1 Introduction

Ocean fronts are narrow transition zones separating water masses with different properties. Depending on the physical characteristics, ocean fronts can be classified into sea surface temperature (SST) fronts, sea surface density fronts, and sea surface salinity fronts. Taking SST fronts as an example, they typically represent a sharp boundary current and can persist for extended periods, separating materials and energy on both sides of the front (Mauzole 2022). Owing to the challenge particles face in crossing these fronts, the water masses on each side of an SST front display exhibit significantly different temperatures (Zheng et al. 2022).

The presence and dynamic evolution of ocean fronts can influence the biogeochemical and physical properties of the surrounding ocean areas by altering material and energy transport (Corredor-Acosta et al. 2020; Siegelman et al. 2020; Nicholson et al. 2022; Song et al. 2022; Zhou and Cheng 2022). Therefore, the exploration of ocean fronts holds significant implications for various fields, including marine ecology, oceanic climatology, marine fisheries, and maritime military operations (Pattiaratchi et al. 2022; Prants 2022; Salim et al. 2023). For instance, the influx of freshwater into the Alboran Basin from the Atlantic (with a salinity of about 36.5) occupies the southern part of the basin, while the northern region is dominated by saltier water (with salinity levels exceeding 38). The ocean front that forms between these two surface water masses, known as the North Balearic Front, plays a pivotal role in determining the physical and biogeochemical characteristics of the western Mediterranean. Therefore, accurately identifying and analyzing the spatiotemporal distribution and dynamic changes in ocean fronts from a large volume of remote sensing images holds significant importance.

In recent years, extensive research has been conducted on the variability and interaction of ocean processes, resulting in significant research breakthroughs (Strobach et al. 2022; Xie et al. 2022). For instance, Siegelman et al. (2020) employed advanced methodologies, including the deployment of seals as data-collecting instruments, for in situ observations of the ocean’s mixed layer and underlying regions. Their findings elucidated the widespread presence of ocean fronts within the ocean mixed layer, especially near the edges of eddies. Similarly, Su et al. (2020) and colleagues utilized high-resolution ocean models to demonstrate the pivotal role of ocean fronts in influencing the global ocean’s heat absorption through heat transfer mechanisms. Furthermore, Corredor-Acosta and his research team have shown that ocean fronts significantly enhance nutrient concentrations within the ocean’s euphotic layer, thereby fostering phytoplankton growth and affecting community dynamics (Corredor-Acosta et al. 2020). Nevertheless, it is worth noting that in eastern boundary upwelling systems and various other ocean regions, the underlying mechanisms are not fully understood owing to the absence of in situ observational data (Jian et al. 2019, 2021; Wang et al. 2021b; Luo et al. 2022, 2023).

Despite significant advancements in oceanographic research, experts often resort to manual analysis techniques to investigate the variability and interactions of ocean fronts. These conventional methodologies require the acquisition of pertinent data through specialized instruments and take months to years for comprehensive data processing, making the process both time-intensive and laborious. Nevertheless, there is a burgeoning interest in leveraging deep learning methodologies to explore the dynamics and interactions of ocean fronts using remote sensing data. This emerging field represents a nascent yet pivotal avenue within oceanographic research (Yang et al. 2017, 2022c). Currently, in the realm of remote sensing, deep learning techniques span diverse domains. These encompass detecting ocean eddies, change detection in remote sensing imagery, enhancing remote sensing images, classifying agricultural land, and predicting crop yields, among others (Dong et al. 2018; Yang et al. 2021, 2022b).

