas ‘-1’ indicates negative correlation. The area of the cyclone is determined as follows.

$$\begin{equation*} {Area}_{C}=\pi \ast {[\max {(ED)}]}^{2} \tag {6}\end{equation*}$$

Here, ${Area}_{C}$ denotes an area of the cyclone, $\pi$ is the mathematical constant (3.14) and $\max {(ED)}$ denotes maximum Euclidean distance from center to edge. The area of the eye is estimated as given below,

$$\begin{equation*} {Area}_{E}=\pi \ast {[ED (c,p) ] }^{2} \tag {7}\end{equation*}$$

$ED (c,p)$ denotes Euclidean distance from center to calm weather. To determine the strength of the cyclone, the wind speed of the storm is collected every hour. In this way, spatial and temporal features are extracted from the preprocessed image.

Algorithm 3 Bivariate correlative Spatio-temporal Feature extraction

Input:

Preprocessed image ${PI}_{i}={PI}_{1},{PI}_{2},..P$

Output:

extract the spatial and temporal features

Begin

Step 1:

For each preprocessed image ${PI}_{i}$

Step 2:

Apply bivariate correlation ‘$\rho$ ’

Step 3:

Extract the spatial representation of pixel intensity

Step 4:

Extract spatial features ‘${Area}_{C}$ ’, ${Area}_{E}$ , and wind speed

Step 5:

End for

Step 6:

Return (Features of Spatio-temporal)

Step 7:

End for

End

The algorithm above outlines the sequential steps of spatial and temporal feature extraction. For each preprocessed image, the spatial representation of the pixel intensity, area of the cyclone, eye area and speed of the wind are estimated at each time instance. This process minimizes the cyclone intensity estimation.

D. Multidimensional Deep Belief Network Classification Model

Finally, in our proposed technique, a multidimensional deep belief network classification model estimates the cyclone intensity prediction accurately. A multidimensional deep belief network classification model is a machine learning technique that consists of multiple layers for learning the given input (i.e. Spatio-temporal features). The deep learning model takes spatial and temporal features in the feature space to predict cyclone. Hence, it is called a multidimensional deep structure learning classification model.

Figure 5 illustrates the proposed multidimensional deep belief network classification model which consists of an input, output, and multiple hidden layers with nonlinearly activating nodes. The input and output layers are said to be visible layers. Each node in one layer is connected to consecutive layers with adjustable weights. Each layer performs a certain process to attain the final output.

FIGURE 5.

Schematic diagram of multidimensional deep belief network classification model.

The estimate for the neuron’s activity at layer ‘$\omega \left ({{ t }}\right)$ ’ of the input is shown below,

$$\begin{equation*} \omega \left ({{ t }}\right )=p+\sum \nolimits _{i=1}^{m} {i_{i}\left ({{ t }}\right )\delta _{0}} \tag {8}\end{equation*}$$

In (8), $i_{i}\left ({{ t }}\right)$ denotes an input feature, $\delta _{0}$ shows the regulating weights, and p indicates a bias value as ‘1’. The hidden layer is then used to do the feature analysis on the incoming input features. The feature analysis is performed using Ruziicka Indexive Regression Function. A machine learning method called regression is utilized to assess the relationship between the two variables. The Ruzicka Similarity Index is used to assess the relationship between the variables.

The Extracted Spatio-temporal feature is considered as input. The similarity between Spatio-temporal features is estimated as given below,

$$\begin{equation*} S=\frac {F_{E}\cap {F}_{T}}{\sum {F_{E}+\sum {{F}_{T}-F_{E}\cap {F}_{T}}}} \tag {9}\end{equation*}$$

Here, S denotes similarity coefficient, $F_{E}$ represents extracted wind speed feature, ${F}_{T}$ denotes testing cyclone wind speed feature, $F_{E}\cap {F}_{T}$ denotes mutual dependence between features. The Ruzicka similarity coefficient (S) provides the similarity value which lies between 0 and 1. If the extracted feature is close to the testing cyclone feature, then Ruzicka similarity coefficient (S) returns 1. Otherwise, the coefficient returns 0.

The intensity level is categorized as depression, deep depression, cyclonic storm, severe cyclonic storm, very severe cyclonic storm, very severe cyclonic storm, super cyclonic storm, etc. based on the calculation of wind speed.

