Sea island reef remote sensing image substrate classification data set construction method and device, electronic equipment and storage medium

Abstract

The invention discloses an island remote sensing image substrate classification data set construction method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining island remote sensing data; based on the priori island substrate classification system, a target island substrate classification system is built according to the feature condition of each substrate category; obtaining a pre-selected ROI sample in the island remote sensing data; carrying out substrate category analysis based on the ROI sample to obtain a low-separation-degree target substrate category set; performing preliminary classification by using a support vector machine based on the ROI sample to obtain a preliminary classification result; and based on the target substrate category set, performing correction processing on the preliminary classification result in response to a manual correction instruction of the target object to obtain an island substrate classification data set. The method provides an effective and feasible technical path for realizing automatic and standardized construction of the high-precision substrate classification data set of the open sea island reef, and can be widely applied to the technical field of substrate classification.

Description

Technical Field

[0001] This invention relates to the field of sediment classification technology, and in particular to a method, apparatus, electronic device, and storage medium for constructing a sediment classification dataset of remote sensing images of islands and reefs. Background Technology

[0002] Islands and reefs are an important component of the marine environment, possessing abundant biological, mineral, and tourism resources. They are also crucial units for marine ecological health and sustainable development. Furthermore, some islands and reefs are important battlegrounds for safeguarding national maritime rights. Therefore, research on islands and reefs is a vital part of marine scientific research. Accurate classification of the substrate types of remote islands and reefs has always been a key task in marine scientific research. Precise island and reef substrate data and classification are fundamental to island and reef development and ecological protection, providing crucial support for marine ecological environment protection, natural resource management, and the safeguarding of national maritime rights. The earliest classification of island and reef substrate types can be traced back to the coral island and reef classification system proposed by Goreau et al. (1959). This system proposed 11 categories: coastal zone, lagoon, adjacent lagoon zone, reef flat, wave zone, barren zone, mixed zone, buttress zone, reef foreshore, shallow reef foreshore slope, and deep reef foreshore slope. This classification combined the geographical location, marine dynamic environment, and substrate composition characteristics of each coral reef substrate type, laying the foundation for later island and reef substrate classification. However, existing technologies for classifying the seabed of islands and reefs have the following problems: There are numerous classification systems for island and reef substrates, depending on different research objectives and application scenarios. There are many versions of classification systems for islands and reefs that can be referenced. However, there is currently a lack of a unified island and reef substrate classification system that can effectively and reasonably describe the substrate types of remote islands and reefs (such as the Xisha Islands), and that is highly comprehensive and suitable for the production of efficient and automatic identification datasets of island and reef substrate remote sensing data.

[0003] The cost of on-site surveys of islands and reefs located in remote offshore areas is high, making pixel-by-pixel classification difficult. Furthermore, different substrate types, such as sand, coral sediments, living corals, and reef ridges, often exhibit an intermingled or scattered distribution. This results in a spatial distribution pattern where different substrate types within the same image contain each other. Traditional labeling methods in these areas, especially when accurately delineating the boundaries of these intermingled substrate types, often face high labor costs, leading to labeling difficulties. Summary of the Invention

[0004] The main objective of this invention is to provide a method, apparatus, electronic device, storage medium, and program product for constructing a seabed classification dataset of remote sensing images of islands and reefs, aiming to solve at least one problem in the prior art.

[0005] To achieve the above objectives, one aspect of this invention proposes a method for constructing a seabed classification dataset for remote sensing images of islands and reefs, the method comprising: Acquire remote sensing data of islands and reefs; wherein, the remote sensing data of islands and reefs adopts remote sensing observations of the target band; Based on the prior island and reef sediment classification system, a target island and reef sediment classification system is constructed according to the characteristics of each sediment category. The target island and reef substrate classification system includes substrate categories such as deep sea, reef pond, shallow reef foreslope, deep reef foreslope, reef ridge, coral deposition area, living coral, sandy land, vegetation, and land. Obtain pre-selected ROI samples from the remote sensing data of islands and reefs; wherein, the ROI samples are labeled with the substrate category in the substrate classification system of the target islands and reefs; Based on ROI samples, sediment category analysis was performed to obtain a target sediment category set with low separation. Preliminary classification was performed using a support vector machine based on the ROI samples to obtain preliminary classification results. Based on the target substrate category set, the preliminary classification results are corrected in response to the manual correction instructions of the target object to obtain the island and reef substrate classification dataset.

[0006] In some embodiments, acquiring remote sensing data of islands and reefs includes the following steps: Preliminary remote sensing data was downloaded from a pre-designed remote sensing information website; Remote sensing observations of the target bands were obtained from the initial remote sensing data and compiled into island and reef remote sensing data. The target bands include the blue band, green band, red band, and near-red band.

[0007] In some embodiments, based on a priori reef sediment classification system, a target reef sediment classification system is constructed according to the characteristics of each sediment category, including the following steps: Based on the biological characteristics of spotted reefs, spotted reefs are classified as living corals. Based on the substrate composition, formation conditions and spatial location characteristics of the foreshore terraces, the foreshore terraces are classified as reef ridges. Based on the similarity in composition and spectral curve characteristics between the inner edge of atolls and coral deposition areas, the inner edge of atolls is classified as a coral deposition area. Based on the common origin and evolutionary transformation relationship of the substrate composition of sandbars and sand flats, sandbars and sand flats are merged into sandy land; Based on living corals, reef ridges, coral sedimentary areas, and sandy areas, and combined with prior reef substrate categories such as deep sea, reef ponds, shallow reef foreshore slopes, deep reef foreshore slopes, vegetation, and land, a target reef substrate classification system was constructed.

[0008] In some embodiments, the method further includes the following steps: In response to the region selection and division command for the target object, the region of interest is selected in the island and reef remote sensing data, and then the seabed category of the region of interest is labeled to obtain ROI samples.

[0009] In some embodiments, sediment category analysis is performed based on ROI samples to obtain a target sediment category set with low separation, including the following steps: Spectral analysis was performed on the ROI samples corresponding to each substrate category to obtain the spectral curves for all substrate categories; Spectral features corresponding to each substrate category are extracted from the spectral curves. Based on the spectral features, the JM distance is used to measure the separation degree between different substrate categories, and the separation degree relationship between different substrate categories is obtained. The target substrate category set with low separation degree is obtained by screening based on the separation degree relationship.

[0010] In some embodiments, a preliminary classification is performed using a support vector machine based on ROI samples to obtain a preliminary classification result, including the following steps: Hyperplanes for different substrate categories are constructed based on ROI samples corresponding to each substrate category; Based on the hyperplane, support vector machines are used for margin-constrained supervised classification to obtain the optimal decision boundary for each substrate category; Preliminary classification results are determined based on the optimal decision boundary.

[0011] In some embodiments, based on the target substrate category set, the preliminary classification results are corrected in response to a manual correction instruction for the target object to obtain an island / reef substrate classification dataset, including the following steps: Push key classification reminders to target objects based on the target sediment category set; The preliminary classification results are raster data; Convert each pixel of the raster data into vector point data centered on the pixel; Obtain manual correction instructions for the target object; In response to manual correction instructions, the attribute values ​​of target point data in vector point type data are adjusted, and then the vector point type data is converted back to raster data, resulting in an island and reef bottom sediment classification dataset.

