Deep Convolutional Network Based Machine Intelligence Model for Satellite Cloud Image Classification

The deep convolutional network has been implemented on Python with the help of keras API as the front end and TensorFlow as the backend. The keras and TensorFlow provide ease of implementing the deep convolutional network, as it contains predefined layers, models, and optimizers^{[ 19 ]}. The implementation platform is GPU NVIDIA MX250, type GDDR5, and has 1518 MHz and a RAM of 8 GB, which enhances the speed of training the model on a 64-bit Windows platform. A sequential model, which contains layers (Conv2D, Max-pooling, Flattening, Dropout) with

the input of 224 pixel $\times$ 224 pixel image, is illustrated in Fig. 3 . All training and validation images after processing (loading, resizing, and labeling) pass through it. The models are trained and validated in every epoch (once a total dataset passes through forward and backward in the network or sequential model). Some sample images for cyclone and non-cyclone are displayed in Fig. 4 and Fig. 5 , respectively. An image of 224 pixel $\times$ 224 pixel is given as the input to the sequential model where two convolutional 2D networks are applied, taking 64 filters, defining the kernel of 3 $\times$ 3, and keeping the padding the “same”. Each convolutional feature extraction takes place in a feature matrix, and the matrix is followed to the next layer for further processing.

10.26599/BDMA.2021.9020017.F003 Fig. 3Proposed sequential model using ReLU.

10.26599/BDMA.2021.9020017.F004 Fig. 4Bulbul cyclone images in the dataset (sample images)^{[ 20 , 21 ]}.

10.26599/BDMA.2021.9020017.F005 Fig. 5Normal images in the dataset (sample images)^{[ 20 , 21 ]}.

As previously mentioned, convolution is applied to three segments, and the filters double in each segment. Therefore, feature detection gives high accuracy to predict the testing images. Dropout is mainly introduced for avoiding overfitting. The model performance is evaluated using a confusion matrix^{[ 43 , 44 ]}.

As mentioned above, the model starts with an input image where the first convolutional layer extracts 64 feature matrices taking a kernel size of 3 $\times$ 3. The input of the second convolutional layer is the output of the first convolutional layer. Here the convolution again extracts 64 features from the feature matrices of the previous layer, taking a kernel of 3 $\times$ 3. Down-sampling images is the main objective of the CNN model, and therefore, a MaxPool2D, which is also called a max-pooling layer, reduces images without reducing their features. The max-pooling layer retrieves all the features into a feature map/matrix. The pool size of 2 $\times$ 2 reduces the feature matrices to half of the original size. The output of the max-pooling layer is the input of the third convolutional layer. In the second segment of the model, which contains the third and fourth convolutional layers, 128 feature matrices are extracted, followed by a dropout of 0.5, which prevents the model from overfitting. The second max-pooling layer again down-samples the feature matrices into half of the previous size by using a pool size of 2 $\times$ 2. The third segment of the model, which contains three convolutional layers with 256 feature matrices each and a kernel size of 3 $\times$ 3, again extracts the appropriate features for classifying the problem. The max-pooling layer again comes into the role of down-sampling. The input feature matrices, which are the outputs of the previous convolutional layer, are finally reduced into many feature matrices. The model at this phase deals with the matrices of size 28 $\times$ 28 and 256 feature matrices or filters. Thus, converting these 2D arrays or matrices into a single vector is carried by the flattening layer whose output is 28 $\times$ 28 $\times$ 256, which is equal to 200 704. The model now deals with 200 704 parameters, which are being distinguished by a dense layer of two units, as the model is a binary classifier. The SoftMax classifies the parameters into a probability of a cyclone or not. The epochs “ $E_p$ ” defined in this article are 10, and the batch size is 10. The percentages of accuracy, loss, validation loss, and validation accuracy are presented in Table 1 .

