A Multichannel-Based Deep Learning Framework for Ocean SAR Scene Classification

Abstract

High-resolution synthetic aperture radars (SARs) are becoming an indispensable environmental monitoring system to capture the important geophysical phenomena on the earth and sea surface. However, there is a lack of comprehensive models that can orchestrate such large-scale datasets from numerous satellite missions such as GaoFen-3 and Sentinel-1. In addition, these SAR images of different ocean scenes need to convey a variety of high-level classification features in oceanic and atmospheric phenomena. In this study, we propose a multichannel neural network (MCNN) that supports oceanic SAR scene classification for limited oceanic data samples according to multi-feature fusion, data augmentation, and multichannel feature extraction. To exploit the multichannel semantics of SAR scenes, the multi-feature fusion module effectively combines and reshapes the spatiotemporal SAR images to preserve their structural properties. This fine-grained feature augmentation policy is extended to improve the data quality so that the classification model is less vulnerable to both small- and large-scale data. The multichannel feature extraction also aggregates different oceanic features convolutionally extracted from ocean SAR scenes to improve the classification accuracy of oceanic phenomena with different scales. Through extensive experimental analysis, our MCNN framework has demonstrated a commendable classification performance, achieving an average precision rate of 96%, an average recall rate of 95%, and an average F-score of 95% across ten distinct oceanic phenomena. Notably, it surpasses two state-of-the-art classification techniques, namely, AlexNet and CMwv, by margins of 23.7% and 18.3%, respectively.

1. Introduction

Synthetic aperture radar (SAR) has proven itself as a vital technology for a wide variety of air–sea interaction processes [1,2] to exploit its multi-dimensional constructional capability. It allows various satellite missions to obtain thousands of spatiotemporal images, e.g., ERS-1/2 (1991–2003), Environmental Satellite (ENVISAT)/Advanced Synthetic Aperture Radar (ASAR) (2002–2012), TerraSAR-X (2007–present), RADARSAT-1/2 (1995–present), GaoFen (GF)-3 (2016–present), and Sentinel-1 (S-1) (2014–present). Due to the independence between atmosphere and sunlight, the accessibility of SAR images offers unprecedented convenience for exploration in research directions such as bathymetric mapping [3], ocean current analysis [4], wind direction estimation [5], ship target detection [6], and ice shelf monitoring [7]. Moreover, SAR data have been used in diverse applications, including fish species recognition [8], urban land classification [9], marine spatial planning [10], and earthquake disaster prediction [11]. To further broaden the versatility of SAR technology and highlight the potential to contribute to an array of scientific and practical domains, geophysical researchers have employed machine learning and deep learning techniques to support fish species recognition, military objective protection, and natural hazard prevention [12,13,14]. However, SAR images inherently exhibit speckle noise due to the attenuation of echo signals, which is characterized by a distribution of granular patterns [15]. This noise is typically characterized as multiplicative, not only affecting SAR images but also influencing all coherent images. Therefore, the primary challenge with the automated classification of SAR images becomes speckle noise interference. On the one hand, the noise poses difficulties in multiple tasks such as feature extraction and edge detection [16]. This interference not only diminishes the contrast of the images but also modifies the spatial statistics of the underlying scene backscatter, thereby compromising the interpretation and recognition of targets. On the other hand, the speckle noise interference poses a big challenge for oceanic scene classification with multiple SAR images and deteriorates the reliability of prediction models like predicting meteorological conditions and analyzing atmospheric patterns [17,18,19]. As depicted in Figure 1, the filtered image (Figure 1b) reveals enhanced details, in contrast to the original SAR image (Figure 1a), illustrating improvements achieved through noise removal.

Due to the small differences in backscattering characteristics [21], it is difficult to precisely distinguish the specific ocean phenomenon from similar marine environments. Moreover, the surface current features cannot be extracted because of the wind to backscatter data in dual-polarized mode [22]. Furthermore, there is presently a lack of comprehensive annotated datasets for SAR targets. This limitation arises from the expensive nature of acquiring SAR data and the associated high costs involved in ensuring quality annotation [23]. Hence, relatively insufficient labeled data are used for training and modeling due to the long acquisition period and expensive labeling process [24]. Two recent works report that most previous ocean SAR image-based applications just involve limited regional scene case studies such as those conducted in the East China Sea and the Malacca Strait, where the SAR images are employed for the identification of ocean eddies and the analysis of the spatial-temporal distribution to investigate the characteristics of internal solitary waves [25,26].

To tackle the aforementioned issues, deep neural network techniques have been well-studied to alleviate the influence of speckle noise and backscattering features [27]. For example, a deep encoder–decoder convolutional neural network (CNN) architecture is established for the speckle filtering tasks [28], which outperformed classical algorithms on both simulated and real data. A noisy reference-based SAR deep learning filter was proposed [29], achieving a better despeckling performance over certain state-of-the-art deep learning techniques. But when these deep learning methods are simply applied to SAR image-based detection, some important SAR domain knowledge and corresponding oceanic features are eliminated by specific feature extraction and transformation [30]. Compared to deep learning strategies, manual classification and recognition significantly depend on visual features [31], introducing inefficiencies in processing marine big data. Each visual task necessitates predefined artificial features, resulting in time-consuming efforts. Furthermore, these manually defined features are relatively low-level when compared to those extracted by deep learning methods, thereby constraining the comprehensive representation of marine objects. Therefore, for effective utilization of SAR images by scientific users, it is critical to develop an appropriate deep learning framework for standardizing the use of SAR images so that they can be seamlessly integrated into the repertoire of the public oceanic datasets for category information analysis.

Recently, data augmentation has been extensively studied to synthetically enhance the diversity of training datasets by a magnitude of 1000 within image rotation, random transformation, and horizontal flip [32]. Furthermore, several efforts have explored the use of data augmentation to further boost the performance of artificial intelligence methods for automatic SAR image classification and oceanic phenomenon detection [33,34]. In addition, the transfer-learning strategy has been verified as an efficient way to develop a robust CNN model for specific applications in the case of limited datasets [23]. Hence, an effective oceanic phenomenon recognition system also needs to provide an augmentation mechanism for SAR scene detection applications to efficiently expand the data size without costly data collection.