For example, Sun et al. (2020) and collaborators devised a deep learning framework integrating contextual semantics with sea surface height data from remote sensing to precisely delineate eddy boundaries. Shaffque et al. (2022) conducted a comparative analysis of prevalent deep learning methodologies for change detection in remote sensing imagery, analyzing their effectiveness across diverse change detection datasets. Wang et al. (2021a) introduced a novel multisensor optical remote sensing dataset and evaluated mainstream super-resolution techniques on this dataset. Yao et al. (2022) applied deep learning methodologies to crop classification using remote sensing imagery, introducing a framework that combines deep neural networks with a random forest classification approach. Their model can swiftly and accurately discern crop types across various spatial and temporal scales. Muruganantham et al. (2022) conducted a comprehensive review of the progress and limitations of using deep learning approaches in crop yield prediction leveraging remote sensing data over the preceding decade, elucidating various factors influencing yield prognostication. Gawlikowski et al. (2022) introduced distribution uncertainty principles into remote sensing research, aiming to mitigate challenges associated with labeling and training standardized remote sensing image data sourced from disparate sensors, categories, and geographic regions.

With the exponential proliferation of available data volumes, deep learning technology has experienced rapid advancements and proven effective in elucidating the variability and interactions of ocean fronts. A research team under the guidance of Prof. Junyu Dong leveraged physics-based methodologies for detecting ocean fronts to develop an ocean front recognition network based on an annotated dataset. This effort utilized convolutional neural networks for ocean front identification in sensing images, facilitated by transfer learning strategies (Lima et al. 2017). To gain deeper insights into the ocean front dynamics, the team introduced a novel ocean front tracking algorithm, employing the GoogLeNet Inception network (Yang et al. 2022a). Furthermore, in a bid to bolster the precision of ocean front detection and tracking, Sun et al. (2018) proposed a multiscale deep learning architecture that improves detection accuracy and refines localization of ocean fronts. Subsequently, Yang et al. (2022c) amalgamated prior domain knowledge and devised an evolutionary trend recognition network framework using deep learning, empowering the classification of ocean front evolution trends with discerning precision.

Nevertheless, the efficacy of deep learning technology heavily hinges on the abundance of trainable data (Ju et al. 2023a, b, c; Zhang et al. 2023b). Insufficient data can impede achieving satisfactory results (Ju et al. 2022; Yeung and Lam 2023; Zhang et al. 2023a, c). Despite the increase in SST remote sensing imagery owing to advancements in remote sensing technology, current deep learning methodologies have not yet leveraged these images for oceanographic research. For instance, studies on SST front variability and interactions using remote sensing imagery, such as front tracking and classification, are confined to specific regions and lack the capability to accurately track and classify the nuanced variability and interactions of individual ocean fronts (Yang et al. 2022a, c). A fundamental impediment in comprehending ocean front dynamics lies in the absence of methodologies for processing ocean front remote sensing data, notably the lack of a comprehensive database enabling direct exploration of the evolutionary patterns of individual ocean fronts.

To address this limitation, this paper introduces methodologies for constructing an ocean front dynamic evolution dataset, catering to the requisites for training deep learning networks. The structure of this paper is organized as follows: Section 2 introduces the remote sensing data used and outlines the dataset construction alongside a succinct exposition of the ocean front detection algorithm employed. Section 3 expounds upon the methodologies employed for dataset construction. Section 4 showcases samples selected from the ocean front dynamic evolution dataset. Finally, Section 5 summarizes the main points of the paper and suggests directions for future research.

2 Data

The research data in this study was obtained from an advanced very-high-resolution radiometer, with a data accuracy of 5 km. The data covers a specific timeframe from 2010 to 2020, covering a geographical span from 10\(^{\circ }\)N to 40\(^{\circ }\)N in latitude and from 112.5\(^{\circ }\)E to 135\(^{\circ }\)E in longitude. The study area belongs to the Kuroshio Current System, characterized by regular and frequent dynamic changes in SST.

2.1 Dataset introduction

This paper intends to construct an ocean front dynamic evolution database for the study of ocean front dynamic processes. This dataset consists of 1000 sample data, with each sample containing a sequence of ocean front images. These sequences document the dynamic evolution of selected ocean fronts, including phases of enhancement, attenuation, splitting, and merging. The observation of ocean fronts capturing phases of enhancement or attenuation includes the inevitable phenomena of splitting and merging. These processes significantly influence the intensification and diminution of ocean front systems. Therefore, recording both the enhancement/attenuation and the splitting/merging of ocean fronts simultaneously in the samples provides a more objective and comprehensive representation of the ocean front dynamic processes.