We can also define model loss and accuracy with reference to the image shown in Fig. 6 . Here, accuracy and loss are optimal with respect to per epoch. From the first epoch, the training data show a condition called underfitting. However, while the epoch increases, the training data tend to fit the model. In $Ep$ -1, the accuracy is nearly 93%, and the loss is considered nearly 8%. It suggests that during the first batch of the dataset with 494 images, 93% are trained accurately; meanwhile, 7% are not trained accurately. In $Ep$ -2, the accuracy is approximately 98%, and the loss percentage is approximately 2.0. The model has significantly trained many images accurately from the batch size. As soon as the model is showing a good result at a point exactly at $Ep$ -8, the model starts to overfit and thus signifies that the training accuracy is more, whereas the validation accuracy is less. However, in $Ep$ -9, the accuracy reduces to 1%, and the model perfectly fits. In the last epoch, the accuracy level of model training is nearly 98%, and the accuracy level of validation is 99%. With these results, we can easily confirm that the model fits the data. The underfit problem is also eradicated.

10.26599/BDMA.2021.9020017.F006 Fig. 6Training and validation accuracy with respect to per epoch.

Figure 6 displays the graph between validation accuracy and training accuracy. The deviation between the two lines is evident at the beginning of epochs. However, the deep neural network during the training increases its accuracy, and the deviation between the two lines decreases. The lines shown in Fig. 6 tend to merge at the ninth epoch, justifying the fitting of the proposed model to the training and validating datasets.

Now, the confusion matrix helps us determine the accuracy of the total model. The formula of the confusion matrix is determined in Eq. ( 3 ). It can be defined by the ratio of the addition of true positive and false positive with respect to the total number of instances. The total number of testing images is 1008, out of which 408 are random cyclone images of AMPHAN and OCKHI. These images are not considered before in the training or validation dataset, and the model has predicted 352 images to be cyclones^{[ 20 , 21 ]}. Thus, it has also predicted 587 images to be non-cyclones from 600 images. The accuracy can be defined by the following formula:

(3) $$\text{Accuracy}=\frac{\alpha +\beta }{\alpha +\beta +{\alpha }^{\prime }+{\beta }^{\prime }}$$

where $\alpha$ is denoted as a predicted “cyclone” of a cyclone class, ${\alpha }^{\prime }$ is denoted as predicted “not a cyclone” but belongs to the cyclone class, $\beta$ is denoted as predicted “not a cyclone” of the no_cyclone class, ${\beta }^{\prime }$ is denoted as a predicted “cyclone” which belongs to the no_cyclone class. Putting the values of $\alpha =352,$ $\beta =597$ , $\alpha =56$ , and ${\beta }^{\prime }=3$ , the model obtains an accuracy of 94.14%, and an error rate of less than 6%. The confusion matrix of the proposed model is illustrated in Fig. 7 . The corresponding experimental results are outlined in Table 2 .

10.26599/BDMA.2021.9020017.F007 Fig. 7Confusion matrix of the proposed model.

From the summarized results in Table 2 , the proposed model shows an appealing detection accuracy of 94.2%. The model reveals not only detection accuracy but also a sensitivity of 86.3%, which appears to be unique in the field of cyclone detection. Another point of observation is that the false positive rate is quite impressively low, with only 0.005. It shows the real power of our proposed system, as it seems to be stable in all the performance measures.

In the final stage of analysis, we compare our proposed cyclone detection model with other state-of-the-art cyclone detection models. Given that our model is based on a deep learning technique, we shortlist a few deep learning based models for comparison in Table 3 . In this regard, DLR–FH, Deep CNN, and DAV–T cyclone detection models implemented by Rajesh et al.^{[ 7 ]}, Deep CNN (2D) model proposed by Matsuoka et al.^{[ 8 ]}, Multilayer CNN proposed by Kovordányi and Roy^{[ 17 ]}, and PCA-based CNN model proposed by Rai et al.^{[ 22 ]} are considered. The proposed model exhibits at par performance with other models in terms of detection accuracy. The proposed multilayer CNN model shows 12.20% and 5.1% better detection accuracy than the deep learning approach proposed by Matsuoka et al.^{[ 8 ]} and the DAV-T approach implemented by Rajesh et al.^{[ 7 ]}, respectively. Our system also slightly lacks less than DLR-FH, Multilayer CNN, and PCA-CNN approaches.