The output of the hidden layer is expressed as follows,

$$\begin{equation*} H\left ({{ t }}\right )= \left ( \sum \nolimits _{i=1}^{m} {i_{i}\left ({{ t }}\right )\delta _{0}} \right)+\left(\delta _{1}\ast h_{t-1} \right ) \tag {10}\end{equation*}$$

Algorithm 4 Multidimensional Deep Belief Network Classification

Input:

extracted features

Output:

Increase prediction accuracy

Step 1:

Collect the extracted features as input // input layer

Step 2:

Input is transferred into the first hidden layer

Step 3:

For each $\mathbf {'}F_{E}$ ’

Step 4:

For each ${F}_{T}$

Step 5:

Measure the similarity ‘S’

Step 6:

if($S=+1$ )then

Step 7:

Classifies the cyclone intensity

Step 8:

end if

Step 9:

end for

Step 10:

end for

End

$H\left ({{ t }}\right)$ denotes output of the hidden layer at the instance of time ‘t’, $i_{i}\left ({{ t }}\right)$ represents input, and $h_{t-1}$ denotes previous hidden layer output. The weight $\delta _{0}$ , $\delta _{1}$ denotes weight between the input and hidden layer are multiplied with inputs received. The output layer then returns a classification of the cyclone’s intensity. This process is repeated with different weights until the classification technique attains the minimum error. This process increases the accuracy of cyclone intensity prediction and minimizes the rate of false positive.

$$\begin{equation*} Z= \delta _{2}\ast H\left ({{ t }}\right ) \tag {11}\end{equation*}$$ Z Indicates output of the output layer at a time ‘t’, $\delta _{2}$ symbolizes between the output layer and hidden layer in terms of weight.

Above algorithm classifies cyclone intensity using extracted features. Initially, the extracted features are given as input. Then the similarity-based regression is applied to analyze the extracted features and testing features. The severity of the cyclone is forecasted using the similarity value with the least amount of time and false positive rate.

SECTION III. Experimental Settings

Experimental evaluation of proposed CDHEKMF-BCDSLC technique is carried out using MATLAB with the hurricane cyclone image dataset taken from tropical cyclone repository of the Marine Meteorology Division of U.S. Naval Research Laboratory https://www.nrlmry.navy.mil/tcdat/. Cyclone images are collected from 68 Atlantic cyclones and 30 Pacific cyclones. The repository consists of 8,138 images for every 2 hours from 98 cyclones. 25 to 250 images are taken for conducting the simulation. The qualitative process of CDHEKMF-BCDSLC is shown in Figure 6.

FIGURE 6.

Qualitative results of CDHEKMF-BCDSLC technique.

Figure 7 illustrates the intensity estimation of classification at different stages such as Depression, Deep Depression, Cyclonic storm, severe cyclonic storm, very severe cyclonic storm, extremely severe cyclonic storm, super cyclonic storm.

FIGURE 7.

Intensity estimation of classification.

SECTION IV. Quantitative Analysis

Several quantitative measures are used to estimate the quantitative analysis of the new CDHEKMF-BCDSLC approach and the existing TCICENet [5] and TCNN [6]. These metrics are described below. Cyclone prediction accuracy: it describes the percentage of cyclone photos that are successfully predicted compared to the entire sample of photographs used as input. Therefore, the prediction accuracy is estimated as given below,

$$\begin{equation*} PA=\left [{{\frac {No. of~cyclone~ images~correctly~predicted}{n}}}\right ]*100 \tag {12}\end{equation*}$$ where, PA denotes a cyclone prediction accuracy ‘n’ corresponding to the number of input images. Higher the cyclone prediction accuracy is measured in percentage (%).

Precision: It is measured based on the number of true positives and false positives. The formula for calculating the precision is expressed as follows,

$$\begin{equation*} P= \left [{{ \frac {T_{p}}{T_{p}+F_{p}} }}\right ]\ast 100 \tag {13}\end{equation*}$$ where, P denotes a precision measure, $T_{p}$ indicates true positive, $F_{p}$ refers the false positive. The precision is measured in terms of percentage (%).