In addition, the routine SAR wave mode (WM) measurements from multiple satellite missions typically exhibit a series of fine spatial resolution (4 m), large scene footprints (20 km × 20 km), and high signal-to-noise ratios, which improves SAR imagery despeckling effects [35]. Because the massive wave mode data at the global scale, approximately 120 k images per month, contain redundant geophysical information, it is feasible to dynamically capture the useful geophysical properties by a dynamic method that combines feature extraction utilizing the wavelet–radon transform (WR) and classification employing the neural network technique, specifically the dynamic neural network (DNN) [36]. The conventional image processing methods are commonly based on visual inspection and handcrafted features [37,38], which are notably impractical for the generalization of the structural patterns of different geophysical phenomena [39,40].

CNN architectures serve as an automated framework for feature learning and representation, facilitating the extraction of features from images in a unified spatiotemporal manner of the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) [41]. AlexNet, introduced by Krizhevsky et al. in 2012 [42], emerges as the pioneering large-scale convolutional neural network (CNN) in the field of computer vision. It achieves a notable image classification performance, attaining an 83.6% top-five accuracy compared to the manual classification methods (e.g., histograms of oriented gradients (HoG) [43], scale invariant feature transform (SIFT) [44], and local binary pattern (LBP) [45]). The architecture known as GoogLeNet or Inception-v1, as detailed in the work by [46], won the ILSVRC 2014 challenge with an accuracy rate of 93.7%. Then, the classical Inception architecture evolved into both Inception-v2 and Inception-v3 [47], achieving an incredible performance with a 94.4% accuracy on the ILSVRC 2012 classification dataset, in contrast to the manual classification method SIFT [44]. A variant of Inception-v3 CNNs (denoted CMwv by [48]) has been employed for automatic SAR image classification and achieved the optimal accuracy for geophysical phenomenon detection on the public dataset. However, CMwv exhibits limited generalization capability in capturing oceanic signatures, as evidenced by its notably low precision level, approaching zero, in the detection of oceanic fronts (OF), and a recall level below 50% in the identification of pure ocean waves (PW). Moreover, the utilization of Inception-v3’s image input techniques by CMwv leads to a lack of consideration for distinctive features associated with various oceanic phenomena and the domain-specific knowledge embedded in SAR images.

Taken together, for a greater good on improving model reliability and guiding the spatiotemporal analysis, it is imperative to integrate oceanic scene classification with sophisticated characteristics of natural images for a combined scheme that can overcome the inefficiency of geophysical phenomenon prediction observed in several WV datasets. In this study, we present a multichannel neural network (MCNN) management scheme to mitigate the impacts of insufficient data samples on ocean SAR scene classification. Based on multichannel data augmentation, MCNN involves multiple versions of standard CNN models equipped with convolution kernels of different sizes ($1\times 1$, $3\times 3$, $5\times 5$), conducting convolution operations to capture image features at different scales. This design enables the model to emphasize both local and global features, thereby improving its capacity to recognize ocean SAR scenes of varying scales. Moreover, MCNN employs a comprehensive approach to extract multiple features from the original SAR images. These extracted features are then organized into multiple channels within the intricately designed Inception CNN module, effectively fulfilling the goal of capturing deeper features. Additionally, several techniques have been incorporated into MCNN, such as multi-feature fusion, data augmentation, and multichannel feature extraction, to address the challenges arising from the aforementioned inadequate data. We use a number of SAR image datasets to evaluate the performance of MCNN and validate our design strategies through experimental analysis.

In summary, our paper makes the following contributions:

• We first illustrate and describe the wave mode data from the Sentinel-1 and TenGeoP-SARwv datasets. To cover the whole stereoscopic structure of the geophysical phenomena, we construct a new multi-dimensional dataset and data processing module that efficiently extracts, concatenates, and processes different area data from multiple satellite missions.
• We propose a novel deep learning framework, namely, MCNN, including multi-feature fusion for capturing valuable information, data augmentation for perturbing feature embedding, and multichannel feature extraction for accurately distinguishing different types of oceanic phenomena.
• We evaluate the performance of MCNN with a broad set of WV data and deep learning methods. Our experimental results demonstrate that MCNN achieves perfect classification accuracy and significantly improves the data quality, outperforming two state-of-the-art methods by 23.7% and 18.3%, respectively.

The rest of the paper is organized as follows: Section 2 illustrates the Sentinel-1 wave mode and TenGeoP-SARwv datasets and Section 3 details the design of the MCNN framework. Section 4 provides an analysis of ocean SAR scene classification accuracy and presents the experimental results, followed by Section 5, which concludes this paper.

2. Dataset and Data Processing

In this study, we mainly use ocean SAR images from Sentinel-1 (S-1) in different wave modes because of its routine wave mode measurements at the global scale. To train a greater deep learning classification architecture, we construct training datasets from the TenGeoP-SARwv database. All the related datasets are illustrated in the following.

2.1. Sentinel-1 Wave Mode

The Sentinel-1 (S-1) mission is designed as a two-satellite constellation including polar-orbiting and sun-synchronous satellites [35,49], namely, Sentinel-1A and Sentinel-1B, launched by the European Space Agency (ESA) in April of 2014 and 2016, respectively. In order to provide an effective 6-day repeat cycle, the two satellites share the same orbital plane offset by a 180° phase difference. The S-1 microwave SAR instruments operate at a wavelength of approximately 5.5 cm. These satellites are equipped with identical microwave SAR instruments that operate four imaging modes: WV, strip map, extra-wide swath, and interferometric. WV is the primary imaging mode used over open ocean, unless there are specific requests for other imaging modes. S-1 WV small SAR image scenes are acquired in a leapfrog scheme at two alternating center incidence angles of 23° (WV1) and 36.5° (WV2). The small SAR image scenes are referred to as imagettes in this research. Each imagette has dimensions of 20 km × 20 km and a spatial resolution of 5 m. Considering that certain horizontal (HH) images have been acquired during special phases, the two incidence angles frequently apply in linear vertical (VV) image transmission and polarization receipt. This study mainly focuses on S-1A WV data, considering that S-1B imagettes exhibit essentially equivalent characteristics to those of S-1A imagettes [48]. The joint utilization of S-1A and S-1B satellites will enhance the coverage in both temporal and spatial dimensions for various applications related to geophysical phenomena. Furthermore, the designed classification models can also be applied to S-1B.