2.2 Ocean front detection method

This project employs the microcanonical multiscale fractal model (MMF) to detect ocean fronts (Yang et al. 2016). Originally a significant breakthrough in signal processing, MMF has been successfully applied areas across various fields, such as signal compression, inference, prediction, satellite imaging, computer graphics, and natural image processing. By precisely computing the local parameters of each moment in the signal domain, referred to as singular exponents (SE), MMF allows for the study of the local geometric statistical properties of complex signals from a multiscale perspective. SE quantifies the signal regularity at each moment and provides valuable information about the local dynamics of complex signals. The SE factor associated with a multiscale function can be determined through an evaluation of power-law scaling behavior, utilizing a function that observes the power-law distribution at a fine scale.

$$\begin{aligned} h(\overrightarrow{x}) = \frac{\frac{\log (\tau _\psi \mu (\overrightarrow{x},r_0))}{<\tau _\psi \mu (.,r_0)>}}{\log r_0} + o\left( \frac{1}{\log r_0}\right) , \end{aligned}$$
(1)

where \(r_0\) is used for image normalization, and the SE factor \(h(\overrightarrow{x})\) is scale-independent. For small-scale behaviors, the additive term \(o(\frac{1}{\log r_0})\) is negligible compared to the factor \(\frac{\frac{\log (\tau _\psi \mu (\overrightarrow{x},r_0))}{<\tau _\psi \mu (.,r_0)>}}{\log r_0}\). To address this long-distance correlation, filtering is necessary. However, if a fixed scale is set in the filter, it may restrict the range of effective scales, or even worse, obscure the values of scaling exponents. Continuous wavelet transform \(\tau _\psi \mu (x,r_0)\) is defined as the projection of the signal at different resolutions and is suitable for filtering multiscale functions without breaking scale invariance. Presuming the scale invariance of the multiplicative factor, the estimation formula for the most singular values \(F_\infty\) is delineated as follows:

$$\begin{aligned} F_\infty = \overrightarrow{x}:h(\overrightarrow{x})=h_\infty =\max (h(\overrightarrow{x})). \end{aligned}$$
(2)

This calculation method for determining the most singular values is particularly suitable for ocean front detection applications. Points exhibiting the most singular values always superimpose singular behaviors at different scales, indicating that this algorithm can accurately reflect oceanic physical processes with multiscale characteristics, such as ocean fronts.

3 Dataset construction method

The overall approach for dataset construction mainly includes the following algorithms: ocean front segmentation, key ocean front extraction, ocean front variation classification, ocean front interaction classification, and dynamic ocean front tracking.

3.1 Ocean front segmentation method

Following the detection of ocean fronts, it is necessary to segment and isolate the ocean fronts of interest. The ocean front segmentation method is detailed in Algorithm 1. Our approach to ocean front segmentation includes the following steps: (1) removing coastal boundaries and islands from ocean front detection results, (2) removing noise from the resulting image, (3) using a morphological method to isolate overlapping ocean fronts, (4) assigning numbers to the segmented ocean fronts, and (5) restoring the position and value of the ocean fronts.

figure a

Algorithm 1 Ocean front segmentation method

The coastline mask M, for an ocean front image I, is generated through statistical analysis of the temperature variation patterns. The equation for the ocean front segmentation method can be simplified as follows:

$$\begin{aligned} Fronts(T(t)) = BW(M(T(t)),r), \end{aligned}$$
(3)

where \(M(T(t)) > T_\text{small}\), r is the minimum size of the ocean fronts. Given a radius r and an ocean front image I, the function BW(I, r) yields the ocean fronts amplified through the application of a morphological structuring element. The function BW(M(T(t)), r) delineates the bounding box of the ocean fronts. The employment of a morphological structuring element may cause changes in the position and value of the isolated ocean fronts Fronts(T(t)). Consequently, it is imperative to adjust the position and value of these isolated ocean fronts. The procedure for the examination and rectification of ocean fronts can be expressed as follows:

$$\begin{aligned} S(T(t)) = V(P(Fronts(T(t)))), \end{aligned}$$
(4)

where \(P(\cdot )\) restores the axes of the ocean fronts, \(V(\cdot )\) restores the value of the ocean fronts.