Recall: It is measured based on the number of true positives and true negatives. The recall rate calculation formula is written as follows,

$$\begin{equation*} R= \left [{{ \frac {T_{p}}{T_{p}+T_{n}} }}\right ]\ast 100 \tag {14}\end{equation*}$$ where, R denotes a recall measure, $T_{p}$ indicates a true positive, $T_{n}$ refers to the true negative. The Recall is quantified as a percentage (%).

F-measure: It is evaluated using recall and precision. The following mathematical equation is used to estimate the f-measure,

$$\begin{equation*} \mathrm {f}_{\mathrm {measure}}\mathrm {= 2\ast }\left [{{ \frac {\mathrm {P\ast R}}{\mathrm {P+R}} }}\right ]\ast 100 \tag {15}\end{equation*}$$ where, $f_{measure}$ denotes an F-measure, P denotes precision R denotes a recall. The F-measure is expressed in percentage (%) terms.

Cyclone prediction time: It refers to the amount of time consumed for predicting the cyclones for experimentation. The cyclone prediction time is estimated as given below.

$$\begin{equation*} \mathrm {PT=n\ast Time}\left [{{ \mathrm {Cyclone~prediction} }}\right ] \tag {16}\end{equation*}$$ where PT denotes a cyclone prediction time, ‘ n’ denotes the number of input images. ‘PT’ is quantified in milliseconds (ms).

False positive rate: It is calculated in percentage (%) and evaluated as the ratio of the number of satellite cyclone photos that were incorrectly forecasted to the total number of input images. Where, False positive rate ‘FPR’ is estimated.

Specificity: It measures the ratio of negatives that are correctly identified as not having the cyclone image. It is estimated in percentage (%)

$$\begin{equation*} Specificity=\frac {TN}{\left ({{ TN+FP }}\right )}\ast 100 \tag {17}\end{equation*}$$

Peak Signal to Noise Ratio: PSNR is described as the input image quality and based on Mean Square Error (MSE) after the preprocessing task. MSE is calculated as follows and is defined as the ratio between the original cyclone image and the denoised images,

$$\begin{align*} PSNR& =10\ast {log}_{10}\left [{{ \frac {R^{2}}{MSE} }}\right ] \tag {18}\\ MSE& = \left ({{ I-I^{\prime } }}\right )^{2} \tag {19}\end{align*}$$ where ‘PSNR’ refers to the ratio of peak signal-to-noise, ‘R’ indicates a greatest pixel value achievable (i.e.255), ‘MSE’ refers a mean square error, ‘I refers the input cyclone image original size, ‘I’ refers to the sized denoised images. Decibel measurements are used to calculate the PSNR (dB).

The performance results of the three techniques—CZHEKMF-BCDSLC, TCICENet, and TCNN—about a range of 25 to 250 photos taken are shown in Table 2 and Figure 8 for cyclone prediction accuracy. When compared to conventional classification algorithms, the performance result shows that the suggested CDHEKMF-BCDSLC has a greater cyclone prediction accuracy. With the aid of statistical estimation, this is demonstrated.

TABLE 1 Example of Intensity Estimation

TABLE 2 Comparison of Prediction Accuracy

FIGURE 8.

Performance results of prediction accuracy.

The overall average value shows that, when compared to cutting-edge techniques, CDHEKMF-prediction BCDSLC’s accuracy is greatly increased by 5% and 7%. This is due to the application of deep belief network architecture. The regression function is applied to deep belief network architecture to analyze the extracted features and testing features. This results in the estimation of different intensity levels of the cyclone with higher accuracy.

Table 3 and figure 9 illustrate the performance results of precision of cyclone intensity estimation using CDHEKMF-BCDSLC technique and existing techniques. The acquired results show that the CDHEKMF-BCDSLC technique and existing techniques are shaded as blue, red, and green, respectively, depending on their degree of precision. Ten iterations of each approach are carried out with various numbers of input images.

TABLE 3 Comparison of Precision

TABLE 4 Comparison of Recall

FIGURE 9.

Performance results of precision.

FIGURE 10.

Performance results of the recall.

The results of CDHEKMF-BCDSLC technique are compared with existing methods. It is observed that the precision of the CDHEKMF-BCDSLC technique has considerably improved by 3% when compared to [5] and 6% when compared to [6]. Before classification, image segmentation and preprocessing are carried out. This ensures accurate classification besides resulting in minimization of false positive rate and maximization of true positive rate.