2.2. TenGeoP-SARwv Dataset

TenGeoP-SARwv is a comprehensive dataset containing over 37,000 labeled SAR images (20 km × 20 km) captured over the ocean [50], each associated with ten well-defined geophysical phenomena described in Figure 2 that are frequently observed and expertly categorized. This dataset holds significant utility for researchers and practitioners engaged in the fields of oceanography, meteorology, and remote sensing, as well as those involved in deep learning endeavors. Additionally, this dataset comprises 16,068 imagettes acquired in WV1 mode and 21,485 imagettes acquired in WV2 mode. Each imagette within this dataset is annotated to correspond to ten geophysical phenomena, encompassing both oceanic and meteorological characteristics. For instance, Figure 2 demonstrates ten conventional geophysical phenomena from Figure 2a–j, including atmospheric fronts (AF), biological slicks (BS), icebergs (IB), low-wind areas (LWA), microconvective cells (MCC), oceanic fronts (OF), pure ocean waves (PW), rain cells (RC), sea ice (SI), and windstreaks (WS). Furthermore, data quality assurance is meticulously conducted through a two-stage process. Initially, the raw data undergo transformation through the application of the European Space Agency’s developed nominal calibration method, resulting in the generation of the normalized radar cross section (NRCS). Secondly, the NRCS executes recalibration, employing the CMOD5n model function [51] to mitigate the impact of incidence angle variations.

This dataset is generously supplied by the French Research Institute for the Exploitation of the Sea (IFREMER). It is available at http://www.seanoe.org/data/00456/56796/ (accessed on 7 February 2024). Figure 3 displays the distribution of the TenGeoP-SARwv dataset in a 5° × 5° global spatial grid. Although this SAR dataset includes the majority of WV acquisition such as the Pacific, Indian, and South Atlantic oceans, the SAR image density in each region is quite low, considering that the largest number of WV images is less than 40. The task of recognizing multiple oceanic phenomena can be easily hindered by this lack of high-quality data. To satisfy the requirement of training an effective classification model, these imagettes are chosen with the criteria that one geophysical phenomenon dominates with its special characteristics. It is worth mentioning that the format of SAR images is important for visual interpretation and dynamic detectability.

2.3. Data Processing

In order to avoid overfitting in deep neural networks for oceanic phenomenon prediction in the ocean, we expand the geographic scope with the range of the operational maritime domain. In the TenGeoP-SARwv dataset of Figure 2a–j provided by the ESA, these ten grayscale images collectively encompass an extensive span of the ocean’s surface, encapsulating a diverse array of features. Employing despeckling techniques such as a Lee filter, enhanced Lee filter, and supervised denoising method [20,52], these images strive to achieve an optimal equilibrium between mitigating speckle noise and retaining intricate feature details. However, it is imperative to acknowledge that the pixel intensity within these images inherently exhibits a limited range, a characteristic attributed to the underlying principles of SAR imaging system theory. This means that their features are not readily apparent because of low contrast. All original global data records are mixed in a disorderly manner. Thus, firstly, we create nested loops and conditionals to iterate over all records and judge whether the oceanic phenomenon is within the above geographic region. Each oceanic phenomenon is identified according to a specific signature or pattern. Due to the complicated nature of the C-band radar scatter response from the sea surface, it is predominantly contingent upon the incidence angle and the relative azimuthal angle between the radar system and the directional vector of the surface wind. In certain atmospheric circumstances, notably under conditions characterized by elevated wind speeds exceeding 15 m/s, the backscatter phenomenon is influenced by the combined effects of winds and oceanic wave states. As a result, these mentioned ten phenomena other than ocean waves (i.e., wind wave, swell, surface wave) may be inadequately represented or captured in the dataset.

SAR images exhibit inherent statistical characteristics, and a majority of statistical models employed in their analysis have originated from the multiplicative noise model. Hence, these imagettes are characterized by the presence of multiplicative noise, which leads to variability and interference in the form of undesired fluctuations, thereby affecting the fidelity of the acquired data [53]. It is obvious that the varied noise causes rapid changes in image grayscale and serious speckle in the bright regions. To completely develop SAR image speckle, we construct a noise model and image intensity definition based on coherent radiation as follows:

$$\begin{array}{cc}\hfill & N\left(t\right)=B\left(t\right)\ast s\left(t\right),\hfill \end{array}$$

(1)

$$\begin{array}{cc}\hfill & C(i,j)=x(i,j)\ast n(i,j).\hfill \end{array}$$

(2)

In Equation (1), the variable $N\left(t\right)$ denotes the signal with noise embedded into it, while $B\left(t\right)$ represents the radar backscatter property without noise for ground targets. The selection of an optimal distribution for each of these random variables relies on the types of ground targets. The variable $s\left(t\right)$ denotes speckle noise, which is independent of $B\left(t\right)$. During the modeling process, we opt to model SAR image speckle noise as multiplicative noise with a mean of 1. This modeling principle ensures that the multiplicative noise does not introduce an overall bias, preserving the average brightness and intensity of the image unaffected by speckle noise. This is advantageous for various image processing tasks, allowing us to assume that the average reflection intensity remains unaltered during image processing. Thus, in Equation (1), the mean value of $s\left(t\right)$ is one and its variance is relevant to the equivalent number of imagettes. In addition, in consideration of the multiplicative noise model, we define the intensity of an image with speckle noise by Equation (2). $C(i,j)$ stands for the intensity of SAR image pixels with the coordinate $(i,j)$. The variable $x(i,j)$ represents the intensity of SAR image pixels without speckle noise, and $n(i,j)$ represents the intensity of speckle noise.

To eliminate speckle noise in SAR images, a Lee filter is designed for preserving edges and point features in speckle noise reduction [54]. In a Lee filter, a variable-sized window is moving consistently from left to right and top to bottom across the SAR image to compute the local statistic. To sustain the characteristics of oceanic phenomena, our Lee filter is based on a linear speckle noise model and minimum mean square error methods to enhance the data quality as follows:

$$\widehat{Y}\left(t\right)=\overline{I}\left(t\right)+W\left(t\right)\left[I\left(t\right)-\overline{I}\left(t\right)\right],$$

(3)

where $\overline{I}\left(t\right)$ stands for the mean value of the intensity within the filter window; $W\left(t\right)$ represents the adaptive filter coefficient defined as Equation (4):

$$W\left(t\right)=1-\frac{C_v}{C_I}.$$

(4)

The variable $\widehat{Y}\left(t\right)$ represents the intensity of the filtered SAR images in Equation (3). $I\left(t\right)$ is the intensity of global pixels without noise. Typically, the value of $W\left(t\right)$ tends to zero in homogeneous regions, yielding analogous outcomes to that achieved by the mean filter. Meanwhile, the value of $W\left(t\right)$ tends towards edges, leading to pixel alteration in close proximity to these edges. The impact of the Lee filter is clearly exhibited in Figure 1.

