A method and system for identifying a type of multi-rule-constrained offshore net cage culture

Abstract

The application provides a multi-rule-constrained offshore net cage culture type identification method and system, and belongs to the technical field of remote sensing image analysis, interpretation and mapping. The method comprises the following steps: using a net cage culture facility classifier to classify the facility types of net cage culture patches that have been interpreted in a remote sensing image, to obtain a facility type classification result; based on the facility type classification result, combining sample points and water depth environmental characteristics, using a culture type classifier to preliminarily identify the culture biological types of each patch, and outputting a net cage culture type classification result; constructing a multi-rule-constrained aquaculture type identification model, which is used to introduce statistical distribution rules, township distribution rules and water depth ecological suitability rules for multi-rule-constraint, to adjust the net cage culture types of each patch, and form a final net cage culture type identification result. The method can accurately identify the biological types of each patch in the case of lacking direct observation means.

Description

Technical Field

[0001] This application relates to the fields of remote sensing image analysis, interpretation, and mapping, and in particular to a method and system for identifying nearshore cage aquaculture types under multiple rule constraints. Background Technology

[0002] Nearshore cage aquaculture refers to a marine aquaculture method in which artificially constructed net structures (cages) are set up in nearshore waters at a certain distance from the coast to cultivate fish, shellfish, or other aquatic organisms. Quickly and efficiently understanding the spatial distribution and corresponding aquaculture types of nearshore cage aquaculture is beneficial for scientific planning, rational layout, strict control of stocking density, enhanced environmental monitoring, promotion of environmentally friendly technologies, and strict supervision, thereby providing decision-making support for cage aquaculture to effectively address environmental risks and management issues.

[0003] In my country, cage aquaculture encompasses a wide range of marine organisms, including fish, abalone, and sea cucumbers. The environmental impact and ecological benefits of different aquaculture species vary significantly. For example, fish farming typically involves high-density feeding, easily leading to eutrophication in localized sea areas. In contrast, the cultivation of filter-feeding or benthic organisms like abalone and sea cucumbers causes minimal disturbance to water bodies and can even have ecological restoration functions. Therefore, accurately identifying the types of aquaculture organisms is crucial. However, current technologies primarily focus on classifying and identifying cage facilities, failing to provide identification of the specific aquaculture organisms. This makes it difficult to meet the needs for accurate pollution load calculations and precise environmental risk assessments. Furthermore, events such as the escape of invasive species or the spread of diseases (e.g., abalone heatstroke outbreaks) are highly species-specific. If remote sensing interpretation only identifies the presence of a certain type of cage facility but not the actual aquaculture organisms within, it hinders the early deployment of targeted prevention and control measures, severely weakening the intelligence and foresight of marine fisheries governance. In addition, official data such as the China Fisheries Statistical Yearbook usually record detailed information on the output, area and other attributes of different aquaculture (biological) types and facilities, but lack spatial location information.

[0004] From the perspective of existing technologies, the academic and engineering communities mainly use remote sensing interpretation technology, employing methods such as CNN, U-Net, and traditional image processing, to identify the spatial location of net cages or the type of aquaculture facility (such as standard net cages and hanging cages) by utilizing texture, shape, and arrangement patterns. However, remote sensing interpretation can usually only identify the precise spatial distribution of net cage aquaculture areas in large regions. For example, the solution in Chinese patent CN120014401A can classify and extract nearshore aquaculture areas and distinguish between raft aquaculture areas and net cage aquaculture areas. However, since the principle of remote sensing technology is to collect relevant features of the nearshore surface (such as texture, spectrum, shape, etc.), it is difficult to obtain underwater information. Therefore, it is difficult to effectively determine what kind of organisms are actually being farmed underwater (such as fish, abalone, and sea cucumbers).

[0005] In summary, existing remote sensing interpretation data only provides identification at the level of aquaculture facilities and fails to provide detailed spatial distribution information of aquaculture organisms. This makes it impossible to combine remote sensing data with precise location information with statistical data with rich attribute information, thus failing to form a method for quickly obtaining precise location and corresponding aquaculture type of cage aquaculture, which is not conducive to improving the level of intelligent management of cage aquaculture. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for identifying nearshore cage aquaculture types under multiple rule constraints. This addresses the technical problem that existing remote sensing interpretation methods generally stop at facility classification, failing to effectively integrate statistical information and prior knowledge to infer aquaculture (organism) types, resulting in a "last mile" gap between aquaculture identification results and the needs of refined management. This solution, in the absence of direct observation methods, can automatically and accurately identify the aquaculture types in each cage aquaculture patch based on remote sensing interpretation, combined with statistical yearbooks and prior aquaculture knowledge. This provides technical support for improving the informatization level of coastal aquaculture management and for research on related coastal issues.

[0007] To achieve the above objectives, this application provides the following technical solution: This application provides a method for identifying nearshore cage aquaculture types under multiple rule constraints, including: S1. For the interpreted cage aquaculture patches in the remote sensing image, based on spectral features, texture features and shape features, a cage aquaculture facility classifier is used to classify the cage aquaculture patches by facility type, and the cage aquaculture facility type classification results are obtained. S2, based on the facility type classification results, combined with the characteristics of sample points and water depth environment, the aquaculture type classifier is used to preliminarily identify the aquaculture organism type of each patch, and output the cage aquaculture type classification results; S3. Construct a multi-rule constraint aquaculture type discrimination model. The multi-rule constraint aquaculture type discrimination model is used to introduce multiple rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules to adjust the cage aquaculture type of each patch and form the final cage aquaculture type discrimination result.

[0008] Preferably, the classification results of the cage aquaculture facility types include standard cages and hanging cages.

[0009] Preferably, the cage culture types include: fish, standard cage-sea cucumber, hanging cage-sea cucumber, and abalone.

[0010] Preferably, before introducing multi-rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules, the multi-rule constraint aquaculture type discrimination model is also used to initially adjust the cage aquaculture type classification results based on hard constraints. The specific steps are as follows: Based on the facility type classification results, a hard constraint is imposed on the cage aquaculture type classification results output in step S2: if the facility type of a certain cage aquaculture patch is a standard cage according to the facility type classification results, then only the aquaculture organism types of fish and standard cage-sea cucumber are allowed; if the facility type of the cage aquaculture patch is a hanging cage cage, then only the aquaculture organism types of hanging cage-sea cucumber and abalone are allowed. When the predicted category in the cage aquaculture type classification result does not meet the hard constraint with the facility type classification result, its predicted probability is removed from the probability matrix, and the category with the highest probability is selected from the remaining predicted probabilities as the cage aquaculture type prediction result for the cage aquaculture patch. At the same time, the remaining predicted probabilities are normalized to obtain the initial adjustment result that meets the hard constraint.