[0012] To achieve the above objectives, another aspect of the present invention proposes a device for constructing a seabed classification dataset from remote sensing images of islands and reefs. The device includes: The data acquisition module is used to acquire remote sensing data of islands and reefs; the remote sensing data of islands and reefs adopts remote sensing observations of the target band. The system construction module is used to build a target island and reef bottom classification system based on the prior island and reef bottom classification system and the characteristics of each bottom category. The target island and reef substrate classification system includes substrate categories such as deep sea, reef pond, shallow reef foreslope, deep reef foreslope, reef ridge, coral deposition area, living coral, sandy land, vegetation, and land. The sample acquisition module is used to acquire pre-selected ROI samples from the island and reef remote sensing data; the ROI samples are labeled with the substrate category in the target island and reef substrate classification system; The separation analysis module is used to perform sediment category analysis based on ROI samples to obtain a target sediment category set with low separation. The preliminary classification module is used to perform preliminary classification based on ROI samples using a support vector machine to obtain preliminary classification results. The manual correction module is used to correct the preliminary classification results based on the target substrate category set and in response to the manual correction instructions of the target object, so as to obtain the island and reef substrate classification dataset.

[0013] In some embodiments, the apparatus further includes a sample selection module for performing the following operations: In response to the region selection and division command for the target object, the region of interest is selected in the island and reef remote sensing data, and then the seabed category of the region of interest is labeled to obtain ROI samples.

[0014] To achieve the above objectives, another aspect of the present invention provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.

[0015] To achieve the above objectives, another aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0016] To achieve the above objectives, another aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0017] The embodiments of this invention include at least the following beneficial effects: This invention provides a method, apparatus, electronic device, storage medium, and program product for constructing a substrate classification dataset for island and reef remote sensing images. This scheme acquires island and reef remote sensing data; wherein the island and reef remote sensing data uses remote sensing observations of the target band; based on a prior island and reef substrate classification system, a target island and reef substrate classification system is constructed according to the characteristics of each substrate category; wherein the substrate categories of the target island and reef substrate classification system include deep sea, reef pond, shallow reef foreslope, deep reef foreslope, reef ridge, coral deposition area, living coral, sandy land, vegetation, and land; pre-selected ROI (Region of Interest) samples are acquired from the island and reef remote sensing data; wherein the ROI samples are labeled with the substrate category labels in the target island and reef substrate classification system; substrate category analysis is performed based on the ROI samples to obtain a target substrate category set with low separation; preliminary classification is performed using a support vector machine based on the ROI samples to obtain preliminary classification results; based on the target substrate category set, the preliminary classification results are corrected in response to manual correction instructions for the target object to obtain the island and reef substrate classification dataset. This invention addresses the problems of chaotic and inapplicable classification systems in existing technologies by constructing a comprehensive target classification system optimized specifically for remote sensing identification of remote islands and reefs. Preliminary automatic classification using Support Vector Machines (SVM) significantly reduces reliance on costly field surveys, enabling large-scale and efficient extraction of seabed information. To address the unavoidable misclassification problem in automatic classification, an innovative manual correction mechanism based on low-discretionary-degree category sets is introduced. This focuses arduous manual labor on the most challenging areas requiring expert intervention, rather than full map annotation, thereby significantly reducing the cost and difficulty of manual annotation while maintaining classification accuracy. This successfully overcomes the annotation difficulties caused by the intermingled distribution of seabed sediment. This invention provides an effective and feasible technical path for the automated and standardized construction of high-precision seabed classification datasets for remote islands and reefs. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an implementation environment for the method of constructing a dataset for classifying the seabed sediment of remote sensing images of islands and reefs provided in this embodiment of the invention; Figure 2 This is a flowchart illustrating a method for constructing a seabed classification dataset for remote sensing images of islands and reefs, as provided in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the relevant characteristic descriptions of each typical substrate category and their true-color sample examples on remote sensing images, provided by embodiments of the present invention. Figure 4 This is a schematic diagram illustrating an example of a spectral curve feature map of a coral deposition area provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating an example of sample selection from Antelope Reef in the Xisha Islands provided in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating various spectral curves of ground features provided in embodiments of the present invention; Figure 7 This is a schematic diagram illustrating the separation relationship of different substrate categories in the selected training sample ROI provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating an example of the preliminary classification of the accuracy of various substrates on a verification sample, provided by an embodiment of the present invention. Figure 9 This is an example diagram of correction using Google Earth high-resolution imagery provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the overall technical route of the method for constructing a seabed classification dataset of remote sensing images of islands and reefs provided in the embodiments of the present invention; Figure 11 This is a schematic diagram illustrating an example of precise classification accuracy verification for each substrate category after correction, provided in an embodiment of the present invention. Figure 12 This is a schematic diagram illustrating an example of a dataset constructed using the method for constructing a seabed classification dataset based on remote sensing images of islands and reefs, as provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0020] It is understood that the terms “first,” “second,” etc., used in this invention may be used herein to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to determination” as used herein may be interpreted as “when…” or “when…” or “in response to determination.”

[0021] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0023] To facilitate understanding of the technical solution of this invention, the technical terms that may be involved in the technical solution of this invention will be explained first: Seabed classification for islands and reefs: The formation and evolution of islands and reefs produce seabed types with different characteristics. Classification methods, standards, and categories are generally established based on the characteristics of the seabed topography, the dynamic environment of the seabed formation process, and the biological environment formed by the seabed sediments. To suit the specific application scenarios of various research projects, when constructing an island and reef seabed classification system, the classification criteria should be comprehensively selected based on the actual application and the characteristics of the seabed in the target area.

[0024] Remote sensing image classification dataset: A structured and organized collection of data whose core component is remote sensing imagery (such as satellite or aerial imagery), and some or all areas of this imagery data have been manually or semi-automatically labeled with land cover category labels (such as "water bodies," "buildings," "forests," "farmland," etc.). The core purpose of this dataset is to provide training samples (for the algorithm to learn features of different categories) and validation / test samples (for evaluating the accuracy and generalization ability of the algorithm's classification) for developing and evaluating automatic remote sensing image classification algorithms (especially machine learning and deep learning-based methods).

[0025] In related technologies, existing techniques for classifying island and reef substrates have the following problems: Numerous island and reef substrate classification systems exist depending on different research objectives and application scenarios, and many versions of classification systems can be referenced. However, there is currently a lack of a unified island and reef substrate classification system that can effectively and reasonably describe the substrate types of remotely located islands and reefs (such as the Xisha Islands), possess strong comprehensiveness, and is suitable for the efficient and automatic identification of island and reef substrate remote sensing data datasets. The cost of on-site surveys of islands and reefs in offshore areas is high, making pixel-by-pixel classification difficult. Furthermore, different substrate types, such as sand, coral sediments, living corals, and reef ridges, often exhibit overlapping and sporadic distributions. This results in a spatial distribution characteristic where different substrate types in the same image contain each other. When creating traditional label data in these areas, especially when accurately delineating the boundaries of these overlapping substrate types, high manual costs often lead to labeling difficulties.

[0026] In view of this, this invention provides a method, apparatus, electronic device, and storage medium for constructing a seabed classification dataset from remote sensing images of islands and reefs. This scheme involves acquiring remote sensing data of islands and reefs; wherein the remote sensing data uses remote sensing observations of the target band; based on a priori island and reef seabed classification system, a target island and reef seabed classification system is constructed according to the characteristics of each seabed category; wherein the seabed categories of the target island and reef seabed classification system include deep sea, reef pond, shallow reef foreslope, deep reef foreslope, reef ridge, coral deposition area, living coral, sandy land, vegetation, and land; acquiring pre-selected ROI samples from the island and reef remote sensing data; wherein the ROI samples are labeled with the seabed category tags in the target island and reef seabed classification system; performing seabed category analysis based on the ROI samples to obtain a target seabed category set with low separation; performing preliminary classification using a support vector machine based on the ROI samples to obtain preliminary classification results; and correcting the preliminary classification results based on the target seabed category set in response to manual correction instructions from the target object to obtain the island and reef seabed classification dataset. This invention addresses the problems of chaotic and inapplicable classification systems in existing technologies by constructing a comprehensive target classification system optimized specifically for remote sensing identification of remote islands and reefs. Preliminary automatic classification using Support Vector Machines (SVM) significantly reduces reliance on costly field surveys, enabling large-scale and efficient extraction of seabed information. To address the unavoidable misclassification problem in automatic classification, an innovative manual correction mechanism based on low-discretionary-degree category sets is introduced. This focuses arduous manual labor on the most challenging areas requiring expert intervention, rather than full map annotation, thereby significantly reducing the cost and difficulty of manual annotation while maintaining classification accuracy. This successfully overcomes the annotation difficulties caused by the intermingled distribution of seabed sediment. This invention provides an effective and feasible technical path for the automated and standardized construction of high-precision seabed classification datasets for remote islands and reefs.