To improve the overall visual quality of imagettes, we apply global histogram equalization in image processing by redistributing the intensity values of the pixels in an image [55], aiming to maximize the full dynamic range available. This method involves computing the cumulative distribution function of the image’s pixel intensities and then adjusting the intensity values based on this cumulative distribution. In our data processing, we first calculate a probability density function (PDF) of the pixel intensity of an image, as defined in Equation (5):

$$p\left(X_k\right)=\frac{n_k}{N},fork=0,1,\dots ,L-1.$$

(5)

The variable $n_k$ stands for the number of image pixels with an intensity that equals $X_k$. The variable N represents the total number of image pixels. To ensure that specific intensity values are assigned to each pixel in our imagettes, we apply a ceiling operation to round the intensity values to the nearest integer. This rounding operation ensures that each pixel is associated with a discrete and unique integer intensity value, contributing to the creation of a PDF based on these rounded values. For instance, the rounding method converts the continuous intensity values (e.g., 0.831, 0.832, 0.842) into discrete integer values (e.g., 1, 1, 1), thereby creating a set of unique and distinguishable intensity values for each pixel. If the intensity values of pixels are rounded up to integers, it is possible that the size of the resulting PDF could be less than the size of the original pixels. Then, we add each PDF element and achieve the cumulative density values by the cumulative density function (CDF). This function is defined as follows:

$$c\left(X_k\right)=\sum _{j=0}^{k}p\left(X_j\right)fork=0,1,\dots ,L-1.$$

(6)

To further enhance the overall brightness and contrast of imagettes, we map the original image intensity into the range [0,$x_{max}$] by the transform function:

$$f\left(x\right)=X_{max}\ast c\left(x\right),$$

(7)

where x is the intensity of image pixels. We apply the intensity transformation function uniformly across all pixels to facilitate the mapping of the original image into a new representation characterized by the desired contrast.

Our dataset comprises TenGeoP-SARwv images characterized by low contrast, frequently manifesting as dark imagery. To enhance the performance of subsequent feature extraction, we employ the aforementioned algorithms to generate brighter images.

3. MCNN Framework for Geophysical Phenomenon Classification

In this section, we introduce a novel neural network, referred to as the multichannel neural network (MCNN), which integrates the multichannel features derived from the original input for the purpose of classifying the aforementioned ten predefined geophysical phenomena shown in Figure 2. The comprehensive architecture of the MCNN algorithm, encompassing multi-feature fusion, data augmentation, and multichannel feature extraction, is visually depicted in Figure 4.

The multi-feature fusion module incorporates an efficient image edge filter to selectively preserve intricate details pertaining to oceanic phenomena present in SAR images, concurrently mitigating the disruptive effects of speckle noise. Hence, the multichannel characteristics of the input are integrated with the original SAR imagettes, along with the associated gradient magnitude derived from the applied filter. And then, feature augmentation such as horizontal translation and reflection is implemented to address the challenge of limited sample size. This approach aims to mitigate the small-sample problem by expanding both the volume and diversity of the training dataset. Furthermore, the feature extraction module utilizes a multichannel approach employing transfer-learning techniques to address constraints imposed by limited dataset sizes. Specifically, the Inception-ResNet architecture [56] is strategically chosen to strike a favorable balance between a substantial parameter count and optimal classification performance. Finally, some important weights are optimized with our softmax function to implement optimal parameter decision for oceanic phenomenon classification. The subsequent sections comprehensively detail the different parts of our MCNN algorithm.

3.1. Multi-Feature Fusion

For each original input SAR image, the MCNN algorithm requires resizing the corresponding dimensions to 299 pixels for both height and width. Nevertheless, in the direct implementation of oceanic phenomenon classification, there is a potential for the loss of certain meaningful features and the attenuation of information that is considered less pertinent [57]. Hence, the an image edge filter is implemented to retain the critical structural attributes inherent in SAR imagettes.

Table 1. Description of four different image edge filters [58].

No.	Edge Filter	Description (x and y Denote the Horizontal and Vertical Directions, Respectively)
1	Central	Weighted difference of neighboring pixels: $\frac{dI}{dy}=\frac{I(y+1)-I(y-1)}{2}$ and $\frac{dI}{dx}=\frac{I(x+1)-I(x-1)}{2}$
2	Prewitt	Convolution mask: $\mathsf{\Delta }y=\left[\begin{array}{ccc}1& 1& 1\\ 0& 0& 0\\ -1& -1& -1\end{array}\right] \mathrm{and}\mathsf{\Delta }x=\left[\begin{array}{ccc}-1& 0& 1\\ -1& 0& 1\\ -1& 0& 1\end{array}\right]$
3	Roberts	Weighted difference between diagonally adjacent pixels: ${\mathsf{\Delta }}_1=\left[\begin{array}{cc}0& 1\\ -1& 0\end{array}\right]$
4	Sobel	Convolution mask: $\mathsf{\Delta }y=\left[\begin{array}{ccc}1& 2& 1\\ 0& 0& 0\\ -1& -2& -1\end{array}\right] \mathrm{and}\mathsf{\Delta }x=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]$

lists the gradient magnitudes of different image edge filters with specified operators. Then the resultant and the initial are concatenated to form the input image channels. Experiments are conducted on the TenGeoP-SARwv dataset to determine the optimal channel combination for oceanic phenomenon classification. The empirical findings indicate that the optimal performance can be achieved when utilizing a multichannel input image reconstructed from both the original image and the central difference gradient. When employing the channel set combined only with the gradient magnitude, the classification accuracy sharply decreases. This can be attributed to the redundancy contained in the gradient magnitude. This does not contribute significantly to the enhancement of classification accuracy. Instead, information pertaining to interference is utilized as an alternative approach. Hence, the input image in this study is resized to the dimensions of $299\times 299\times 3$, incorporating three channels with two original SAR images and a central difference gradient.