[0011] Preferably, in step S3, multi-rule constraints are introduced, including statistical distribution rules, township distribution rules, and water depth ecological suitability rules, to adjust the cage aquaculture type of each patch, forming the final cage aquaculture type discrimination result, including: Calculate the scores for statistical area index, township distribution index, and water depth condition index based on the statistical yearbook's area distribution rules, township distribution rules, and water depth ecological suitability rules. The probability values ​​of different aquaculture types in the cage aquaculture type classification results output in step S2 are weighted and summed with the scores of statistical area index, township distribution index, and water depth index to obtain a comprehensive score. The aquaculture type with the highest comprehensive score is taken as the final cage aquaculture type discrimination result.

[0012] Preferably, the calculation steps for the statistical area index score are as follows: Based on statistical yearbooks and cage aquaculture type classification results, the proportion of aquaculture area of ​​each aquaculture type in each sub-region of the study area was calculated, and the target proportion and the current proportion were obtained accordingly. The relative difference between the target ratio and the existing ratio is quantified, and the statistical area index score is calculated using the following formula: , In the formula, This refers to the serial number of the cage aquaculture patch. As a candidate category for cage aquaculture, To score the area indicator, , The first The target proportion and the existing proportion within the sub-region where each patch is located.

[0013] Preferably, the township distribution index score is calculated according to the following steps: based on the spatial differences in the aquaculture distribution of each sub-region within the study area, the correspondence between different aquaculture organism types and each sub-region is determined; Based on the correspondence between different types of farmed organisms and each sub-region, a binary discrimination mechanism is used to calculate the township distribution index score.

[0014] Preferably, the water depth condition index score is calculated according to the following steps: Based on the constraints of biological habits, the water depth conditions corresponding to different types of cultured organisms were determined; A piecewise function is constructed to calculate the water depth condition index score; the expression of the piecewise function is as follows: , In the formula, This refers to the serial number of the cage aquaculture patch. As a candidate category for cage aquaculture, The score is based on the water depth condition index. This represents the water depth.

[0015] Preferably, the classifier for the cage aquaculture facility is a random forest model, and the classifier for the aquaculture type is an XGBoost classification model.

[0016] This embodiment provides a nearshore cage aquaculture type discrimination system with multiple rule constraints. This system is used to execute the nearshore cage aquaculture type discrimination method with multiple rule constraints provided in any of the above embodiments, including: The facility classification unit is configured to classify cage aquaculture patches in remote sensing images based on spectral, texture and shape features using a cage aquaculture facility classifier to obtain cage aquaculture facility type classification results. The preliminary classification unit for biological categories is configured to use an aquaculture type classifier to preliminarily identify the aquaculture biological types of each patch based on the facility type classification results, combined with the characteristics of the sample points and water depth environment, and output the cage aquaculture type classification results. The multi-rule constraint unit is configured to construct a multi-rule constraint aquaculture type discrimination model. The multi-rule constraint aquaculture type discrimination model is used to introduce multi-rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules to adjust the cage aquaculture type of each patch and form the final cage aquaculture type discrimination result.

[0017] In this embodiment, a three-level discrimination system is constructed, consisting of facility classification, preliminary identification of cage aquaculture type, and multi-rule constraints. Based on remote sensing interpretation and identification of nearshore cage aquaculture facilities (such as standard cages or hanging cages), a preliminary discrimination is performed using an aquaculture type classifier. Then, a multi-rule constraint aquaculture type discrimination model is constructed using statistical information and prior aquaculture knowledge to adjust the results of the preliminary discrimination. This accurately identifies the biological aquaculture type (such as fish, abalone, sea cucumber, etc.) corresponding to each cage aquaculture patch, achieving fine-scale discrimination of aquaculture (biological) types. This truly breaks down the barriers between spatial location information and aquaculture attribute information in statistical yearbooks, providing technical support for intelligent supervision of nearshore aquaculture. Attached Figure Description

[0018] Figure 1 This is a technical logic diagram of a nearshore cage aquaculture type discrimination method with multiple rule constraints provided according to some embodiments of this application.

[0019] Figure 2 This is a flowchart illustrating a method for determining nearshore cage aquaculture types with multiple rule constraints, provided according to some embodiments of this application.

[0020] Figure 3 The diagram shows the location and facility types of the study area. (a) is a partial enlarged view of the suspended cage net cage, (b) is a real-life image of the suspended cage net cage, (c) is a partial enlarged view of the standard net cage, and (d) is a real-life image of the standard net cage.

[0021] Figure 4 This is a schematic diagram of the final classification results of cage aquaculture types in the study area. The upper left figure shows the overall classification results. (a), (b), and (c) are magnified views of the three densely populated cage areas, respectively. (d) shows the number and area ratio of various types of aquaculture patches corresponding to cage aquaculture in Sansha Bay. (e) compares the number and area of ​​various types of aquaculture patches in different sub-regions.

[0022] Figure 5 The diagram shows the structure of an electronic device provided for some embodiments of this application. Detailed Implementation

[0023] Nearshore cage aquaculture includes three elements: aquaculture area (i.e., aquaculture location), aquaculture facilities, and aquaculture species. In some existing literature, aquaculture type is usually understood as aquaculture facility type. However, in this application, aquaculture type specifically refers to aquaculture species type, also known as aquaculture organism type. Therefore, although both use the term "aquaculture type" for description, their substantive connotations are different.

[0024] The embodiments of this application will now be described with reference to the accompanying drawings.

[0025] This embodiment provides a method for identifying nearshore cage aquaculture types under multiple rule constraints, such as... Figure 1 , Figure 2 As shown, the method includes the following steps: S1. For the interpreted cage aquaculture patches in the remote sensing image, based on spectral features, texture features and shape features, a cage aquaculture facility classifier is used to classify the cage aquaculture patches by facility type, and the cage aquaculture facility type classification results are obtained.