[0027] It is understood that the method for constructing a seabed classification dataset for remote sensing images of islands and reefs provided by this invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various types of terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.

[0028] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided by an embodiment of the present invention. (Refer to...) Figure 1 The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0029] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0030] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0031] Terminal 102 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the invention does not impose any limitations.

[0032] For example, based on Figure 1The implementation environment shown in this embodiment of the invention provides a method for constructing a seabed classification dataset of remote sensing images of islands and reefs. The following description uses the application of this method in server 101 as an example. It can be understood that this method can also be applied in terminal 102.

[0033] Reference Figure 2 , Figure 2 This is an optional flowchart of the method for constructing a seabed classification dataset of remote sensing images of islands and reefs provided in the embodiments of the present invention. The execution subject of this method for constructing a seabed classification dataset of remote sensing images of islands and reefs can be any of the aforementioned computer devices (including servers or terminals). Figure 2 The method may include, but is not limited to, steps S100 to S600.

[0034] Step S100: Obtain remote sensing data of islands and reefs; Among them, the remote sensing data for islands and reefs uses remote sensing observations of the target band; It should be noted that in some embodiments, step S100 may include the following steps: downloading preliminary remote sensing data from a preset remote sensing information website; selecting remote sensing observations of target bands from the bands used in the preliminary remote sensing data and organizing them into island and reef remote sensing data; wherein, the target bands include blue band, green band, red band and near-red band.

[0035] For example, in some specific implementations, Sentinel-2 island and reef remote sensing data can be downloaded from remote sensing information websites. Here, Sentinel-L2A level data is used. The provider, the European Space Agency, has released finished data that has undergone radiometric calibration and atmospheric correction. It can be used directly after download. Moreover, this data is free to download and has a large time resolution for Earth observation, providing sufficient experimental analysis material. At the same time, its ease of download and application enhances the practical value of this experiment. In addition, the Sentinel-2 L2A data uses remote sensing observations in the Blue, Green, Red, and RedEdeg bands with a resolution of 10m. These four bands contain most of the solar radiation energy, and their reflectance values ​​carry more spectral information about ground objects, which can clearly distinguish and identify ground objects. Their spectral characteristics can largely summarize the spectral characteristics of various ground objects.

[0036] Specifically, this embodiment of the invention optimizes the data source by selecting specific target bands (blue, green, red, and near-red) from multi-band remote sensing data as classification features. The benefits are twofold: firstly, these bands are most sensitive to the spectral characteristics of water penetration and different substrates, laying a data foundation for subsequent high-precision classification; secondly, it avoids using irrelevant or noisy bands, improving data processing efficiency and enhancing the generalization ability and interpretability of the classification model.

[0037] Step S200: Based on the prior island and reef sediment classification system, the target island and reef sediment classification system is constructed according to the characteristics of each sediment category. The target island and reef substrate classification system includes substrate categories such as deep sea, reef pond, shallow reef foreslope, deep reef foreslope, reef ridge, coral deposition area, living coral, sandy land, vegetation, and land. It should be noted that in some embodiments, step S200 may include the following steps: classifying point reefs as living corals based on their biological characteristics; classifying foreshore terraces as reef ridges based on their substrate composition, formation conditions, and spatial location characteristics; classifying the inner edge of atolls as coral deposition areas based on the similarity in composition and spectral curve characteristics between the inner edge of atolls and coral deposition areas; merging sandbars and sand flats into sandy land based on the homology and evolutionary transformation relationship of their substrate composition; and constructing a target reef substrate classification system based on living corals, reef ridges, coral deposition areas, and sandy land, combined with prior reef substrate categories of deep sea, reef ponds, shallow reef foreshore slopes, deep reef foreshore slopes, vegetation, and land.

[0038] For example, in some specific implementations, the classification system for island and reef sediment can be established as follows: There are already many classification systems for island and reef substrates, each designed for different applications. This invention combines the characteristics and classification rules of previous island and reef substrate classification systems and proposes a new system suitable for remote sensing imagery. The classification system considers the spectral characteristics, composition, spatial distribution, and coral reef evolution of various substrate categories in remote sensing images. Figure 3 The image shows a description of the relevant characteristics of each typical substrate category and an example of its true-color sample on a remote sensing image.

[0039] Based on the spectral characteristics, spatial distribution, and overall area of ​​various sediment types, the following mergers and modifications were made to common sediment types: (1) Pointed reefs are distributed in areas such as reef ponds and lagoon slopes, and their biological nature is that of living corals that are growing normally. Therefore, in order to better count the overall coral reef coverage area of ​​islands and reefs, pointed reefs are regarded as living corals.

[0040] (2) The foreshore terrace is located between the reef ridge and the shallow foreshore slope. The reef ridge is formed by waves and storm surges uplifting and depositing reef blocks or pebbles. These materials originate from or pass through the foreshore terrace. At the same time, the position of the foreshore terrace changes with the sea level and the position of the wave-breaking zone outside the reef. The distinction between the two is usually unclear. Moreover, the area occupied by the foreshore terrace is relatively small. For example, the east-west length of Antelope Reef is about 4,000 meters, while the length of the foreshore terrace in the same direction is only about 20 meters. Furthermore, its bottom composition and formation conditions are similar to those of the reef ridge. Combining the two will not affect the description of the main categories and spatial characteristics of the reef. Therefore, the foreshore terrace is classified as a reef ridge here.

[0041] (3) As coral reefs develop from fringing reefs to atolls, lagoons gradually form, and the central part of the coral reef gradually sinks, turning the originally enclosed area into deep sea. This results in the existence of coral deposition areas along the shoreline inside the original atoll. Taking Yongle Atoll as an example, the area between the outer reef ridge and the inner reef pool is a coral deposition area, such as Antelope Reef. As the interior gradually sinks and evolves into an atoll, such as the entire Yongle Atoll, the inner edge of the atoll was also a central coral deposition area before the interior sinks. Therefore, although these areas are currently judged as reef foreland slopes according to their spatial location, their formation and dynamic characteristics are significantly different from those of traditional reef foreland slopes based on the geological and geomorphological evolution of islands and reefs. Therefore, the essence of the sediments in this area is still the same as that of coral deposition areas. On the other hand, the spectral curve characteristics of the coral deposition areas inside the coral reef and the inner edge of the atoll are highly similar, such as Figure 4 The image shown is an example of a spectral curve characteristic diagram of a coral deposition area. Therefore, these areas are collectively categorized as coral deposition areas.

[0042] (4) Sandbars are relatively stable landforms formed by the accumulation of sand dunes by wind and waves. The bottom composition of sandbars and sand dunes is the same, both formed by biological remains, mainly coral gravel and shell fragments, after being eroded and affected by the external environment. In the dynamic evolution of islands and reefs, the two will transform into each other, so they are merged into sandy land type so as to uniformly monitor the changes in the area of ​​island and reef sandy land.