The combination of multiple SAR images facilitates the creation of an intricate oceanic scene representation, consequently enhancing the precision of decision-making in subsequent tasks. Nonetheless, the efficacy of image fusion critically depends on the stringent geometric alignment of imagettes. Hence, we design and implement the feature-based image registration method to establish an accurate correlation between imagettes and assess the spatial transformation accordingly. In our design, we assume that the spatial resolution of both images is known and remains constant in our scenario. Because the pixel-by-pixel computations and variable resolution correlation can result in a time-intensive implementation, we initially conduct a normalized cross-correlation between the images at a reduced resolution, aiming to approximately delineate the common area of overlap between the two images. High-resolution correlation is subsequently assessed in close proximity to each potential registration point. The variables $I{M}_1$ and $I{M}_2$ denote two SAR images with both images having dimensions $M\times N$ and f represents the image zoom factor. Let $I{M}_{l1}$ and $I{M}_{l2}$ denote the low-resolution versions of the two SAR images with dimensions $Ml\times Nl$. The edged images of the corresponding images are represented by $I{M}_{1e},I{M}_{2e},I{M}_{l1e},I{M}_{l2e}$, respectively. In this context, the generalized cross-correlation can be mathematically expressed as follows:

$${\varphi }_1(u,v)=\frac{{\sum }_{i=0}^{Ml}{\sum }_{j=0}^{Nl}(I{M}_{l1eu,v}(i,j)-I{M}_{l2e}(i,j))}{{\sigma }_{1lu,v}{\sigma }_{2l}},$$

(8)

$$\varphi (k,l)=\frac{{\sum }_{i=0}^M{\sum }_{j=0}^N(I{M}_{1ek,l}(i,j)-I{M}_{2e}(i,j))}{{\sigma }_{1k,l}{\sigma }_2},$$

(9)

where the variable ${\varphi }_1(u,v)$ represents the cross-correlation coefficient derived from the low-resolution image pair $I{M}_{l1e}$ and $I{M}_{l2e}$. The pair $(u,v)$ denotes the coordinate index of the image $I{M}_{l1e}$. The image $I{M}_{l1eu,v}$ is a sub-image situated at the $(u,v)th$ position within the image $I{M}_{l2e}$, with dimensions identical to those of image $I{M}_{l2}$. ${\sigma }_{1lu,v}$ and ${\sigma }_{2l}$ stand for the standard deviations of the respective images. Similarly, $\varphi (k,l)$ corresponds to the cross-correlation coefficient calculated from the low-resolution image pair $I{M}_{1e}$ and $I{M}_{2e}$, where $(k,l)$ signifies the coordinate index of the image $I{M}_{l1e}$. The $I{M}_{1ek,l}$ image is an image segment located at the $(k,l)th$ pixel of $I{M}_2$ with dimensions matching those of $I{M}_{2e}$. ${\sigma }_{1k,l}$ and ${\sigma }_2$ denote the standard deviations of the corresponding images.

In the context of oceanic phenomenon detection, the extraction of edge information is significant for capturing structural features [59]. The basic idea of the image edge filter is to detect changes in image intensity by using a discrete differentiation operator [60]. To preserve more useful oceanic characteristics in SAR imagettes, we modify the standard Sobel operator to enhance the correlation between two fused SAR images [61]. In this study, we utilize four new operators instead of the standard two operators for retaining global edges and disregarding smaller edges as follows (Table 1):

$$S_1={\left(\mathsf{\Delta }y\right)}^T=\left[\begin{array}{ccc}1& 0& -1\\ 2& 0& -2\\ 1& 0& -1\end{array}\right],S_2={\left(S_1\right)}^T=\mathsf{\Delta }y=\left[\begin{array}{ccc}1& 2& 1\\ 0& 0& 0\\ -1& -2& -1\end{array}\right],$$

(10)

$$S_3=\left[\begin{array}{ccc}2& 1& 0\\ 1& 0& -1\\ 0& -1& -2\end{array}\right],S_4=\left[\begin{array}{ccc}0& 1& 2\\ -1& 0& 1\\ -2& -1& 0\end{array}\right].$$

(11)

By applying convolutionally the four operators into the original images, we can obtain:

$$E_l(i,j)=\sum _{m=1}^3\sum _{n=1}^{3}IM(i+m-1,j+n-1)\ast S_k(m,n)k=1,2,3,4$$

(12)

$$E(i,j)=\underset{l}{max}\left(E_l(i,j)\right).$$

(13)

In Equation (13), $E(i,j)$ stands for the appropriate edge of the point $(i,j)$ in the whole image $IM$. Hence, a better correlation of the two fused images can be achieved, in contrast to the correlation between them generated from standard Sobel operators or other edge detection methods.

3.2. Data Augmentation

A sufficient amount of training data is crucial to obtain state-of-the-art results in image classification tasks based on the deep learning method. Unfortunately, many application scenarios, such as oceanography and medicine [62,63], cannot assure enough samples for algorithm training purposes. Data augmentation has been proved to be an effective way not only to expand the quantity, but also to improve the quality of the training progress by diversifying the dataset. In this study, we address the issue of limited sample size in the TenGeoP-SARwv dataset for ocean phenomenon detection by employing image warping and oversampling techniques. When utilizing SAR images for training models, we encountered a challenge due to the scarcity of samples. The shortage of training samples poses obstacles to the model in capturing robust feature representations, ultimately impacting its ability to generalize effectively. This limitation may lead to suboptimal predictive performance, especially for categories with insufficient sample representation.

In this study, we tackle the challenges of limited samples and oversampling, both critical aspects for effective model training. Abundant samples are essential for robust model training and mitigating the risks of overfitting. To enhance dataset diversity, we employ data warping techniques initially. Data warping refers to altering existing samples; typical operations include geometric transformations, random cutting, and other alternations. This type of method is straightforward and effective and a random cropping algorithm is designed for this investigation. Parameters such as the maximum crop size and position are first defined as constraints. Then, the starting point and destination point are randomly generated to reconstruct a new sample. In the majority of interactive warping systems, the researchers specify the warp parameters in a broadly defined manner or indicate a point-to-point correspondence. Then, the automated system conducts geometric interpolation procedures to generate a cartographic representation based on the specified geometric parameters with a mapping $M_s$: ${\mathrm{D}}^2\to {\mathrm{D}}^2$ of the plane to itself. We apply the mapping to our input SAR images $I(x,y)$ and define the output images as $O(x,y)$ = $I\left(M_s^{-1}(x,y)\right)$, where $M_s^{-1}$ is the function inverse of $M_s$. Note that $M_s$ does not depend on the input I and the geometric specification S is based on SAR image features. To obtain $M_s$ from the specification S, we compute the unique thin-plate spline satisfying $M_s$ with the solution u as follows:

$$arg\underset{u(x,y)}{min}\underset{D^2}{\int \int }(u_{xx}^2+2u_{xy}^2+u_{yy}^2)dxdy,$$