Figure 11 and Table 5, which are provided below, show the experimental F-Measure results. The table and graph display the F-Measure results for three different approaches in relation to various numbers of input images. Among the three methods, it is observed that the CDHEKMF-BCDSLC technique provides superior performance when compared to the other two conventional methods. According to the overall comparison performance, the CDHEKMF-BCDSLC approach enhances F-Measure outcomes by 6% when compared to TCNN [6] and by 3% when compared to TCICENet [5]. One reason why the CDHEKMF-BCDSLC technique shows better precision and recall rates, as well as an increase in F-measure, may be because a deep classification technique is being applied.

TABLE 5 Comparison of F-Measure

FIGURE 11.

Performance results of F-measure.

The experimental recall findings for various numbers of cyclone images in the range of 25 to 250 are displayed in Table 4 and Figure 10. The table and graphical data show that the CDHEKMF-BCDSLC methodology performs better than the other two methods currently in use. Overall comparative results show that the CDHEKMF-BCDSLC approach has a 2% recall improvement over TCICENet [5] and a 5% recall improvement over TCNN [6]. Note that CDHEKMF-BCDSLC technique uses Ruzicka similarity index effectively, analyzes the features of the cyclone and identifies different intensity levels.

Table 6 and figure 12 display the cyclone prediction times for three distinct approaches, TCICENet [5], TCNN [6], and CDHEKMF-BCDSLC technique, with respect to a variety of cyclone images. The findings show that the three approaches’ cyclone prediction times get faster as the quantity of cyclone images increase. Among the three methods, the CDHEKMF-BCDSLC technique outperforms the other two techniques with a lesser prediction time. For each method, ten results are obtained using a different number of input images. The average of ten comparison results confirms that the overall cyclone prediction time of the CDHEKMF-BCDSLC technique is considerably reduced by 12% and 22% when compared to TCICENet [5] and TCNN [6] respectively. This might be because Czekanowsky dice Intensity threshold-based Interval Hypergraph model is employed for image segmentation. With the segmented images, preprocessing is Carried out to remove the noise and to accurately extract the features. Finally, deep classification is performed based on the regression analysis which results in minimization of the cyclone prediction time.

TABLE 6 Comparison of Prediction Time

FIGURE 12.

Performance results of prediction time.

The comparison result for the false positive rate is described in Table 7 and figure 13. The number of images lies in the range of 25 to 250 and we consider 10 iterations. From the table, the incorrect prediction of cyclone images using the proposed CDHEKMF-BCDSLC technique is reduced when compared to other methods. The overall results of the false positive rate are minimized using the CDHEKMF-BCDSLC technique by 27% when compared to TCICENet [5], and 39% when compared to TCNN [6]. The reason for the lesser false positive rate is to apply the regression function for examining the extracted features and testing features.

TABLE 7 Comparison of False Positive Rate

FIGURE 13.

Performance results of false positive rate.

Table 8 and figure 14 displays the specificity versus the number of cyclone images collected from the database. The comparison of specificity using the proposed CDHEKMF-BCDSLC is made with existing [5] and [6]. From this comparison, the results of specificity using the proposed CDHEKMF-BCDSLC are higher than the existing methods. This is because of the application of the Ruzicka similarity index to discover the training and testing features with higher specificity. The overall results of specificity are increased using the CDHEKMF-BCDSLC technique by3% and 5% than TCICENet [5], TCNN [6].

TABLE 8 Comparison of Speciality

FIGURE 14.

Performance result of specificity.

Table 9 and figure 15 explains the comparative result analysis of peak signal to noise ratio for existing and proposed methods with respect to different images of various size. The peak signal to noise ratio using the proposed CDHEKMF-BCDSLC technique is compared with two of the existing methods such as TCICENet [5], and TCNN [6]. The proposed CDHEKMF-BCDSLC technique achieves a higher peak signal to noise ratio when compared to the other two methods. With the application of the Kalman Momentum Filter model, noisy pixel images are removed resulting in improved image quality. With the preprocessed images, mean square error is minimized with an increased peak signal to noise ratio. The peak signal-to-noise ratio is increased by 8% and 15% when compared to state-of-the-art methods.

TABLE 9 Comparison of Peak Signal to Noise Ratio

FIGURE 15.

Performance result of Peak Signal to Noise Ratio.