[0043] Based on the above description and analysis of the characteristics of each substrate category, the classification system is finally divided into 10 categories: "deep sea", "reef pond", "shallow reef foreslope", "deep reef foreslope", "reef ridge", "coral deposition area", "living coral", "sandy land", "vegetation" and "land".

[0044] Specifically, this invention simplifies the complexity of the classification system and makes its number of categories more reasonable by scientifically merging and classifying prior categories (for example, merging sandbars and sand flats into "sandy land," and classifying the inner edges of atolls into "coral deposition areas"). This not only enhances the system's comprehensive descriptive ability for remote islands and reefs (especially atolls like the Xisha Islands), but also ensures that each final category has significant and easily distinguishable spectral or spatial characteristics through remote sensing, greatly improving the system's adaptability to automated remote sensing identification methods and solving the core problem of the lack of a suitable unified system in the prior art.

[0045] Step S300: Obtain pre-selected ROI samples from the island and reef remote sensing data; Among them, the ROI samples are labeled with the substrate category in the target island / reef substrate classification system; It should be noted that in some embodiments, the method may further include the following steps: in response to the region selection and division instruction of the target object, selecting the region of interest in the island and reef remote sensing data, and then labeling the region of interest with the seabed category to obtain ROI samples.

[0046] For example, in some specific implementations, the selection of ROI samples can be achieved by: selecting a representative and accurate sample set based on visual interpretation experience and combined with Google Earth high-resolution imagery. During the selection process, it is important to ensure that the selected ROIs represent the typical characteristics of the target land cover category, ensuring the specificity and accuracy of each category's ROIs; avoiding samples from edge areas; ensuring a uniform spatial distribution of the selected ROI samples; and avoiding the selection of mixed pixels. Figure 5 The image shows an example of sample selection at Antelope Reef in the Xisha Islands.

[0047] Specifically, this invention creates training samples by allowing users to flexibly select regions of interest (ROIs) on images for labeling. Its advantages are: this method greatly simplifies the sample labeling process. Labellers do not need to exhaustively depict the entire image; they only need to quickly label typical regions that represent the category. This significantly reduces the initial manual and time costs of creating the training dataset, facilitating subsequent automatic classification model training.

[0048] Step S400: Perform sediment category analysis based on ROI samples to obtain a target sediment category set with low separation. It should be noted that in some embodiments, step S400 may include the following steps: performing spectral analysis on the ROI samples corresponding to each substrate category to obtain spectral curves for all substrate categories; extracting spectral features corresponding to each substrate category from the spectral curves; using the JM distance based on the spectral features to measure the separation degree between different substrate categories to obtain the separation degree relationship between different substrate categories; and filtering out a target substrate category set with low separation degree based on the separation degree relationship. Here, low separation degree represents a state where the separation degree is lower than a preset threshold or meets a preset condition.

[0049] For example, in some specific implementations, the selected ROI samples are subjected to spectral analysis, analyzed from two perspectives: the sample spectral curves and the separation relationship between samples. Specifically: like Figure 6 As shown, the spectral curves of various land features are examples. It is found that the spectral curve of "reef pond" is between the spectral curves of "deep reef front slope" and "shallow reef front slope", and it shows similar spectral characteristics to the spectral curves of both.

[0050] Next, we analyze the spectral features using the separation relationship between samples. Separability is an indicator used to describe the degree of difference in classification features between categories, and its commonly used metric is the Jeffries-Matusita (JM) distance.

[0051] in Distance to Bhattacharyya:

[0052] in This represents the Mahalanobis distance of the mean differences in the covariance space, while the latter half... This represents the impact of covariance matrix differences on distribution overlap. The separation degree defined above is commonly used in supervised remote sensing classification to measure the similarity between samples. A higher separation degree indicates a higher degree of distinguishability between land cover types, resulting in better accurate classification. When performing supervised classification on remote sensing images based on a reasonable selection of ROI, a separation degree greater than 1.9 indicates accurate classification between two land cover types. The separation degree of spectral characteristic values ​​for each substrate type in the four bands (Blue, Green, Red, and RedEdeg) is calculated using the Jeffries-Matusita (JM) distance. Figure 7 The figure shows an example of the separation degree relationship of each substrate category in the selected training sample ROI.

[0053] Based on the separation degree relationship, the separation degree between "reef pool" and "deep reef foreslope" and "shallow reef foreslope" is relatively low, which is consistent with the changing trend of the above spectral curves.

[0054] Specifically, this invention uses JM distance to quantitatively measure the spectral separation between different substrate categories. Its advantages lie in transforming traditional category analysis, which relies on subjective experience, into an objective and quantitative scientific analysis. It can accurately and automatically identify which categories are most prone to confusion in automatic classification (i.e., low-separation category sets), providing a clear target and direction for subsequent targeted manual correction. This avoids the blindness of manually inspecting the entire map, further improving the efficiency and accuracy of the entire method.

[0055] Step S500: Based on the ROI samples, a support vector machine is used to perform preliminary classification to obtain preliminary classification results; It should be noted that in some embodiments, step S500 may include the following steps: constructing hyperplanes for different substrate categories based on ROI samples corresponding to each substrate category; performing supervised classification with margin constraints using a support vector machine based on the hyperplanes to obtain the optimal decision boundary for each substrate category; and determining the preliminary classification result based on the optimal decision boundary.

[0056] For example, in some specific embodiments, the supervised classification method used in the embodiments of the present invention is Support Vector Machine (SVM). This algorithm is a commonly used machine learning classification algorithm with excellent classification performance. Its goal is to find hyperplanes between each category, using these boundaries to divide the data into different categories. Here, the dataset is assumed to be:

[0057] in For data samples, For the corresponding category label, the label value is the value corresponding to the category. The equation of the hyperplane is:

[0058] in The normal vector determines the direction of the hyperplane. The bias term determines the distance from the hyperplane to the origin.

[0059] The goal of SVM is to find a hyperplane that maximizes the margin to the support vectors.

[0060] in The original objective function for maximizing the margin is very complex because it involves both max and min operations, making it extremely difficult to solve directly. Therefore, the objective of maximizing the margin can be transformed into an optimization problem:

[0061] The optimization interval condition is constrained by:

[0062] here It is the length of the normal vector. Let represent the square of the margin, and define the minimum margin numerator as 1. The optimal decision boundary for splitting each category is obtained by analyzing this optimization equation.

[0063] Preliminary classification results were obtained through SVM supervised classification, such as Figure 8 The image shows an example of the verification accuracy of the preliminary classification of various substrates on the verification sample. Based on the verification accuracy between categories, it can be found that both deep and shallow reef foreslopes are misidentified as reef ponds to varying degrees, and similarly, some reef ponds are also identified as deep or shallow reef foreslopes. Furthermore, living corals, reef ridges, and sandy areas are spatially adjacent, and the growth and aging of corals, along with their biodebris and other residues, gradually transform into components of sandy areas and reef ridges. Similarly, sandy areas, reef ridges, and coral sediments can transform into each other due to wave transport and aggregation. Therefore, these substrate categories within the reef platform all exhibit varying degrees of mutual interference in the preliminary classification.

[0064] Specifically, this embodiment of the invention employs Support Vector Machine (SVM) for supervised classification. Its advantages lie in the fact that the SVM algorithm is particularly suitable for processing high-dimensional data such as remote sensing imagery, and by finding the optimal decision boundary, it can achieve good generalization performance even with a limited number of samples. This is crucial for research on remote islands and reefs where on-site samples are scarce. This step enables rapid and large-scale preliminary automatic classification, and is a core technical step in reducing reliance on on-site surveys and improving overall efficiency.