3.2 Key ocean front extraction method

From a macroscopic viewpoint, the intensity of ocean fronts exhibits regular fluctuations. However, at the micro-scale, specific geographic locations maintain perennial ocean fronts, whereas others experience ephemeral manifestations. Acknowledging the significance of identifying the prolonged presence of ocean fronts, our methodology for ocean front extraction is devised to segregate areas distinguished by the persistent occurrence of ocean fronts. The key ocean front extraction method is outlined in Algorithm 2. Our principal methodology for extracting key ocean fronts encompasses the subsequent steps: (1) conducting a statistical analysis to determine the frequency of ocean front occurrences at the pixel level, (2) identifying the most recurrent ocean fronts to determine the key ocean fronts, (3) employing morphological techniques to delineate the key ocean fronts.

figure b

Algorithm 2 Key ocean front extraction method

The calculation of the key ocean fronts Key(T) is executed as follows:

$$\begin{aligned} Key(T) = S(E(F(T)),p)), \end{aligned}$$
(5)

where Key(T) includes \(n_\text{key}\) key ocean fronts, T represents a time-series of ocean fronts. \(F(\cdot )\) represents the function that quantifies the frequency of ocean front occurrences, \(E(\cdot )\) denotes the operation that orders and extracts these occurrences by frequency, p is the extraction percentage, and \(S(\cdot )\) delineates the operation that segments key ocean fronts.

Fig. 1
figure 1

Percentage for extracting key ocean fronts: a p=5%; b p=10%; c p=20%; and d p=30%

Full size image

Our methodology for extracting key ocean fronts is implemented on time-series data of ocean fronts. Initially, images of ocean fronts undergo pre-processing to eliminate coastal boundaries. Subsequently, statistical analysis is used to determine the frequency of ocean front occurrences and identify the key ocean fronts. The outcomes of this extraction process are illustrated in Fig. 1a at \(5\%\), b at \(10\%\), c at \(20\%\), d at \(30\%\). The extraction percentage can be adjusted based on practical requirements. Expert experience indicates that the most commonly utilized percentage falls below \(30\%\).

3.3 Ocean front variation classification method

We juxtapose the ocean fronts at time step t with those at the preceding time step \(t-1\) and catalog the variation states of the ocean fronts. In this paper, the variation states of ocean fronts can be divided into two categories: enhancement and attenuation. The overall classification method is described in Algorithm 3.

Fig. 2
figure 2

Results of ocean front variation classification method: a ocean front attenuation time series and b ocean front enhancement time series

Full size image
figure c

Algorithm 3 Ocean front variation classification method

The calculation functions are simplified as follows:

$$\begin{aligned} R(T(t),T(t-1)) = V_{OI}(S(T(t)))/V_{OI}(S(T(t-1))), \end{aligned}$$
(6)
$$\begin{aligned} S_V(T,t) = {R(T(t),T(t-1))>1 : \text{Enhancement, attenuation}}, \end{aligned}$$
(7)

where \(R(T(t),T(t-1))\) gives the ratio between the ocean front size at time step t and that at time step \(t-1\), \(S_V(T,t)\) labels the variation state of the ocean fronts at time step t. As shown in Fig. 2, the displayed ocean front is one of the segmented ocean fronts and exhibits two variation trends according to the ocean front size. These trends are attenuation (Fig. 2a), and enhancement (Fig. 2b).

Fig. 3
figure 3

Key ocean fronts and their variation process: a example of an isolated key ocean front and b–i variation processes of selected key ocean fronts

Full size image

Given the ocean front variation classification algorithm, the key ocean front variation states can be tracked and recorded. As shown in Fig. 3, the key ocean front extraction percentage is set at \(5\%\), while Figs. 3b–i depicts the key ocean front variation process.