[0026] Remote sensing imagery refers to images formed by sensors recording electromagnetic wave information from the sea surface in the study area. In this embodiment, the remote sensing imagery is a high-resolution optical imagery covering aquaculture waters, which can be acquired via high-resolution satellites or drones. This embodiment does not limit the acquisition method of the remote sensing imagery.

[0027] A cage aquaculture patch refers to a spatial region unit (i.e., a "patch") with cage aquaculture characteristics extracted from original remote sensing images using remote sensing interpretation methods (such as Object-Oriented Interpretation (OBIA), threshold segmentation, or supervised classification algorithms such as SVM and Random Forest). This patch is a target-level object precise down to each individual cage, rather than a coarse extraction of a large-scale aquaculture area, and the patch has been identified as cage aquaculture land or facility type. The specific implementation steps for interpreting cage aquaculture patches using remote sensing interpretation methods can refer to existing technologies. For example, in this application, the patch-level cage aquaculture area vector data is derived from the research results of Liu et al., whose 2017 cage patch data extracted based on a deep learning RCF model achieved a comprehensive accuracy of 93%, providing precise spatial boundary constraints for the aquaculture area in this embodiment and ensuring the targeting of remote sensing image feature extraction.

[0028] The net cage aquaculture patches have different reflectivity or radiation characteristics in different electromagnetic bands (such as red, green, blue, and near-infrared bands), which are called spectral characteristics.

[0029] Texture features refer to the spatial variation patterns of pixel grayscale or brightness values ​​in remote sensing images, used to describe the uniformity, directionality, and roughness of cage distribution. In this embodiment, texture features may include grayscale co-occurrence matrix (GLCM) statistics, such as energy, contrast, homogeneity, entropy, etc.

[0030] Shape characteristics refer to the spatial geometric properties of net cage aquaculture patches, such as aspect ratio, perimeter-to-area ratio, compactness, and rectangularity, which are used to distinguish net cage facilities of different structural types (such as square, circular, or strip-shaped distribution).

[0031] In this embodiment, the cage aquaculture facility classifier refers to a classification model used to further divide the identified cage aquaculture patches into different facility types. The classifier can be a model based on machine learning or deep learning algorithms (such as random forest, support vector machine or convolutional neural network), with the input being the spectral, texture and shape features of the patch, and the output being the facility type label.

[0032] Preferably, the classifier for cage aquaculture facilities is a Random Forest (RF) model. The Random Forest classifier, by integrating multiple decision trees and employing a Bagging strategy, can effectively reduce model variance and simultaneously handle both continuous and discrete features. It is also insensitive to feature collinearity. Therefore, in this embodiment, based on the vector boundaries of the cage aquaculture patches, the remote sensing image is extracted using a mask, and combined with the constructed feature space, the Random Forest model is used to classify the existing cage patches as facilities.

[0033] The classification result of cage aquaculture facility type refers to the classification output obtained after calculation by the cage aquaculture facility classifier, which reflects the facility type to which each cage aquaculture patch belongs, such as: standard cage, hanging cage. There is a certain correspondence between different facility types and aquaculture types. By classifying cage aquaculture facilities, input features are provided for the aquaculture type classifier in step S2.

[0034] S2, based on the facility type classification results, combined with the characteristics of sample points and water depth environment, uses an aquaculture type classifier to preliminarily identify the aquaculture organism type of each patch, and outputs the cage aquaculture type classification results.

[0035] Among them, the facility type classification result refers to the facility type information of each cage aquaculture patch obtained by the cage aquaculture facility classifier in step S1, that is, the cage aquaculture facility type classification result, which includes two facility types: standard cage and hanging cage.

[0036] In this embodiment, the sample points are spatial point data with known aquaculture organism type labels obtained through field surveys, records from fisheries management departments, or existing databases. Each sample point has geographical coordinates and aquaculture object attributes (such as fish, sea cucumber, abalone, etc.), and can be overlaid with remote sensing images to obtain spectral, texture, and other features at the corresponding location. These features are used to train or validate aquaculture type classifiers. Different aquaculture types (such as fish, sea cucumber, abalone, etc.) differ in water depth environment, facility type, etc. The classifier learns the above features to achieve preliminary identification.

[0037] In this embodiment, the water depth environmental characteristics refer to the hydrological and topographical attributes related to the environmental conditions of the cage aquaculture area, used to reflect the differences in the adaptation of different cultured organisms to the water depth environment. Specifically, the water depth environmental characteristics can be taken as the average water depth of the water area where the cages are located. This characteristic can be obtained by processing water depth spatial distribution data, which can be obtained, for example, from water depth measurement data, digital bathymetry models (DEMs), or multi-source remote sensing extrapolation results. In this embodiment, the marine water depth information is derived from the local interpolation results (resolution approximately 500 meters) of the GEBCO (General Bathymetric Chart of the Oceans) global ocean bathymetry DEM.

[0038] In this embodiment, the aquaculture type classifier is used to identify the classification model of aquaculture organisms in each net cage patch based on multi-source feature information (including: spectral features, texture features, shape features, facility type, sample point information and environmental features). This establishes a nonlinear mapping model from facility type, water depth and environmental features to aquaculture organism type, enabling the identification of aquaculture organisms from remotely visible facilities to remotely invisible ones, and obtaining a preliminary discrimination result based on physical and ecological principles.

[0039] Furthermore, the aquaculture type classifier can specifically employ machine learning algorithms (such as random forest, support vector machine, XGBoost) or deep learning models (such as fully connected neural networks). The input features are spectral features, texture features, shape features, facility type (i.e., the output of step S1) and water depth environment features, and the output is the aquaculture organism category.

[0040] Preferably, the aquaculture type classifier is an XGBoost classification model. The XGBoost classifier iteratively optimizes the loss function through a gradient boosting framework (supporting second-order Taylor expansion), which can significantly improve prediction performance and effectively prevent overfitting. Based on the vector boundaries of the net cage aquaculture patches, the remote sensing image is extracted by masking. Combined with the constructed feature space, the XGBoost model is used to classify the aquaculture (biological) type of the pre-sorted net cage facility patches (i.e., net cage aquaculture patches whose facility types have been distinguished in step S1).

[0041] The aquaculture type classifier is used to make a preliminary identification of the aquaculture organism type. That is, the aquaculture type classifier is used to classify each net cage aquaculture patch for the first time (as opposed to reclassification, i.e., a fine reclassification of net cage aquaculture patches by multi-rule comprehensive judgment). The classification results are used to make a preliminary prediction of the main organism species cultured in the patch.