(14)

where $u_{xx}$, $u_{xy}$, $u_{yy}$ represent $\frac{{\partial }^{2}u}{\partial x^2}$, $\frac{\partial u}{\partial x}\frac{\partial u}{\partial y}$, $\frac{{\partial }^{2}u}{\partial y^2}$, respectively. In fact, $M_s$ is independent of the input parameters I, causing suboptimal outcomes. The mapping function ensures the functional minimization across the specified domain. The mapping procedure serves as an intermediate component and then is applied to image I after iterative computation. To enhance the robustness and separability of SAR images, we take contemporary warping procedures with a pair of bivariate real functions as $M_s=(M_s^X,M_s^Y)$ independently. When invoking the thin-plate regularization paradigm, the procedure divides the interpolation constraints S into x-constraints and y-constraints and calculates the two mapping functions $M_s^X$ and $M_s^X$ to reduce the warping function of Equation (14). To further squeeze the edges of SAR images when stretching other areas, we use the following image warp $M_{S,I}(x,y)$ to optimize the image quality:

$$\underset{u(x,y)}{min}\underset{{[0,1]}^2}{\int \int }[u_x^2+u_y^2+\mu {(u(x,y)-m(x,y))}^2]\ast \varphi (x,y)dxdy.$$

(15)

For the x-component, we replace $m(x,y)$ with $mx$ and set $\varphi (x,y)=1+\lambda E_x(x,y)$. For the y-component, we have $m(x,y)=my$ and $\varphi (x,y)=1+\lambda E_y(x,y)$. $E_x(x,y)$ and $E_y(x,y)$ are the implementation of the x- and y-components, respectively, in the Canny edge detector [59].

To adjust the sample distribution of all classes, we apply the oversampling technique. Additionally, oversampling approaches generate synthetic instances to further benefit this task, whose typical techniques include image mixing, feature augmentations, and deep learning-based methods like generative adversarial networks (GANs) [63]. For the sake of diversity, we choose GANs as our data oversampling method. This story for data augmentation follows a typical neural network development process of data preparation, training, and generating new samples. To perform oversampling on a SAR dataset, an initial step involves training a generative adversarial network (GAN) to estimate the underlying distribution of oceanic data. Subsequently, the trained generator component of the GAN is employed to generate supplementary samples specifically focused on augmenting the representation of the minority class. In this design, we use a GAN to achieve the conditional distribution $Py|x$. The GAN facilitates the training process by enabling robust discriminators (D) that yield informative gradients to the generator, even in scenarios where the quality of the generated samples is suboptimal. This characteristic enhances the training stability, contributing to more effective and reliable model convergence. To further optimize the GAN’s loss function for categorical variables on GAN-based oversampling, we adopt the Wasserstein-1 distance (WD) by modifying the GAN objective to the following equation:

$$\underset{G}{min}\underset{D}{max}{\mathbb{E}}_{x\sim p_{data}}\left[D\left(x\right)\right]-{\mathbb{E}}_{y\sim p_y}\left[D\left(G\left(y\right)\right)\right]-\lambda {\mathbb{E}}_{\widehat{x}\sim p_{\widehat{x}}}\left[\right({\Vert {\nabla }_{\widehat{x}}D\left(\widehat{x}\right)\Vert }_2-1{)}^2].$$

(16)

In this model, the generator G produces synthetic data samples that are indiscernible from authentic data and the discriminator D differentiates authentic instances from the training dataset and synthetic instances generated by the generative model G. The variable $\lambda$ is the penalty coefficient. D is a k-Lipschitz function, and this Lipschitz continuity condition is ensured through the process of clipping [64], where the weights of D are constrained to a bounded interval [−c, c]. The gradients are computed based on linear interpolations $\widehat{x}\sim p_{\widehat{x}}$ between actual and synthetic samples, where $p_{\widehat{x}}$ represents the probability distribution of these linear interpolations.

3.3. Multichannel Feature Extraction

Despite previous endeavors to augment the dataset, the imperative remains to devise an innovative feature extraction architecture. While employing SAR images for identifying ocean phenomena, we confronted the challenge of the model lacking robustness. The model proved sensitive to the impact of noise and abnormal samples. Given the constraints of the constructed dataset with limited original training data, the construction of a medium-sized network, such as Inception-ResNet, becomes indispensable to fulfill the demands of multichannel feature extraction.

To capture the structural characteristics of ocean SAR imagettes, we implement the Inception-ResNet architecture, which is widely recognized for its intricate multi-branch design in Figure 5. As depicted in Figure 5, the data first pass through a stem network composed of pooling layers and convolutional layers. Then, they undergo dimension reduction in the network’s Inception-Resnet and reduction parts, which serve to reduce the spatial dimensions.

After feature extraction achieved by the aforementioned structure, the output section maps the one-dimensional vector to the predicted results. The stem configuration assumes a critical function in both feature extraction and the consequential reduction in spatial dimensions in the input. Within this network, a series of convolutional layers and filters are integrated through concatenation, indicating the network’s capacity for complicated feature representation. Inception-ResNet architectures denoted as A, B, and C in Figure 5 consist of both pure and residual inception blocks. The pure configuration, characterized by a split–transform–merge architecture, endows a potent representational capability owing to the presence of dense layers.

To expedite gradient propagation through multiple layers, we leverage the residual function, renowned for its efficacy in facilitating information flow and accelerating gradient propagation through layers. It achieves this by approximating a mapping function from the input:

$$y=F\left(x\right)+x.$$

(17)

Following the establishment of the MCNN architecture, it becomes imperative to specify the loss function, quantifying the disparity between the model predictions and the ground truth associated with the input data. In order to mitigate potential class imbalance issues inherent in the classification task, a weighted balanced cross-entropy loss function, as proposed by [21], is adopted for training purposes in Equation (18):

$$L_x(y,\widehat{y})=-\frac{1}{M}\left(\sum _{i=1}^M\left(\sum _{j=1}^C{\theta }_j{y}_{ij}log{\widehat{y}}_{ij}\right)\right).$$

(18)

Let y and $\widehat{y}$ represent the true label and the label predicted by the neural network, respectively, on the training dataset X sampled from the TenGeoP-SARwv database. For the true label, we assign numerical values to categorical classes (i.e., AF → 0, BS → 1, IB → 2, LWA → 3, MCC → 4, OF → 5, PW → 6, RC → 7, SI → 8, WS → 9, shown in Figure 2). For the predicted label, the whole MCNN model takes input data and produces output predictions between 0 and 9. Here, M signifies the total number of samples in X, and C denotes the number of distinct oceanic phenomena types. Additionally, ${\theta }_j$ denotes the weight assigned to the samples belonging to the j-th class. The computation of a loss value for the entire set X is a computationally intensive task. To address this challenge, a transfer-learning strategy is implemented in this study. Specifically, half of the layers within the Inception-ResNet architecture are preserved in a frozen state. Subsequently, the parameters ${\theta }_j^t$ of the unfrozen layers in Equation (19) and the ultimate classifier layer undergo iterative optimization using the gradient descent method, taking into account the Adam optimization algorithm, as described in [65]:

$${\theta }_j^t={\theta }_j^{t-1}-\frac{a\widehat{m_t}}{\sqrt{\widehat{v_t}}+\epsilon }.$$