3.4 Ocean front interaction classification method

Ocean fronts lack a fixed shape and exhibit variations not only in size and shape but also undergo processes such as merging with adjacent fronts or dividing themselves. The interaction of ocean fronts in a time series can be divided into splitting and merging. As shown in Fig. 4a, the ocean fronts selected from an ocean front image at time step \(t-1\) contain two or three isolated ocean fronts. These isolated ocean fronts merged at time step t (Fig. 4b). The overall method for tracking ocean front interactions is detailed in Algorithm 4.

Fig. 4
figure 4

Example of the ocean front merge process: a ocean fronts detected from a remote sensing image before merging and b merged ocean fronts

Full size image
figure d

Algorithm 4 Ocean front interaction classification method

The calculation functions are simplified as follows:

$$\begin{aligned} C_{OI}(T(t),T(t-1)) = C(O(S(T(t)),S(T(t-1)))), \end{aligned}$$
(8)
$$\begin{aligned} C_{IO}(T(t),T(t+1)) = C(O(S(T(t)),S(T(t+1)))), \end{aligned}$$
(9)
$$\begin{aligned} S_I(T,t) = {C_{OI}(T(t),T(t-1))>1: \text{Merging}; \; C_{IO}(T(t),T(t+1))>1: \text{Splitting}}, \end{aligned}$$
(10)

where \(O(\cdot )\) classifies whether ocean front i at time step t overlaps with another one at time step \(t-1\). The function \(C(\cdot )\) counts the number of overlapping ocean fronts. \(C_{OI}(T(t),T(t-1))\) provides the count of ocean fronts that overlap at time step \(t-1\), while \(C_{IO}(T(t),T(t+1))\) indicates the number of overlapping ocean front at time step \(t+1\) and \(S_I(T,t)\) labels the interaction state of the ocean fronts at time step t.

Given a segmented ocean front i at time step t (Fig. 5a), we calculate its interaction result at time step \(t+1\). The interaction result \(S_I(T,t+1)\) for ocean front i is displayed at Fig. 5b. Then, we calculate the interaction of ocean front i at time step \(t+2\), and the interaction result, \(S_I(T,t+2)\), is displayed in Fig. 5c. The interactions of ocean front i from time step \(t+1\) to \(t+7\) are tracked, and the tracking results are displayed in Figs. 5b–h.

Fig. 5
figure 5

Ocean front interaction process: a example of an isolated ocean front and b–h interaction process at the subsequent ocean front time series

Full size image

3.5 Comprehensive ocean front dynamic evolution tracking method

The ocean front dynamic evolution tracking method is designed based on algorithms for key ocean front extraction, ocean front variation classification, and ocean front interaction classification. As illustrated in Algorithm 5, our method can track the key ocean front variations and label the ocean front interaction states simultaneously. The calculation functions are simplified as follows:

$$\begin{aligned} Tracking(T) = {S_V(O(Key(T),S(T)),t_1), S_I(O(Key(T),S(T)),t_1)}, 1 \ge t_1 \ge L, \end{aligned}$$
(11)
$$\begin{aligned} D_{com}(T) = {S_V(Tracking(T),t_2), S_I(Tracking(T),t_2)}, 1 \ge t_2 \ge L, \end{aligned}$$
(12)

where L is the length of the time series, while \(Tracking(\cdot )\) tracks the ocean front variation and interaction process throughout the entire time series T.

figure e

Algorithm 5 Complicated ocean front tracking method

Fig. 6
figure 6

Results of key ocean front variation process from time step t to time step \(t+7\) in key ocean fronts: a number 1; b number 4; c number 11; and d number 22

Full size image

4 Results

Illustrative representations of key ocean front variation process, ocean front interaction process, and comprehensive ocean front dynamic evolution process are presented in Figs. 6, 7, and 8, respectively. Following our methodology, we systematically enumerated all discernible isolated key ocean fronts. Subsequently, from this pool of identified fronts, a subset comprising four key ocean fronts was randomly chosen for sampling, specifically ocean fronts 2, 4, 5, and 22, as delineated in Fig. 6. Employing our designated key ocean front tracking method, we proceeded to track the selected ocean fronts across sequential time steps ranging from t to \(t+7\). The outcomes of this tracking endeavor are visually represented in the provided figures, with successfully tracked instances highlighted in yellow, while untracked fronts are depicted in blue.