[0042] Furthermore, considering the differences in the types of cultured organisms between suspended cages and standard cages, in this embodiment, the cage culture types include: fish, standard cage-sea cucumber, suspended cage-sea cucumber, and abalone. That is, the labels for the culture type classifier are: fish, standard cage-sea cucumber, suspended cage-sea cucumber, and abalone. This setting enables the differentiation of different cultured organisms within the same facility, effectively solving the problem of cultured different organisms in the same facility. By establishing a non-linear mapping between facility type and water depth environmental characteristics and the above labels, a preliminary and refined classification of culture types is achieved.

[0043] It should also be noted that the output of step S2 is the preliminary aquaculture organism type result (i.e., the aquaculture type classification result) for each net cage aquaculture patch. This result includes not only the predicted aquaculture type, but also the probability of each net cage aquaculture patch belonging to each aquaculture organism category. Since each patch outputs multiple probability values, the corresponding prediction probability matrix can be obtained by using the predicted aquaculture organism category as the column and the patch as the row. This output provides initial data for the final multi-rule comprehensive discrimination of aquaculture type.

[0044] S3. Construct a multi-rule constraint aquaculture type discrimination model. The multi-rule constraint aquaculture type discrimination model is used to introduce multiple rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules to adjust the cage aquaculture type of each patch and form the final cage aquaculture type discrimination result.

[0045] In this embodiment, the Multi-Rule Constrained Aquaculture Type Discrimination Model (MRC-ATDM) refers to a model that, based on the preliminary classification results, incorporates various external rule information (including statistical yearbooks, aquaculture planning, water depth suitability, etc.) to comprehensively correct and optimize the classification results. By constructing a multi-rule constraint system, the aquaculture organism types in each patch are adjusted and optimized, thereby achieving accurate identification of underwater aquaculture organism types in cage aquaculture.

[0046] In a further improved scheme, before introducing multi-rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules, the multi-rule constraint aquaculture type discrimination model is also used to initially adjust the cage aquaculture type classification results based on hard constraints. The specific steps of the hard constraints are as follows: S91, apply hard constraints to the cage aquaculture type classification results output in step S2 based on the facility type classification results: if the facility type of a certain cage aquaculture patch is a standard cage according to the facility type classification results, then only the aquaculture organism types of fish and standard cage-sea cucumber are allowed; if the facility type of the cage aquaculture patch is a hanging cage cage, then only the aquaculture organism types of hanging cage-sea cucumber and abalone are allowed.

[0047] It should be noted that, here, "hard constraints" refer to constraints based on the deterministic correspondence between facility types and aquaculture types, used to forcibly correct illogical criteria in the preliminary classification results. In this embodiment, hard constraints are determined based on prior knowledge of aquaculture and are considered preliminary rules, possessing the characteristic of "certain prohibitions," used to exclude obviously erroneous or unreasonable combinations.

[0048] Furthermore, the correspondence between facility type and aquaculture type is as follows: When the facility type is a standard net cage, the permitted types of cultured organisms are fish or standard net cage-sea cucumbers; When the facility type is a hanging cage, the permitted aquaculture species are hanging cage-sea cucumber or abalone.

[0049] The above correspondence between facility types and aquaculture types is derived from prior knowledge of the structural characteristics of aquaculture facilities and the habits of different aquaculture organisms, reflecting the applicable conditions for different facilities to be suitable for different aquaculture organisms.

[0050] In this embodiment, the hard constraint is executed before the multi-rule constraint stage. As a preliminary correction mechanism, it can effectively reduce the conflict and error accumulation in the subsequent statistical distribution, township distribution and environmental ecological constraint process by eliminating or correcting samples in the preliminary judgment results that do not conform to the correspondence between facility type and aquaculture type.

[0051] S92, when the predicted category in the cage aquaculture type classification result and the facility type in the facility type classification result do not meet the hard constraints, their predicted probabilities are removed from the probability matrix, and the category with the highest probability is selected from the remaining predicted probabilities as the cage aquaculture type prediction result for the cage aquaculture patch. At the same time, the remaining predicted probabilities are normalized to obtain the initial adjustment result that meets the hard constraints.

[0052] In step S92, based on the establishment of hard constraint rules, the model execution mechanism and algorithm logic are further clarified, namely, at the probability matrix level of the prediction results, categories that do not meet the constraints are eliminated, renormalized, and the categories are redefined.

[0053] As mentioned earlier, the probability matrix refers to the set of predicted probabilities of each aquaculture category output by the aquaculture type classifier when it identifies the aquaculture type of each net cage aquaculture patch. Each row of the matrix corresponds to a patch, and each column corresponds to a possible aquaculture type category (also known as a candidate category). The matrix elements represent the predicted probability that the patch belongs to a certain category.

[0054] In this embodiment, eliminating prediction categories that do not meet the hard constraints means that when the prediction category of a patch does not meet the logical correspondence between its facility type and the prediction category defined in S91, the model will set the prediction probability of that category to invalid (e.g., set it to 0 or delete it directly) in the probability matrix to exclude unreasonable type results.

[0055] After removing categories that do not meet the hard constraints, the predicted probabilities of the remaining categories are renormalized so that their sum is 1, ensuring that the probability distribution remains valid and can be used for subsequent rule constraints or final classification. From the normalized remaining predicted probabilities, the category with the highest probability is selected as the new prediction result for that patch, thus obtaining the initial adjustment result that meets the hard constraints.

[0056] For example, for a given patch, the Random Forest model identifies the facility as a standard net cage. The XGBoost model predicts four categories: fish, standard net cage-sea cucumber, hanging cage-sea cucumber, and abalone, with corresponding probabilities of [0.4, 0.3, 0.2, 0.1]. Since standard net cages only allow the cultured organisms to be either fish or standard net cage-sea cucumber, the probabilities 0.2 and 0.1 corresponding to hanging cage-sea cucumber and abalone are removed from the probability matrix (e.g., set to 0). The resulting probabilities are [0.4, 0.3, 0, 0]. Renormalizing these probabilities yields new normalized probabilities of [0.57, 0.43, 0, 0]. Among the remaining probabilities, the highest is 0.57, corresponding to the fish category. Therefore, the new prediction is fish (since this is the same as the original prediction from the XGBoost model, the category remains unchanged; only the probability value is adjusted). In another example, if the Random Forest model identifies a facility as a hanging net cage in a certain patch, the XGBoost model predicts four categories: fish, standard net cage-sea cucumber, hanging net cage-sea cucumber, and abalone, with probabilities of [0.4, 0.3, 0.2, 0.1]. Since hanging net cages only allow the cultured organisms to be either hanging net cage-sea cucumber or abalone, the probabilities for fish and standard net cage-sea cucumber need to be removed. After removal, the remaining probabilities are [0, 0, 0.2, 0.1]. Renormalization yields [0.0, 0.0, 0.67, 0.33]. Among the remaining probabilities, the highest probability is 0.67, corresponding to the hanging net cage-sea cucumber category. Therefore, the prediction is adjusted from fish to hanging net cage-sea cucumber.