[0065] Step S600: Based on the target substrate category set, the preliminary classification results are corrected in response to the manual correction instructions of the target object to obtain the island and reef substrate classification dataset; It should be noted that in some embodiments, step S600 may include the following steps: pushing key classification reminders to the target object based on the target substrate category set; wherein, the preliminary classification result is raster data; converting each pixel of the raster data into vector point type data centered on the pixel; obtaining the manual correction instruction of the target object; adjusting the attribute values ​​of the target point data in the vector point type data in response to the manual correction instruction, and then converting the vector point type data back to raster data to obtain the island and reef substrate classification dataset.

[0066] For example, in some specific implementations, precise label classification can be performed based on the preliminary classification results, and the process is as follows: Preliminary classification results are obtained through supervised classification using machine learning. However, since machine learning classification algorithms have limited ability to extract high-dimensional features from data, further manual correction is performed on the basis of supervised classification to further improve classification accuracy.

[0067] To facilitate the editing of pixel label values, each pixel in the raster data is converted into vector point data centered on the pixel. The raster labels are manually corrected by editing the attribute values ​​of each point data. Geoscience software such as ArcGIS and ArcGIS Pro are used for attribute editing, and then the corrected point data is converted back into raster data.

[0068] Raster values ​​are integer label values ​​corresponding to the land feature categories. Each pixel is converted using the pixel center point method. When converting vector point types back to raster data, the same resolution as the original raster label is used. Therefore, labels will not experience resampling, positional shifts, or other errors before and after conversion and modification. It is worth noting that because the separation and spectral curve differences of the six categories "reef pond," "shallow reef foreslope," "deep reef foreslope," "living coral," "vegetation," and "land" in remote sensing impacts are relatively small, manual correction is necessary. "Deep sea," "sandy land," "coral deposition area," and "reef ridge" have higher separation and larger spectral curve differences compared to other categories, resulting in higher initial classification accuracy. When correcting easily confused categories, experience in manually interpreting reef substrate is used, combined with comprehensive conditions such as spatial distribution, texture characteristics, and spectral features to determine the substrate category. For example, the reef foreshore slope type will not appear inside the reef; the presence of wave patterns on the outermost part of the reef indicates the presence of water at that location, thus identifying the boundary between the reef ridge and the shallow reef foreshore slope; live corals grow in all locations on the reef, their spectral appearance is brownish-black or grayish-black and their growth is discrete, often confused with sandy areas, reef ridges, etc. Figure 9 The image shown is an example of correction performed using high-resolution Google Earth imagery.

[0069] Specifically, this embodiment of the invention first guides experts to focus on the most error-prone areas by pushing key point reminders, greatly improving the efficiency of manual review. Secondly, by converting raster data into vector point data for correction, the object of operation changes from a group of pixels to a single point, simplifying the logic and complexity of manual interaction. This allows correction personnel to make fine adjustments through "error checking and correction," which reduces the workload by orders of magnitude compared to redrawing the entire polygon boundary, effectively solving the pain point of "difficult annotation" in the background technology. Finally, the data is converted back to raster format, ensuring the standardization and usability of the data results.

[0070] In some preferred embodiments, before performing step S500, the method of the present invention may further include the following steps: Substrate category analysis was performed on each ROI sample to obtain the separation degree of each ROI sample. Then, based on the separation degree, ROI samples with high separation degree were selected as the sample data basis for preliminary classification.

[0071] It should be understood that the purpose of step S400 is to analyze and obtain multiple substrate categories with small differences in separation, so as to focus on manual correction. The purpose of the above embodiment to select ROI samples with high separation is to select a batch of samples with large differences in separation of each substrate category from all ROI samples as the sample data basis for subsequent preliminary classification. Based on the separation advantage of the ROI samples themselves, the workload of manual correction can be reduced to a certain extent, which not only ensures the classification accuracy of substrate categories in the dataset, but also improves the construction efficiency of the dataset.

[0072] In some optional implementations, the island and reef substrate classification dataset constructed in the embodiments of the present invention can be applied to the training of the island and reef substrate classification model, which can effectively improve the classification accuracy of the model.

[0073] To explain in detail the principle of the technical solution of the present invention, the overall process of the present invention will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present invention and should not be regarded as a limitation of the present invention.

[0074] First, it should be noted that the current classification standards for island and reef substrates are as follows: (1) classification based on the characteristics of the seabed topography and geomorphology; (2) classification based on the dynamic environment of the substrate formation process; and (3) classification based on the biological environment composed of the substrate sediments. In actual research, a suitable substrate classification system should be derived by combining the spatial distribution, substrate composition, and environmental characteristics of each substrate type in the target area.

[0075] Due to their unique geographical location and complex geological environment, obtaining large-scale information on the seabed of remote islands and reefs is challenging. The optical characteristics of different seabed types vary significantly, and satellite remote sensing technology has effectively alleviated this difficulty. High-resolution remote sensing technology is increasingly used in Earth observation, enabling rapid, large-area observation of the types and changes in the seabed of islands and reefs based on high-resolution satellite optical remote sensing image data. Consequently, a classification system for island and reef seabed based on remote sensing data has emerged. This classification system primarily focuses on the optical characteristics of each seabed type in remote sensing images. By describing the spectral characteristics of each seabed type in remote sensing images and combining this with visual interpretation, a high-precision island and reef seabed classification label dataset is created. This lays a crucial foundation for rapid and accurate classification and spatial distribution analysis of large-area island and reef seabeds.

[0076] In some specific implementations, to address the shortcomings of existing technologies, embodiments of the present invention provide a method for constructing a seabed classification dataset from remote sensing images of islands and reefs, such as... Figure 10 The diagram illustrates the overall technical approach of this invention. Specifically, the method of this invention can be implemented through the following process: 1. Data preparation and processing: Download Sentinel-2 island and reef remote sensing data from the remote sensing information website. This example uses Sentinel-L2A level data, provided by the European Space Agency (ESA), which is pre-calibrated and atmospherically corrected, ready for immediate use. This data is available for free download and offers ample experimental analysis material due to its large temporal resolution. The ease of download and application enhances the practical value of this experiment. The Sentinel-2 L2A data uses remote sensing observations in the Blue, Green, Red, and RedEdeg bands, with a resolution of 10m. These four bands contain most of the solar radiation energy, and their reflectance values ​​carry significant spectral information about ground features, allowing for clear identification and generalization of these features. Their spectral characteristics can largely summarize the spectral features of various ground features (data acquisition method: https: / / dataspace.copernicus).

[0077] 2. Establishment of a classification system for the seabed sediment of islands and reefs: There are already many classification systems for island and reef substrates, each designed for different applications. This invention combines the characteristics and classification rules of previous island and reef substrate classification systems and proposes a new system suitable for remote sensing imagery. The classification system considers the spectral characteristics, composition, spatial distribution, and coral reef evolution of various substrate categories in remote sensing images. Figure 3The image shows a description of the relevant characteristics of each typical substrate category and an example of its true-color sample on a remote sensing image.

[0078] Based on the spectral characteristics, spatial distribution, and overall area of ​​various sediment types, the following mergers and modifications were made to common sediment types: (1) Pointed reefs are distributed in areas such as reef ponds and lagoon slopes, and their biological nature is that of living corals that are growing normally. Therefore, in order to better count the overall coral reef coverage area of ​​islands and reefs, pointed reefs are regarded as living corals.