(19)

Here, the terms $\widehat{m_t}$ and $\widehat{v_t}$ represent the bias-corrected first and second moments, respectively, and are iteratively updated according to the following expressions (20):

$$\widehat{m_t}=\frac{{\beta }_1{m}_{t-1}+(1-{\beta }_1)g_t}{1-{\beta }_1}.$$

(20)

Both ${\beta }_1$ and ${\beta }_2$ represent the exponential decay rates associated with moment estimates. The gradient $g_t$ can be calculated by using the stochastic objective function $f\left(\theta \right)$:

$$g_t={\nabla }_{\theta }{f}_t\left({\theta }_j^{t-1}\right).$$

(21)

Appropriate parameters for the classification of ocean SAR imagettes have been determined through a series of experiments, resulting in the optimal settings of $\epsilon ={10}^{-8}$, $\alpha ={10}^{-3}$, ${\beta }_1=0.9$, and ${\beta }_2=0.999$. The choice of employing the Adam optimization algorithm is grounded in its efficacy in handling nonstationary objectives and effectively addressing challenges associated with highly noisy and sparse gradients commonly encountered in the context of ocean SAR imagery.

Following the extraction of multichannel features using the Inception-ResNet architecture, feature concentration is performed to derive classification outcomes. In this investigation, the utilization of the softmax function is employed, leveraging its capacity to acquire discriminative yet generative compact vector representations for image classification, as discussed in prior work [66]. Let there be C classes in the assessment dataset. Denoting the original image input as K and the MCNN parameter as $\theta$, the softmax function is applied to ascertain the probability of y belonging to the k-th class:

$$P(y_k=1|K,\theta )=\frac{e^{K^T{\theta }_k}}{{\sum }_{j=1}^C{e}^{K^T{\theta }_j}}.$$

(22)

Here, $P(y_k=1|K,\theta )$ represents the probability of the image belonging to the k-th class, e denotes the exponential function, and ${\theta }_k$ signifies the parameters corresponding to the k-th class. This formulation allows for the derivation of a probability distribution over the multiple classes for effective image classification. For inputs, denoted as K, not falling within the confines of any predefined classes ($P(y_k=1|K,\theta )<0.5,k=1,2,\dots ,C$), a categorization process is implemented to assign such inputs to a distinct and specific category.

4. Experimental Evaluation

To assess the performance of our MCNN method, we choose AlexNet [42] as the baseline due to its noteworthy achievements in image classification tasks. Additionally, for implementing comparative analysis, we incorporate another popular method, CMwv [48], to evaluate the effectiveness of the MCNN technique. In order to ensure an equitable comparison between MCNN and the two alternative CNN-based methodologies, as delineated in Section 4.1, we adopt identical training methodologies, following the protocols introduced by Wang et al. [48]. A subset of 320 images per class is randomly selected as input from the annotated TenGeoP-SARwv dataset, categorized by WV1 and WV2 modes. It is pertinent to acknowledge that the input size is relatively modest in comparison to the overall dataset. Seventy percent (70%) and thirty percent (30%) of the input dataset are designated for the training and validation subsets, respectively, in accordance with the previously outlined methodologies for feature extraction and neural network weight optimization. A distinct evaluation dataset, consisting of 5000 vignettes captured under the WV1 and WV2 modes, has been established to quantitatively evaluate the effect of different classification models. Subsequently, a geophysical application is exemplified through the utilization of a classified map depicting rain cells (RCs). This map is then juxtaposed with precipitation data obtained from the Global Precipitation Measurement (GPM) Program (available on the NASA data archive website: https://pmm.nasa.gov/data-access/downloads/gpm (accessed on 7 February 2024)).

4.1. Experimental Results of Ocean SAR Scene Classification

The TenGeoP-SARwv dataset has been employed for training our MCNN algorithm through a series of iterations exceeding 2000. As depicted in Figure 6, the comprehensive accuracy and loss metrics exhibit a notable and rapid evolution within the initial 270 iterations, stabilizing thereafter around the 1000th iteration. Consequently, the neural network weights derived from the training process up to 1100 iterations are incorporated into the ultimate configuration of the MCNN model.

Figure 7 illustrates the normalized confusion matrix obtained from the MCNN model. The diagonal and off-diagonal elements of the matrix correspond to accurately and inaccurately classified observations, respectively, normalized with respect to the total number of observations. The percentages indicated in the blue cells within the row (column) summaries on the right (bottom) pertain to recall (precision) metrics. Additionally, F-score parameters, as defined by Sokolova et al. [67], are employed as evaluation metrics, offering a comprehensive assessment considering both precision and recall statistics. The anticipated values for these three parameters are all in proximity to unity.

Table 2. Classification performance in WV1 (denoted by expansion) and WV2 (denoted by abbreviation) imagette detection of the different methods (the highest values are indicated in bold).

Class	Classification Results of the Different Methods
	AlexNet			CMwv			MCNN
	Recall	Precision	F-Score	Recall	Precision	F-Score	Recall	Precision	F-Score
Atmospheric fronts	0.46	0.72	0.56	0.95	0.40	0.56	0.93	0.83	0.88
(AFs)	0.13	0.81	0.22	0.95	0.38	0.54	0.92	0.91	0.91
Biological slicks	0.97	0.67	0.79	0.95	0.88	0.91	0.89	0.99	0.94
(BSs)	0.63	0.7	0.66	0.89	0.91	0.9	0.93	1	0.96
Icebergs	0.48	0.66	0.56	0.97	0.16	0.27	0.96	0.98	0.97
(IBs)	0.04	0.82	0.08	0.92	0.18	0.3	0.97	0.97	0.97
Low-wind areas	0.76	0.98	0.86	1	0.87	0.93	0.92	0.94	0.93
(LWAs)	0.74	0.78	0.76	1	0.79	0.88	0.96	0.95	0.95
Microconvective cells	0.38	0.88	0.53	0.8	0.76	0.78	0.98	0.98	0.98
(MCCs)	0.41	0.82	0.55	0.85	0.94	0.89	0.95	0.93	0.94
Oceanic fronts	0.41	0.88	0.56	1	0.05	0.1	0.91	0.92	0.91
(OFs)	0.67	0.48	0.56	1	0.05	0.1	0.95	0.87	0.91
Pure ocean waves	0.91	0.81	0.86	0.47	1	0.64	0.98	0.95	0.96
(PWs)	0.12	0.95	0.21	0.39	0.98	0.56	0.96	0.95	0.95
Rain cells	0.66	0.88	0.75	0.93	0.88	0.9	0.97	0.99	0.98
(RCs)	0.73	0.88	0.8	0.93	0.8	0.86	0.98	0.96	0.97
Sea ice	0.75	0.93	0.83	0.9	0.96	0.93	0.96	1	0.98
(SI)	0.85	0.74	0.79	0.96	0.96	0.96	0.99	0.98	0.98
Windstreaks	0.53	0.72	0.61	0.83	0.77	0.8	0.99	0.98	0.98
(WSs)	0.79	0.58	0.67	0.83	0.96	0.89	0.83	1	0.91
Average	0.63	0.81	0.71	0.88	0.67	0.76	0.95	0.96	0.95
	0.51	0.76	0.61	0.87	0.7	0.78	0.94	0.95	0.95