[0057] The above steps ensure that the prediction results for each patch simultaneously satisfy the model's classification logic and hard constraints within the probability space, providing a consistent and reliable input foundation for subsequent multi-rule constraint optimization. This initial adjustment result is then input into the multi-rule constraint logic, where comprehensive optimization is performed using statistical distribution rules, township distribution rules, and water depth ecological suitability rules. Ultimately, a cage aquaculture type determination result that conforms to reality in terms of logical consistency, spatial rationality, and ecological suitability is obtained.

[0058] Among them, the statistical distribution rule is an area constraint rule established based on the proportion of the total aquaculture area of ​​each type of aquaculture (such as fish, shrimp, and shellfish) within the administrative region in the official fisheries statistical yearbook (such as the "China Fisheries Statistical Yearbook" or various provincial statistical yearbooks). This rule is used to constrain the statistical distribution of the classification results to be consistent with the macro statistical data, and to prevent the model from over-identifying or under-identifying a certain type of aquaculture.

[0059] Township distribution rules are spatial distribution constraints established based on the actual aquaculture distribution characteristics of different townships. These rules reflect the spatial heterogeneity of aquaculture types in each township and can constrain the spatial distribution characteristics of aquaculture organisms in different townships, thus guiding the model's category adjustment in different regions.

[0060] The water depth ecological suitability rule is an ecological constraint rule established based on the ecological characteristics of different aquaculture organisms and their suitable range for water depth environment, to ensure the ecological rationality of the classification results.

[0061] Multi-rule constraints refer to the comprehensive application of three types of rules—area distribution from statistical yearbooks, township distribution, and water depth ecological suitability—to the preliminary classification results. This adjusts the aquaculture (biological) types of each patch, thereby obtaining the final cage aquaculture type identification result under the multi-rule constraints, and ensuring that the aquaculture biological types corresponding to each patch in the result conform to statistical laws, spatial distribution, and ecological conditions.

[0062] In a further improved scheme, step S3 introduces multi-rule constraints, including statistical distribution rules, township distribution rules, and water depth ecological suitability rules, to adjust the cage aquaculture type for each patch, forming the final cage aquaculture type discrimination result, including: S31. Calculate the scores for statistical area indicators, township distribution indicators, and water depth condition indicators according to the statistical distribution rules, township distribution rules, and water depth ecological suitability rules, respectively.

[0063] S32, the probability values ​​of different aquaculture types in the cage aquaculture type classification results output in step S2 are weighted and summed with the scores of statistical area index, township distribution index, and water depth condition index to obtain a comprehensive score, and the aquaculture type with the highest comprehensive score is taken as the final cage aquaculture type discrimination result.

[0064] In this embodiment, based on the introduction of statistical distribution rules, township distribution rules, and water depth ecological suitability rules, a multi-index weighted fusion mechanism based on strict mathematical definitions is constructed. The indicators include four categories: model probability indicators, statistical area indicators, township distribution indicators, and water depth condition indicators. The model probability indicator score uses the initial adjustment result obtained through hard constraints, i.e., normalized predicted probability; the statistical area indicator score uses an exponential function to characterize the deviation between the target ratio and the existing ratio; the township distribution indicator score uses a binary discriminant mechanism to reflect spatial distribution matching; and the water depth condition indicator score constructs a piecewise function based on the deviation between the water depth of the patch and the target water depth.

[0065] In the multi-rule constrained aquaculture type discrimination model provided in this embodiment, the statistical area index score is used to balance the deviation between the preliminary discrimination result of the XGBoost model and the target proportion in the statistical yearbook at the regional scale, so as to ensure that the final aquaculture type discrimination result conforms to macroscopic statistical laws. Specifically, the calculation steps of the statistical area index score are as follows: Based on statistical yearbooks and classification results of cage aquaculture types, the proportion of aquaculture area for each aquaculture type (such as fish and sea cucumber) in each sub-region of the study area was calculated, and the target proportion and the current proportion were obtained accordingly.

[0066] The target proportion, extracted from statistical yearbooks, represents the desired state of each aquaculture type (e.g., fish, sea cucumber, abalone) within each sub-region (e.g., different townships). The existing proportion refers to the area proportion predicted by the aquaculture type classifier within the same sub-region after initial classification and hard constraint adjustment. It should also be noted that in this embodiment, the area proportion uses a relative proportion, that is, the ratio of the aquaculture area of ​​any two aquaculture types.

[0067] The relative difference between the target ratio and the existing ratio is quantified, and the statistical area index score is calculated using the following formula: , In the formula, This refers to the serial number of the cage aquaculture patch, such as 1~5507. Candidate categories for cage aquaculture types include, for example, fish in standard cages for sea cucumbers, or abalone in hanging cages for sea cucumbers. To score the area indicator, , The first The target and current proportions of any two aquaculture types (such as fish and sea cucumber) within a sub-region (such as a district or county) where a patch is located. "Meets the target" means that the current proportion has reached or exceeded the target proportion, while "does not meet the target" means that the current proportion is lower than the target proportion.

[0068] For example, for a specific sub-region, the area of ​​fish and sea cucumber farming is extracted from statistical yearbooks, and the ratio of fish to sea cucumber farming area is calculated to obtain the target ratio. Assuming the calculated result is 1.2, then 1.2 is the macroscopic expected ratio (i.e., the target ratio) of fish to sea cucumber farming in this sub-region. After preliminary identification by the farming type classifier and adjustment for hard constraints, the total area of ​​all fish patches and all sea cucumber patches (i.e., the sum of patches consisting of standard net cages and hanging cages) in this sub-region is calculated. For example, if the statistical results are 30 hectares and 50 hectares respectively, then the current fish to sea cucumber area ratio is (30 hectares / 50 hectares = 0.6), which is lower than the target ratio of 1.2, belonging to the non-compliant category. At this point, the statistical area index score of each patch in this sub-region is calculated using the above formula. This increases the likelihood that patches predicted to be sea cucumbers within the sub-region will eventually transform into fish.