Fig. 7
figure 7

Results of ocean front interaction process from time step t to time step \(t+7\). Interaction process of isolated ocean fronts a number 1; b number 4; and c number 22 at time step t

Full size image

As shown in Fig. 6a, ocean front 1 at time step t overlaps with the key ocean front. At time step \(t+1\), ocean front 1 still overlaps with the key ocean front and merges with ocean front 2. At time step \(t+2\), ocean front 1 continues to overlap with the key ocean front, and ocean fronts 1 and 4 at time step \(t+1\) merge to create ocean front 1 at time step \(t+2\). At time step \(t+3\), ocean front 1 still overlaps with the key ocean front, and ocean front 1 at time step \(t+2\) split into ocean fronts 1 and 2. In the following time steps, we track the key ocean front by calculating its overlap with other ocean fronts at those time steps. The key ocean front variation results shown in Figs. 6b–d follow the same principle.

In Fig. 7, three samples of the ocean front interaction processes are displayed. As shown in Fig. 7a, ocean front 1 at time step t merges with ocean front 2 at time step \(t+1\). At time step \(t+3\), ocean front 1 at time step \(t+2\) splits into ocean fronts 1 and 2. In the following time steps, we track the interactions between ocean fronts 1 and 2. At time step \(t+4\), ocean fronts 1 and 2 at time step \(t+3\) split into ocean fronts 1, 2, 5, and 6. In the following time steps, we track the interactions among ocean fronts number 1, 2, 5 and 6. At time step \(t+6\), ocean fronts 1, 2, and 14 at time step \(t+5\) merge into ocean front 1. In the following time steps, we track the interaction of ocean fronts 1 and 4. The ocean front interaction tracking results shown in Figs. 7b–c follow the same principle.

Fig. 8
figure 8

Results of key ocean front dynamic evolution process from time step t to time step \(t+7\): a evolution of key ocean front 22 and isolated ocean front 19 at time step t; b evolution of key ocean front 11 and isolated ocean fronts 2, 9, and 10 at time step t; c evolution of key ocean front 4 and isolated ocean fronts 2 and 6 at time step t; d evolution of key ocean front 1 and isolated ocean front 1 at time step t

Full size image

In Fig. 8, four samples from the ocean front dynamic evolution dataset are displayed. As shown in Fig. 8a, ocean front 19 at time step t overlaps with key ocean front 22, and ocean front 19 splits into ocean fronts 15 and 16 at time step \(t+1\). In the following time steps, we track key ocean front 22 and ocean fronts 14, 17, and 18. At time step \(t+3\), ocean front 16 at time step \(t+2\) splits into ocean fronts 14, 17, and 18; besides, ocean fronts 17 and 18 overlap with key ocean front 22. Therefore, in the following time steps, we track key ocean front 22 and ocean fronts 14, 17, and 18. Ocean fronts 13, 17, and 18 at time step \(t+6\) merge into ocean front 15 at time step \(t+6\). In the following time steps, we track the key ocean front 22 and ocean front 15. The ocean front dynamic evolution process shown in Figs. 8b–d follow the same principle.

5 Conclusions

This paper introduces a rigorous methodology tailored for constructing a comprehensive dataset on ocean front dynamic evolution, specifically designed to support research endeavors employing deep learning models. The proposed methodology comprises a series of algorithmic components, including ocean front segmentation, key ocean front extraction, ocean front variation classification, ocean front interaction classification, and ocean front dynamic evolution tracking. By establishing this robust framework, researchers are provided with a solid foundation upon which to develop novel models aimed at enhancing the accuracy of ocean front tracking. Furthermore, this dataset holds promise for shedding light on the underlying evolution principles governing ocean front evolution, thus fostering advancements in related scientific inquiries.