[0079] (2) The foreshore terrace is located between the reef ridge and the shallow foreshore slope. The reef ridge is formed by waves and storm surges uplifting and depositing reef blocks or pebbles. These materials originate from or pass through the foreshore terrace. At the same time, the position of the foreshore terrace changes with the sea level and the position of the wave-breaking zone outside the reef. The distinction between the two is usually unclear. Moreover, the area occupied by the foreshore terrace is relatively small. For example, the east-west length of Antelope Reef is about 4,000 meters, while the length of the foreshore terrace in the same direction is only about 20 meters. Furthermore, its bottom composition and formation conditions are similar to those of the reef ridge. Combining the two will not affect the description of the main categories and spatial characteristics of the reef. Therefore, the foreshore terrace is classified as a reef ridge here.

[0080] (3) As coral reefs develop from fringing reefs to atolls, lagoons gradually form, and the central part of the coral reef gradually sinks, turning the originally enclosed area into deep sea. This results in the existence of coral deposition areas along the shoreline inside the original atoll. Taking Yongle Atoll as an example, the area between the outer reef ridge and the inner reef pool is a coral deposition area, such as Antelope Reef. As the interior gradually sinks and evolves into an atoll, such as the entire Yongle Atoll, the inner edge of the atoll was also a central coral deposition area before the interior sinks. Therefore, although these areas are currently judged as reef foreland slopes according to their spatial location, their formation and dynamic characteristics are significantly different from those of traditional reef foreland slopes based on the geological and geomorphological evolution of islands and reefs. Therefore, the essence of the sediments in this area is still the same as that of coral deposition areas. On the other hand, the spectral curve characteristics of the coral deposition areas inside the coral reef and the inner edge of the atoll are highly similar, such as Figure 4 The image shown is an example of a spectral curve characteristic diagram of a coral deposition area. Therefore, these areas are collectively categorized as coral deposition areas.

[0081] (4) Sandbars are relatively stable landforms formed by the accumulation of sand dunes by wind and waves. The bottom composition of sandbars and sand dunes is the same, both formed by biological remains, mainly coral gravel and shell fragments, after being eroded and affected by the external environment. In the dynamic evolution of islands and reefs, the two will transform into each other, so they are merged into sandy land type so as to uniformly monitor the changes in the area of ​​island and reef sandy land.

[0082] Based on the above description and analysis of the characteristics of each substrate category, the classification system is finally divided into 10 categories: "deep sea", "reef pond", "shallow reef foreslope", "deep reef foreslope", "reef ridge", "coral deposition area", "living coral", "sandy land", "vegetation" and "land".

[0083] 2. Creation of the substrate classification label dataset: Based on the aforementioned classification criteria for island and reef substrate types, a substrate classification dataset was created using an optimization strategy of "preliminary classification by machine learning + manual correction by experts." The preset label values ​​for "deep sea," "reef pond," "shallow reef foreslope," "deep reef foreslope," "reef ridge," "coral deposition area," "living coral," "sand," "vegetation," and "land" are 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9, respectively, with the deep sea type (0) used as the background value.

[0084] (1) Preliminary classification and spectral analysis: The main processes in this section include: selection of sample ROI (Region of Interest), analysis of sample spectral features, and verification of supervised classification results and accuracy.

[0085] 1.1 Selection of ROI samples: By combining visual interpretation experience with high-resolution Google Earth imagery, a representative and accurate sample set should be selected. During the selection process, it is important to ensure that the selected Regions of Interest (ROIs) represent the typical characteristics of the target land cover category, guaranteeing the specificity and accuracy of each category's ROIs; avoid samples from edge areas; ensure the spatial distribution of ROI samples is uniform; and avoid selecting mixed pixels. Figure 5 The image shows an example of sample selection at Antelope Reef in the Xisha Islands.

[0086] 1.2 Analysis of sample spectral characteristics: The selected ROI samples were subjected to spectral analysis, and the analysis was conducted from two perspectives: the spectral curves of the samples and the relationship between the separation of the samples.

[0087] like Figure 6 As shown, the spectral curves of various land features are examples. It is found that the spectral curve of "reef pond" is between the spectral curves of "deep reef front slope" and "shallow reef front slope", and it shows similar spectral characteristics to the spectral curves of both.

[0088] Next, we analyze the spectral features using the separation relationship between samples. Separability is an indicator used to describe the degree of difference in classification features between categories, and its commonly used metric is the Jeffries-Matusita (JM) distance.

[0089] in Distance to Bhattacharyya:

[0090] in This represents the Mahalanobis distance of the mean differences in the covariance space, while the latter half... This represents the impact of covariance matrix differences on distribution overlap. The separation degree defined above is commonly used in supervised remote sensing classification to measure the similarity between samples. A higher separation degree indicates a higher degree of distinguishability between land cover types, resulting in better accurate classification. When performing supervised classification on remote sensing images based on a reasonable selection of ROI, a separation degree greater than 1.9 indicates accurate classification between two land cover types. The separation degree of spectral characteristic values ​​for each substrate type in the four bands (Blue, Green, Red, and RedEdeg) is calculated using the Jeffries-Matusita (JM) distance. Figure 7 The figure shows an example of the separation degree relationship of each substrate category in the selected training sample ROI.

[0091] Based on the separation degree relationship, the separation degree between "reef pool" and "deep reef foreslope" and "shallow reef foreslope" is relatively low, which is consistent with the changing trend of the above spectral curves.

[0092] Spectral curve analysis involves plotting typical spectral curves for various geological categories and analyzing the similarity of these curves to determine the separability of different substrate types. This study selected the blue, green, red, and near-infrared bands. The visible light band, being the three primary colors, allows for true-color representation of remote sensing images, which aligns better with visual interpretation experience. Furthermore, these four bands contain a significant portion of the surface radiation energy and carry a wealth of surface information.

[0093] 1.3. Supervise the classification results and verify accuracy: The supervised classification method used is Support Vector Machine (SVM), a commonly used machine learning classification algorithm with excellent classification performance. Its goal is to find hyperplanes between the classes, using these boundaries to divide the data into different categories. Here, the dataset is assumed to be:

[0094] in For data samples, For the corresponding category label, the label value is the value corresponding to the category. The equation of the hyperplane is:

[0095] in The normal vector determines the direction of the hyperplane. The bias term determines the distance from the hyperplane to the origin.

[0096] The goal of SVM is to find a hyperplane that maximizes the margin to the support vectors.

[0097] in The original objective function for maximizing the margin is very complex because it involves both max and min operations, making it extremely difficult to solve directly. Therefore, the objective of maximizing the margin can be transformed into an optimization problem:

[0098] The optimization interval condition is constrained by:

[0099] here It is the length of the normal vector. Let represent the square of the margin, and define the minimum margin numerator as 1. The optimal decision boundary for splitting each category is obtained by analyzing this optimization equation.

[0100] Preliminary classification results were obtained through SVM supervised classification, such as Figure 8 The image shows an example of the verification accuracy of the preliminary classification of various substrates on the verification sample. Based on the verification accuracy between categories, it can be found that both deep and shallow reef foreslopes are misidentified as reef ponds to varying degrees, and similarly, some reef ponds are also identified as deep or shallow reef foreslopes. Furthermore, living corals, reef ridges, and sandy areas are spatially adjacent, and the growth and aging of corals, along with their biodebris and other residues, gradually transform into components of sandy areas and reef ridges. Similarly, sandy areas, reef ridges, and coral sediments can transform into each other due to wave transport and aggregation. Therefore, these substrate categories within the reef platform all exhibit varying degrees of mutual interference in the preliminary classification.

[0101] (2) Accurate label classification and results: 2.1 Manual calibration method: Preliminary classification results are obtained through supervised classification using machine learning. However, since machine learning classification algorithms have limited ability to extract high-dimensional features from data, further manual correction is performed on the basis of supervised classification to further improve classification accuracy.