provides a comparative analysis of distinct classification models based on metrics such as recall, precision, and F-score. The interpretation of the table is as follows:

(1)

Despite the established robustness of AlexNet in image classification, its performance in ocean SAR scene classification falls short of acceptability without the incorporation of sophisticated domain knowledge.

(2)

The experimental results evaluate the effectiveness of the CMwv model in accurately identifying biological slicks (BS), low-wind areas (LWA), microconvective cells (MCC), rain cells (RC), sea ice (SI), and windstreaks (WS). However, suboptimal results are evident when employing the CMwv model to detect other geophysical phenomena, with a notable instance being the lowest precision value recorded at 0.05 for the oceanic front (OF) class.

(3)

The MCNN algorithm demonstrates superior performance in terms of recall, precision, and F-score, all surpassing 0.9 for both WV1 and WV2 imagettes, with the exception of the LWA class. Despite the limited number of samples available for training, the MCNN algorithm leverages data augmentation and transfer-learning strategies effectively, mitigating the risk of model overfitting. Furthermore, the utilization of an image edge filter optimally encodes discriminant information present in SAR imagettes. On average, the MCNN algorithm yields recall, precision, and F-score values that are 7.9% (8.0%), 43.2% (35.7%), and 25.0% (21.7%) higher, respectively, compared to CMwv when applied to WV1 (WV2) imagette classification. These results signify the potential of the proposed CNN-based model as a viable operational tool for automated ocean SAR scene classification.

4.2. Geophysical Application

To further validate the robustness of our MCNN algorithm, this classification model was systematically applied to all SAR imagettes spanning the period from March 2016 to February 2017 within the TenGeoP-SARwv dataset. Subsequently, a geophysical map delineating detected RCs was generated and meticulously compared with precipitation data sourced from the GPM Program. Within the context of S-1 imagettes, a classified RC was defined as comprising several kilometer-scale semicircular to circular-shaped patches, characteristic of rain downdrafts in convective RCs, as elaborated in [50]. Over the specified timeframe, encompassing March 2016 to February 2017, it was observed that approximately 5% of all SAR imagettes were successfully classified as RCs by the MCNN algorithm. The cartographic outcome of the classified RC occurrences, as depicted in the central panel of Figure 8, indicates discernible spatial patterns. The geophysical map presenting the yearly averaged GPM precipitation, along with the classified outcomes derived from the CMwv algorithm, is concurrently displayed in the right and left panels, respectively. It is imperative to acknowledge that the GPM product is designed to furnish gridded global multisatellite precipitation estimates, resulting in notable disparities in coverage resolutions between the SAR-classified RC occurrences and the GPM precipitation results. Despite observable distinctions in extratropical regions, areas characterized by heightened RC occurrences, as identified by the MCNN and CMwv algorithms, exhibit significant alignment with regions characterized by elevated GPM precipitation, particularly within the Intertropical Convergence Zone and South Pacific Convergence Zone. This concordance is consistent with the precipitation climatology outlined in prior studies [68,69].

5. Conclusions

To address the challenges associated with low efficiency in global ocean SAR scene classification, a novel robust classification model, denoted MCNN, that included multi-feature fusion, data augmentation, and multichannel feature extraction was successfully developed based on the Inception-ResNet CNN architecture. In a departure from conventional deep learning methodologies, a meticulously selected image edge filter was employed to distinguish distinct oceanic phenomena. Furthermore, several data augmentation and feature extraction techniques were implemented to mitigate the limitations posed by small-sample training datasets. The model performance was rigorously validated on a meticulously annotated dataset [70], focusing on the detection of ten geophysical phenomena prevalent on the ocean surface. Comparative analyses were conducted against classical and state-of-the-art methods. Additionally, our MCNN algorithm was applied to generate a geophysical map of RCs, demonstrating acceptable consistency with precipitation data obtained from the GPM Program. These results underscore the potential utility of the MCNN tool in obtaining global/regional and annual/seasonal statistics of diverse geophysical phenomena and enhancing specific aspects of numerical ocean models. While this study exclusively utilized S-1 WV SAR imagettes, the adaptability of the MCNN algorithm to other SAR data archives, such as GF-3 and TerraSAR-X, is acknowledged.

Despite the notable contributions of this study, it is essential to acknowledge certain limitations. The analysis is constrained to SAR images featuring ten common types of ocean phenomena, and there may be occurrences beyond the scope of these datasets that could influence the generalizability of the findings. Moreover, the effectiveness of the proposed methods may vary in the context of diverse datasets or under conditions not represented in the current study. As a result, the robustness and accuracy of the model may be limited when recognizing ocean phenomena beyond the scope of our dataset. These limitations emphasize the importance of future research to explore a more comprehensive range of datasets and scenarios for a thorough understanding of the model’s capabilities.

In future studies, our objective is to augment the dataset by including SAR images of additional ocean phenomena, thereby bolstering the model’s robustness. The expansion of the dataset will enable us to tackle challenges associated with imbalanced categories and enhance the model’s adaptability to new samples. This approach seeks to improve the model’s generalization capabilities and contribute to more effective and comprehensive recognition of diverse ocean phenomena.

This methodology holds potential for broader applications, such as weather prediction and hazard forecasting through improved classification rates. The enhanced performance of accurate recognition and classification for ocean phenomena in SAR images may contribute to more reliable scarce data inputs for weather prediction models, ultimately leading to increased precision in forecasting and hazard prediction.