[0069] Based on the above definition, the statistical area index score... At a macro scale, it plays a role in adjusting the predictive structure of the model. When the predicted area of ​​a certain type of aquaculture in a sub-region (such as a specific administrative region) is lower than the target proportion in the statistical yearbook, the category is marked as "not meeting the standard". Its statistical area index score increases exponentially with the increase of deviation, thereby improving the overall score of the type in the subsequent weighted fusion. Conversely, if a certain type of aquaculture has reached or exceeded the target proportion, the score of the category is zero, and its influence weight in the overall score is no longer increased.

[0070] For example, when the predicted area of ​​fish relative to the predicted area of ​​sea cucumber in a certain county is only 0.6, while the target ratio for the county is 1.2 according to the statistical yearbook, the fish type is considered "not up to standard". The system calculates a higher value, thereby increasing the overall weight of fish in the subsequent integration stage, prompting some sea cucumber patches to be adjusted to fish categories, and achieving dynamic rebalancing based on statistical laws.

[0071] Considering the significant spatial heterogeneity of aquaculture distribution across different sub-regions, the multi-rule constrained aquaculture type discrimination model provided in this embodiment introduces township distribution index scores to guide the classification results towards categories that align with local aquaculture planning directions, thereby achieving consistency with actual aquaculture distribution patterns at a spatial scale. Specifically, the township distribution index scores are calculated as follows: based on the spatial differences in aquaculture distribution across various sub-regions within the study area, the correspondence between different aquaculture organism types and each sub-region is determined; based on the correspondence between different aquaculture organism types and each sub-region, a binary discrimination mechanism is used to calculate the township distribution index scores.

[0072] Among them, the spatial difference in the distribution of aquaculture in each sub-region refers to whether a certain type of aquaculture organism exists in a certain township. The mapping relationship (i.e., the correspondence relationship) between different aquaculture types and each township is established, and this correspondence relationship constitutes the basic rule for spatial matching judgment.

[0073] In the process of calculating the township distribution index score, the model adopts a binary discrimination mechanism, that is, by comparing the spatial location of the current patch with the correspondence between the aquaculture type and the aforementioned mapping relationship table, the spatial matching is judged: if the aquaculture type of the patch conforms to the mapping relationship, it is considered that the preliminary discrimination result of the aquaculture type conforms to the township distribution rules (covering the prescribed categories); if it does not conform (not covering the prescribed categories), it is considered that there is a deviation.

[0074] Specifically, the formula for calculating the township distribution index score is as follows: , In the formula, The township distribution index is scored.

[0075] By using a spatial guidance mechanism based on the township scale, the consistency of the aquaculture type identification results in terms of regional distribution is effectively enhanced, making the final classification results more consistent with the actual aquaculture planning zoning characteristics in terms of geographical space.

[0076] Given the significantly different ecological adaptability of different aquaculture organisms to seawater depth conditions, this embodiment introduces a water depth condition index score into a multi-rule-constrained aquaculture type discrimination model to ecologically constrain the preliminary discrimination results from an environmental suitability perspective. Specifically, the water depth condition index score is calculated as follows: based on biological habit constraints, the water depth conditions corresponding to each different aquaculture organism type are determined; a piecewise function is constructed to calculate the water depth condition index score; the expression of the piecewise function is as follows: , In the formula, This refers to the serial number of the cage aquaculture patch. As a candidate category for cage aquaculture, The score is based on the water depth condition index. This represents the water depth.

[0077] This piecewise function, based on the ecological habits of typical aquaculture organisms, transforms water depth suitability into a computable constraint standard, applying it to each patch according to its location and water depth. Target calculate, It is used to measure the deviation of the current patch depth from the target depth (e.g., 15 meters), and dynamically adjusts different depth ranges in the form of a piecewise function.

[0078] After calculating the scores for the statistical area index, the township distribution index, and the water depth index, in step S32, the probability values ​​of different aquaculture types in the cage aquaculture type classification results output in step S2 are weighted and summed with the scores for the statistical area index, the township distribution index, and the water depth index to obtain a comprehensive score. The aquaculture type with the highest comprehensive score is then used as the final cage aquaculture type discrimination result.

[0079] The formula for calculating the overall score is as follows: , In the formula, For the overall score, The normalized predicted probability values ​​of different aquaculture types in the cage aquaculture type classification results output in step S2, after hard constraints. , , , Corresponding to , , , The weights of each indicator can be reasonably determined based on experience. For example, they can be set as follows: , , , .

[0080] In summary, the method provided in this embodiment accurately identifies the biological aquaculture type corresponding to each cage aquaculture patch through a cage aquaculture facility classifier, an aquaculture type classifier, and a multi-rule constraint aquaculture type discrimination model. It not only retains the efficient predictive ability of the data-driven model, but also achieves deep embedding of domain knowledge through multi-rule constraints, thereby improving the spatial accuracy of aquaculture type identification. Furthermore, it enhances the traceability of model decision-making through a transparent comprehensive scoring mechanism.

[0081] This scheme uses remote sensing image interpretation to obtain net cage aquaculture patches, and integrates remote sensing image features, statistical yearbook information and prior aquaculture knowledge to automatically determine the type of net cage aquaculture. It breaks down the barriers between detailed location information and rich attribute information in statistical yearbooks, and plays an important role in improving the informatization level of coastal marine aquaculture management and research on coastal issues.