[0102] To facilitate the editing of pixel label values, each pixel in the raster data is converted into vector point data centered on the pixel. The raster labels are manually corrected by editing the attribute values ​​of each point data. Geoscience software such as ArcGIS and ArcGIS Pro are used for attribute editing, and then the corrected point data is converted back into raster data.

[0103] Raster values ​​are integer label values ​​corresponding to the land feature categories. Each pixel is converted using the pixel center point method. When converting vector point types back to raster data, the same resolution as the original raster label is used. Therefore, labels will not experience resampling, positional shifts, or other errors before and after conversion and modification. It is worth noting that because the separation and spectral curve differences of the six categories "reef pond," "shallow reef foreslope," "deep reef foreslope," "living coral," "vegetation," and "land" in remote sensing impacts are relatively small, manual correction is necessary. "Deep sea," "sandy land," "coral deposition area," and "reef ridge" have higher separation and larger spectral curve differences compared to other categories, resulting in higher initial classification accuracy. When correcting easily confused categories, experience in manually interpreting reef substrate is used, combined with comprehensive conditions such as spatial distribution, texture characteristics, and spectral features to determine the substrate category. For example, the reef foreshore slope type will not appear inside the reef; the presence of wave patterns on the outermost part of the reef indicates the presence of water at that location, thus identifying the boundary between the reef ridge and the shallow reef foreshore slope; live corals grow in all locations on the reef, their spectral appearance is brownish-black or grayish-black and their growth is discrete, often confused with sandy areas, reef ridges, etc. Figure 9 The image shown is an example of correction performed using high-resolution Google Earth imagery.

[0104] 2.2 Accuracy Verification: During the calibration process, the substrate types of each part of the image were compared with the original remote sensing image and the Google Earth high-resolution ground image, and comprehensive corrections were made in conjunction with interpretation experience. The average accuracy for each category on the test sample was 0.989. Figure 11 The image shows an example of verifying the precise classification accuracy of each substrate category after correction.

[0105] 2.3 Island and Reef Substrate Classification Dataset: For example, islands and reefs within the Xisha Islands region can be selected: Huaguang Reef, Panshi Island, Beijiao Reef, Yuzhuo Reef, Langhua Reef, and Yongle Atoll. Yongle Atoll includes: Jinyin Island, Lingyang Reef, Ganquan Island, Coral Island, Quanfu Island, Yagong Island, Yinyu Island, Jinqing Island, and Guangjin Island. The downloaded remote sensing data is Sentinel-2 10m resolution data, with four bands selected: Red, Green, Blue, and RedEdge. Table 1 shows the basic information of the selected island and reef images.

[0106] Table 1

[0107] The total area of ​​the selected island and reef remote sensing images is approximately 1279.2 km², including a small area of ​​deep-sea region near each image. Both the image data and label data are stored as GeoTIFF geographic raster data in the WGS-84 coordinate system. Through the aforementioned label data creation process, these island and reef image data were compiled into an island and reef seabed classification dataset supporting deep learning, such as... Figure 12 The image shown is a partial example of the results from this dataset.

[0108] 2.4 Multi-model testing on the dataset: The island and reef sediment classification dataset was tested on three mainstream deep learning frameworks: U-Net, deeplabv3+, and Vision Transformer (ViT). 70% of the dataset was used for training and 30% for testing. The final accuracy performance of each model is shown in Table 2 below. Table 2

[0109] The accuracy comparison results above show that the U-Net model performs best on the island and reef seabed classification dataset created in this embodiment of the invention. Although subsequent models (such as DeepLabv3+ and ViT) perform better in other application scenarios, U-Net remains irreplaceable in certain specific fields, especially in tasks with limited data scale and critical local details.

[0110] In summary, the purpose of this invention is to create a dataset for classifying island and reef substrates that supports deep learning. The technical solution of this invention has the following characteristics: Numerous classification systems exist for island and reef substrates, catering to diverse research objectives and application scenarios. This invention, building upon the characteristics and classification rules of previous island and reef substrate classification systems, proposes a method and standard for efficient and automatic identification of island and reef substrates using remote sensing data. It constructs a classification system encompassing 10 substrate types: "deep sea," "reef pond," "shallow reef foreslope," "deep reef foreslope," "reef ridge," "coral sedimentary area," "living coral," "sandy," "vegetation," and "land." The system considers the spectral characteristics, composition, spatial distribution, and coral reef evolution of each substrate type in remote sensing images. It covers the main types of island and reef substrates and spatially describes the stratification characteristics of each substrate type within the island and reef region.

[0111] The dataset was created using an optimization strategy of "preliminary machine learning classification + expert manual correction." First, the preliminary classification step involved a detailed analysis of the spectral characteristics of each substrate type and the types and patterns of misclassification among them. Second, supervised classification was performed using Support Vector Machines (SVM) to obtain preliminary classification results. Then, the classification results were combined with high-resolution Google Earth imagery and visual interpretation experience to perform pixel-by-pixel corrections for areas with contradictory spatial distributions and densely mixed pixel areas, ultimately resulting in a high-precision classification label dataset. This dataset achieved an average accuracy of 0.989 for each category on the test samples.

[0112] Specifically, this invention reconstructs the traditional classification system by integrating island and reef evolution patterns, spatial stratification features, and remote sensing spectral characteristics, forming a comprehensive island and reef substrate classification dataset suitable for deep learning. This dataset can greatly enrich the data classification resources related to remote islands and reefs. Furthermore, the island and reef substrate classification dataset product constructed based on the method of this invention covers the Xisha Islands (including six major areas such as Huaguang Reef and Panshi Island, with a total area of ​​approximately 1279.2 km² encompassing ocean and islands and reefs), and the dataset achieves a test sample accuracy of 0.989. For substrate types with similar spectra that are difficult to classify (such as reef ponds and foreshore slopes), a classification dataset is obtained through pixel-by-pixel correction based on manual visual interpretation experience.

[0113] Compared with the prior art, the embodiments of the present invention have at least the following beneficial effects: (1) A unified classification system for the bottom sediments of islands and reefs has been established. This system covers the main types of various categories of bottom sediments in terms of content, and can describe the stratification characteristics of each category of bottom sediments in the island and reef area in terms of space.

[0114] (2) Compared with the small number of training samples that support machine learning in the conventional way, this dataset greatly increases the number of training data samples by creating a pixel-by-pixel semantic segmentation training dataset, and supports deep learning training and testing. It also enriches the resources of the Xisha Islands and reefs bottom sediment classification dataset.

[0115] This invention also provides a device for constructing a seabed classification dataset from remote sensing images of islands and reefs, which can implement the above-described method. The device includes: The data acquisition module is used to acquire remote sensing data of islands and reefs; the remote sensing data of islands and reefs adopts remote sensing observations of the target band. The system construction module is used to build a target island and reef bottom classification system based on the prior island and reef bottom classification system and the characteristics of each bottom category. The target island and reef substrate classification system includes substrate categories such as deep sea, reef pond, shallow reef foreslope, deep reef foreslope, reef ridge, coral deposition area, living coral, sandy land, vegetation, and land. The sample acquisition module is used to acquire pre-selected ROI samples from the island and reef remote sensing data; the ROI samples are labeled with the substrate category in the target island and reef substrate classification system; The separation analysis module is used to perform sediment category analysis based on ROI samples to obtain a target sediment category set with low separation. The preliminary classification module is used to perform preliminary classification based on ROI samples using a support vector machine to obtain preliminary classification results. The manual correction module is used to correct the preliminary classification results based on the target substrate category set and in response to the manual correction instructions of the target object, so as to obtain the island and reef substrate classification dataset.