A. Invariant Extended Kalman Momentum Filter-Based Preprocessing

The proposed Czekanowsky Dice Hypergraphic Extended Kalman Momentum Filterization based Bivariate Correlative Deep Structure Learning Classification (CDHEKMF-BCDSLC) model aims to enhance cyclone prediction accuracy through a structured approach involving segmentation, preprocessing, feature extraction, and classification.

1) Segmentation

Czekanowsky Dice Intensity Threshold-based Interval Hypergraph (CDIT-IH) Model: This model is employed to segment cyclone images effectively. By applying intensity thresholds within an interval hypergraph framework, the segmentation process is optimized to reduce the time required for accurate cyclone prediction.

2) Preprocessing

Invariant Extended Kalman Momentum Filter: This novel filter is designed to enhance the contrast of cyclone images, making the relevant features more distinguishable. Improved image contrast is crucial for accurate feature extraction in subsequent steps.

3) Feature Extraction

Bivariate Correlative Spatio-Temporal Feature Extraction: This step involves analyzing each pixel’s intensity over time and space to extract meaningful features. By correlating spatial and temporal data, the model captures the dynamic nature of cyclones, leading to more accurate predictions.

4) Classification

Multidimensional Deep Belief Network (DBN): This machine learning technique comprises multiple layers that learn and classify the extracted spatio-temporal features. The DBN is designed to handle the complexity of the data, improving prediction accuracy through its deep structure.

The dataset used for the experimental evaluation consists of hurricane cyclone images obtained from the Marine Meteorology Division’s tropical cyclone repository of the U.S. Naval Research Laboratory. Key details of the dataset are as follows:

• Source: NRLMRY Tropical Cyclone Repository

• Composition:

  • 68 Atlantic cyclones

  • 30 Pacific cyclones

• Total Images: 8,138 images taken at 2-hour intervals from 98 cyclones.

• Simulation Subset: 25 to 250 images from the above dataset were used for conducting simulations in MATLAB.

The model’s performance was evaluated using several metrics, including prediction accuracy, precision, recall, F-measure, and prediction time. The results showed significant improvements in these metrics compared to existing state-of-the-art methods:

• Accuracy: 6% improvement

• Prediction Time: 17% reduction

• Precision: 5% improvement

• F-measure: 5% improvement

• Recall: 4% improvement

The CDHEKMF-BCDSLC model leverages advanced image processing and machine learning techniques to enhance cyclone prediction accuracy. By utilizing a comprehensive dataset and optimizing each processing step, the model achieves superior performance, providing a reliable tool for early cyclone prediction and mitigating the impact of cyclones on life and property.

B. Potential Limitations

CDHEKMF-BCDSLC technique uses the extended Kalman momentum filter to eliminate the noise. However, the Spatial-temporal resolution of the images is not improved to get the finer details of the cyclone images. This affects the precise measurements or observations during the cyclone prediction.

Satellite databases are huge in size as it covers different regions of the sea or ocean. Proposed CDHEKMF-BCDSLC technique takes more time to create useful images from the raw images for predicting cyclone. Though proposed CDHEKMF-BCDSLC technique increases the accuracy, the space complexity involved in the satellite image prediction is not reduced.

SECTION V. Conclusion

This paper introduces a reliable and robust technique called CDHEKMF-BCDSLC for intensity estimation of cyclones using a deep neural network. This technique also ensures that cyclone prediction accuracy is achieved. Czekanowsky dice Intensity threshold-based Interval Hyper graph model segments the entire image into different parts. Image preprocessing is then carried out to enhance the quality of the image by removing noisy pixels. Spatio-temporal features are then extracted from the image. With the extracted features, the classification of different intensity levels is performed based on the regression function. The experimental evaluation is carried out using a different cyclone image repository. The quantitative and qualitative performance analyses are carried out using the CDHEKMF-BCDSLC technique and other methods. The quantitative outcome indicates the advantages of the CDHEKMF-BCDSLC technique and shows that this technique is better in terms of 6%, 5%, 4%, and 5% of higher prediction accuracy, precision, recall, F-measure, and 17% of lesser time consumption when compared with other conventional methods.

ACKNOWLEDGMENT

The authors extend their appreciation to Taif University, Saudi Arabia, for supporting this work through project number (TU-DSPP-2024-50).