[0082] It should also be noted that some traditional studies obtain the distribution of farmed organisms by simply equating the type of aquaculture facility with the type of farmed organism. However, although there is a correlation between aquaculture facilities and types of farmed organisms, they are not in a one-to-one correspondence. For example, hanging cages can be used to farm both abalone and sea cucumbers; standard cages may be used to farm large yellow croaker, grouper, or sea cucumbers in different sea areas. Therefore, simply establishing a one-to-one correspondence between aquaculture facilities and types of farmed organisms to obtain information on the types of farmed organisms cannot guarantee the consistency between macroscopic data and microscopic distribution, nor can it take into account the spatial heterogeneity of farmed organisms between different towns, resulting in a significant deviation between the final aquaculture type identification result and the actual distribution of farmed organism types. The technical solution of this application breaks away from the traditional one-to-one mapping approach between aquaculture facilities and types of farmed organisms. It generates preliminary identification probabilities through an aquaculture type classifier and introduces multi-rule constraints to correct the probabilities of the identification results. The final identification results can ensure that the macroscopic data and microscopic distribution logic are consistent, the spatial distribution of towns is more reasonable and conforms to ecological constraints, and significantly improves the accuracy and precision of aquaculture type identification.

[0083] Figure 1 This is a technical logic diagram of a nearshore cage aquaculture type discrimination method with multiple rule constraints provided according to some embodiments of this application. Figure 1 As shown, as an example, the method provided in this embodiment can be executed according to the following exemplary steps: Step 1 is to classify the facilities for cage aquaculture. Cage aquaculture facilities can be divided into standard cages and hanging cages. Different facility types have a certain correspondence with the types of aquaculture organisms. Based on the spectral and texture features of high-resolution remote sensing images, the facility types are first identified. Step two involves the preliminary identification of cage aquaculture types, using sample points from different aquaculture (biological) types to make a preliminary identification of the aquaculture type of the aquaculture facilities; Step three involves a comprehensive assessment of aquaculture types. Based on the preliminary assessment in step two, the aquaculture types are further adjusted by combining statistical yearbook data, township aquaculture type planning and zoning, and water depth spatial distribution data with corresponding information on aquaculture types, to form the final type assessment result.

[0084] The following section uses Sansha Bay in Fujian Province as a case study area (i.e., the research area) to explain the details of each of the above three steps in more detail.

[0085] It should be noted that the method provided in this embodiment is based on the completion of the interpretation of remote sensing image cage facility patches. The method of remote sensing patch interpretation can be performed with reference to existing technology, and will not be described in detail in this embodiment.

[0086] Sansha Bay is a famous mariculture bay in my country, with a large number of cage aquaculture facilities, which are classified into hanging cages and standard cages according to the type of facilities. Figure 3 The diagram shows the Sansha Bay (i.e., the study area) and the types of facilities. (a) is a partial enlarged view of the suspended cage net cage, (b) is a real-life view of the suspended cage net cage, (c) is a partial enlarged view of the standard net cage, and (d) is a real-life view of the standard net cage.

[0087] Step 1: Classification of RF-based cage aquaculture facilities.

[0088] The main types of cage aquaculture in Sansha Bay include fish, sea cucumber, and abalone. The types of cage facilities include suspended cages and standard cages. There are some differences in the types of crops cultured in suspended cages and standard cages. Specifically: Hanging cages are mainly used for abalone and sea cucumber farming. For example... Figure 3 As shown in (a) and (b), the hanging cages are first assembled into a fish raft structure similar to standard cages, with dimensions of 3.0m×3.0m×3.0m, 4.0m×4.0m×4.0m, or 5.0m×5.0m×5.0m, etc., and then divided in parallel by bamboo poles, with a distance of about 0.6m between the poles. The hanging cages, containing abalone or sea cucumbers, are suspended on the dividing horizontal poles, with a distance of about 0.6m between the cages. In remote sensing images, this type of facility appears as a densely arranged combination of narrow units, with the texture complexity significantly increased due to the hanging cages and the bamboo poles that suspend them.

[0089] Standard net cages are mainly used for farming fish and sea cucumbers. For example... Figure 3 As shown in (c) and (d), the standard net cage is mainly constructed using plastic fishing rafts as the basic unit. The size of the net cage varies from 4.0m×4.0m to 6.0m×6.0m. Several net cages are connected to form a net cage panel, and the area of ​​each net cage panel does not exceed 3000m². 2 A cage area consists of no more than 20 cage panels. In remote sensing images, this type of facility appears as a rectangular array composed of multiple regular small squares, with clear unit boundaries and smooth texture.

[0090] To achieve precise classification of cage aquaculture facilities and types, a classification model was constructed based on multi-dimensional features. In terms of feature system construction, multi-source indicators were selected from five dimensions: texture features, shape features, spectral features, environmental features, and facility features, to construct classifiers for cage aquaculture facilities and aquaculture types, respectively.

[0091] (a) Texture features further include: texture mean Texture standard deviation Energy E, entropy H, contrast C, inverse difference moment Autocorrelation .

[0092] (1) Texture mean The calculation formula is: , In the formula, , These are the row and column indices of the Gray-Level Co-occurrence Matrix (GLCM), respectively. For line numbers, For column number, The number of gray levels in the image. In GLCM ( The joint probability of the position.

[0093] (2) Texture standard deviation The calculation formula is: , In the formula, The mean texture value is calculated above.

[0094] (3) The formula for calculating energy E is: , (4) The formula for calculating entropy H is: , (5) The formula for calculating the contrast ratio C is: , (6) Inverse difference moment The calculation formula is: , (7) Autocorrelation The calculation formula is: , In the formula, This represents the marginal mean along the GLCM row direction. This represents the marginal mean along the column direction of GLCM. This represents the marginal standard deviation along the GLCM row direction; This represents the marginal standard deviation along the column direction of GLCM.

[0095] (ii) Shape characteristics include: area (i.e. polygon area) S, perimeter (total length of polygon boundary) L, aspect ratio AR, compactness index CI, and shape index SI.

[0096] The formula for calculating the aspect ratio AR is as follows: , The formula for calculating the Firmness Index (CI) is: , The formula for calculating the shape index SI is: , In the formula, This represents the length of the longest side of the smallest bounding rectangle that can completely enclose a certain aquaculture vector patch. This represents the length of the shorter side of the smallest bounding rectangle that can completely enclose a certain aquaculture vector patch. This represents the perimeter of a certain aquaculture vector patch. This represents the area of ​​a certain aquaculture vector patch.

[0097] (iii) Spectral characteristics include: spectral average value Spectral standard deviation .

[0098] spectral average The calculation formula is: , In the formula, This represents all valid pixels within the remote sensing image region corresponding to each vector patch (polygon), the i-th The first band The reflectance value of each pixel. The total number of all valid pixels within the image area corresponding to each vector patch (polygon).