[0116] In some embodiments, the apparatus further includes a sample selection module for performing the following operations: In response to the region selection and division command for the target object, the region of interest is selected in the island and reef remote sensing data, and then the seabed category of the region of interest is labeled to obtain ROI samples.

[0117] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0118] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0119] It is understood that the content of the above method embodiments is applicable to the embodiments of this electronic device. The specific functions implemented by the embodiments of this electronic device are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0120] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0121] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0122] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0123] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0124] The present invention provides a method, apparatus, electronic device, storage medium, and program product for constructing a seabed classification dataset for remote sensing images of islands and reefs. This method involves acquiring remote sensing data of islands and reefs, specifically remote sensing observations of the target band. Based on a priori island and reef seabed classification system, a target island and reef seabed classification system is constructed according to the characteristics of each seabed category. The seabed categories in the target island and reef seabed classification system include deep sea, reef ponds, shallow reef foreslopes, deep reef foreslopes, reef ridges, coral deposition areas, living corals, sandy areas, vegetation, and land. Pre-selected Regions of Interest (ROIs) samples are acquired from the island and reef remote sensing data. These ROI samples are labeled with the seabed category tags from the target island and reef seabed classification system. Seabed category analysis is performed based on the ROI samples to obtain a target seabed category set with low separation. Preliminary classification is performed using a support vector machine based on the ROI samples to obtain preliminary classification results. Based on the target seabed category set, the preliminary classification results are corrected in response to manual correction instructions from the target object to obtain the island and reef seabed classification dataset. This invention addresses the problems of chaotic and inapplicable classification systems in existing technologies by constructing a comprehensive target classification system optimized specifically for remote sensing identification of remote islands and reefs. Preliminary automatic classification using Support Vector Machines (SVM) significantly reduces reliance on costly field surveys, enabling large-scale and efficient extraction of seabed information. To address the unavoidable misclassification problem in automatic classification, an innovative manual correction mechanism based on low-discretionary-degree category sets is introduced. This focuses arduous manual labor on the most challenging areas requiring expert intervention, rather than full map annotation, thereby significantly reducing the cost and difficulty of manual annotation while maintaining classification accuracy. This successfully overcomes the annotation difficulties caused by the intermingled distribution of seabed sediment. This invention provides an effective and feasible technical path for the automated and standardized construction of high-precision seabed classification datasets for remote islands and reefs.

[0125] The embodiments described in this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0126] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. A method for constructing a data set of seabed classification of remote sensing images of islands and reefs, characterized in that, The method comprises the following steps: Obtaining island reef remote sensing data; wherein the island reef remote sensing data adopts remote sensing observation values of target wave bands; Based on the prior island reef bottom classification system, the target island reef bottom classification system is built according to the characteristics of each bottom category; Wherein, the bottom categories of the target island reef bottom classification system include deep sea, reef pond, shallow reef front slope, deep reef front slope, reef ridge, coral deposition area, living coral, sand, vegetation and land; Obtaining the preselected ROI sample in the island reef remote sensing data; wherein the ROI sample is labeled with the label of the bottom category in the target island reef bottom classification system; Based on the ROI sample, the bottom category analysis is carried out to obtain a low separation degree target bottom category set; Based on the ROI sample, the support vector machine is used for preliminary classification to obtain a preliminary classification result; Based on the target bottom category set, the preliminary classification result is corrected in response to the artificial correction instruction of the target object to obtain an island reef bottom classification data set.

2. The method of claim 1, wherein, The method for obtaining island reef remote sensing data comprises the following steps: Downloading the preliminary remote sensing data from a preset remote sensing information website; From the use wave band of the preliminary remote sensing data, the remote sensing observation values of the target wave band are screened to obtain the island reef remote sensing data; Wherein, the target wave band includes blue wave band, green wave band, red wave band and near red wave band.

3. The method of claim 1, wherein, The method for building the target island reef bottom classification system based on the prior island reef bottom classification system according to the characteristics of each bottom category comprises the following steps: Based on the biological nature of the point reef, the point reef is classified into the living coral; Based on the bottom composition, formation conditions and spatial location characteristics of the reef front terrace, the reef front terrace is classified into the reef ridge; Based on the same composition of the ring reef inside edge and the coral deposition area and the high similarity of the spectral curve characteristics, the ring reef inside edge is classified into the coral deposition area; Based on the bottom composition homology and evolution conversion relationship of the sandbank and the sand flat, the sandbank and the sand flat are combined into the sand; According to the living coral, the reef ridge, the coral deposition area and the sand, the target island reef bottom classification system is built in combination with the prior island reef bottom categories of the deep sea, the reef pond, the shallow reef front slope, the deep reef front slope, vegetation and land.

4. The method of claim 1, wherein, The method further comprises the following steps: In response to the region selection and division instruction of the target object, a region of interest is selected in the island reef remote sensing data, and then the region of interest is labeled with the bottom category to obtain the ROI sample.

5. The method of claim 1, wherein, The method for obtaining a low separation degree target bottom category set based on the ROI sample comprises the following steps: Spectral analysis is performed on the ROI sample corresponding to each bottom category to obtain the spectral curve of all bottom categories; From the spectral curve, the spectral characteristics corresponding to each bottom category are extracted, and the JM distance is used to measure the separation degree between different bottom categories based on the spectral characteristics to obtain the separation degree relationship between different bottom categories; Screening the target substrate category set with low separation degree based on the separation degree relationship.

6. The method of claim 1, wherein, The preliminary classification result is obtained by using a support vector machine based on the ROI sample, including the following steps: Constructing a hyperplane of different substrate categories based on the ROI sample corresponding to each substrate category; Based on the hyperplane, the support vector machine is used for interval constraint supervised classification to obtain the optimal decision boundary of each substrate category; Based on the optimal decision boundary, the preliminary classification result is determined.

7. The method of claim 1, wherein, The preliminary classification result is obtained by using a support vector machine based on the ROI sample, including the following steps: Based on the target substrate category set, the preliminary classification result is corrected in response to the artificial correction instruction of the target object to obtain an island reef substrate classification data set, including the following steps: Based on the target substrate category set, the key division reminder is pushed to the target object; The preliminary classification result is a raster data; Each pixel point of the raster data is converted into a vector point type data as a center pixel; An artificial correction instruction of the target object is obtained; 8. An island and reef remote sensing image bottom classification dataset construction device, characterized in that, In response to the artificial correction instruction, the attribute value of the target point data in the vector point type data is adjusted, and then the vector point type data is converted back to the raster data to obtain the island reef substrate classification data set. The device comprises: A data acquisition module is configured to acquire island reef remote sensing data; wherein the island reef remote sensing data adopts remote sensing observation values of target bands; A system building module is configured to build a target island reef substrate classification system based on a prior island reef substrate classification system and according to the feature conditions of each substrate category; The substrate categories of the target island reef substrate classification system include deep sea, reef pond, shallow reef front slope, deep reef front slope, reef ridge, coral deposition area, living coral, sand, vegetation and land; A sample acquisition module is configured to acquire preselected ROI samples in the island reef remote sensing data; wherein the ROI samples are labeled with labels of the substrate categories in the target island reef substrate classification system; A separation analysis module is configured to perform substrate category analysis based on the ROI samples to obtain a target substrate category set with low separation degree; A preliminary classification module is configured to perform preliminary classification by using a support vector machine based on the ROI samples to obtain a preliminary classification result; 9. An electronic device, comprising: An artificial correction module is configured to correct the preliminary classification result in response to an artificial correction instruction of a target object based on the target substrate category set to obtain an island reef substrate classification data set.

10. A computer-readable storage medium, characterized in that, The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method of any one of claims 1 to 7 when executing the computer program. The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1 to 7.

Images