[0099] Spectral standard deviation The calculation formula is: , For each vector patch in the first The spectral average value across each band.

[0100] (iv) Environmental characteristics include: water depth The calculation formula is as follows: , In the formula, M represents the total number of valid pixels within the depth sounding raster data area corresponding to each vector patch (polygon). This represents the water depth value of each valid pixel within the depth measurement raster data area corresponding to each vector patch (polygon).

[0101] (v) Facility characteristics include: Facility category Its expression is: , In the formula, This represents the probability prediction function of the random forest classifier when classifying aquaculture facilities, where c represents the facility category. {"Standard wire mesh cage", "Suspended wire mesh cage"} This represents the feature vector input to the random forest classifier, including texture features, spectral features, and shape features.

[0102] Combining field investigation records and manual visual interpretation, samples of two types of cage aquaculture facilities, namely standard cages and hanging cages, were created from vector patches of cage aquaculture in nine high-resolution remote sensing images. Finally, 100 samples of each type of aquaculture facility were selected.

[0103] Step 2: Preliminary identification of cage aquaculture type based on XGBoost.

[0104] Based on the differentiation of cage aquaculture facility types, further samples of aquaculture types were selected, and the XGBoost model was used to achieve preliminary identification of cage aquaculture types. 150 fish samples, 50 standard cage-sea cucumber samples, 70 hanging cage-sea cucumber samples, and 60 abalone samples were selected.

[0105] Step 3: Comprehensive identification of cage aquaculture type based on MRC-ATDM.

[0106] The Rule-Constrained Aquaculture Type Discrimination Model (MRC-ATDM) constructs a comprehensive weighted scoring decision framework by integrating a machine learning model driven by remote sensing data with business rules driven by domain knowledge. This model uses the inherent correspondence between facility type and aquaculture type as a hard constraint, limiting standard cage facilities to fish and sea cucumber aquaculture, and hanging cage facilities to sea cucumber and abalone aquaculture. It discriminates the predicted categories and their probability matrices from the XGBoost model in step two. When the initial prediction result of the XGBoost model does not match the facility type, the MRC-ATDM model forcibly converts the prediction result to the candidate category with the highest probability based on the probability matrix, and normalizes the probability distribution of the candidate categories to form a baseline probability matrix that conforms to the domain rules.

[0107] Among them, the business rules driven by domain knowledge, also known as prior constraint rules, are exemplified by introducing three types of prior constraint rules based on prior knowledge such as survey data of the study area and statistical yearbook data: statistical area rules, township distribution rules, and water depth condition rules. These rules apply multi-rule constraints to the output of the XGBoost model after hard constraints, as detailed below: (1) Statistical Area Rule: According to the Fujian Statistical Yearbook, the proportion of marine aquaculture area for fish and sea cucumber in Jiaocheng District of Sansha Bay is about 63.2%, the proportion in Fu'an City is about 23.6%, and the proportion in Xiapu County is about 3.2%. These figures are used as the target proportions. The core of the statistical area rule is to adjust the preliminary judgment results by utilizing the deviation between the existing proportions and the target proportions.

[0108] (2) Township distribution rules: Township aquaculture distribution shows significant spatial heterogeneity. Sea cucumbers are distributed in Xinan, Xiahu, Shajiang, Changchun and Beibi. Fish are distributed in Xinan, Xiahu, Yantian, Beibi, Changchun and Feiluan. Abalone are distributed in Beibi, Xinan, Xiahu and Jianjiang.

[0109] (3) Water depth conditions rule: Based on the constraints of biological habits, abalone farming must meet the environmental conditions of water depth exceeding 15 meters.

[0110] The above rules are transformed into calculable discrimination criteria through four types of indicators, including model probability indicators, statistical area indicators, township distribution indicators, and water depth condition indicators.

[0111] The calculation process for each indicator follows strict mathematical definitions, as follows: (1) Model probability index score We directly use normalized prediction probabilities, with a weighting of 50%, and the expression is: , In the formula, This represents the patch number, specifically the 1-5507 net cage aquaculture patches in Sansha Bay. The representative candidate categories are either fish and standard cage-sea cucumber, or abalone and hanging cage-sea cucumber. For the first Candidate categories for plaques Normalized prediction probability.

[0112] (2) Statistical area index score By quantitatively analyzing the relative differences between the target and current proportions of fish and sea cucumbers in sub-regions under the jurisdiction of Sansha Bay, such as Jiaocheng District, Fu'an City, and Xiapu County, an exponential function is constructed with a weight of 16.7%, and its expression is as follows: , (3) Township distribution indicators The scoring employs a binary discrimination mechanism. Based on the distribution rules of townships for each type, candidate categories that conform to the spatial distribution rules are assigned 0 points, while candidate categories that do not conform to the spatial distribution rules are assigned 1 point, with a weight of 16.7%. The expression is as follows: , (4) Water depth condition index score Then, based on the type of facility and the water depth of the patch. The piecewise function is dynamically adjusted; the greater the absolute deviation from the target water depth of 15 meters, the higher the score, with a weight of 16.7%. The expression is: , Finally, a linear weighted fusion mechanism is used to synthesize the above four categories of indicators to calculate the comprehensive score for each candidate category. : , The maximum value is selected as the final aquaculture type for the patch. This model retains the efficient predictive ability of data-driven models, achieves deep embedding of domain knowledge through multi-rule constraints, improves the spatial accuracy of aquaculture type identification, and enhances the traceability of model decisions through a transparent scoring mechanism.

[0113] Furthermore, the method also includes an accuracy verification step, which involves introducing a control experiment, using the XGBoost model to classify the types of organisms in net cage aquaculture without imposing multi-rule constraints, comparing the results of the control experiment with the classification results of MRC-ATDM, and performing accuracy verification according to the following steps: The confusion matrix was used to evaluate classification performance, and metrics such as precision (also known as user precision), recall (also known as producer precision), F1 score, overall precision, and Kappa coefficient were calculated. Precision and recall reflect the classifier's ability to predict and identify a particular class, respectively, while the F1 score, overall precision, and Kappa coefficient measure the classifier's integration performance. The calculation formulas for each metric can be found in existing techniques and will not be elaborated here. The calculation results are shown in Table 1. Table 1. Accuracy verification results of classification models for different cage aquaculture types

[0114] The data above shows that, without imposing multiple rule constraints, the XGBoost model demonstrates robust discriminative ability in the four categories (fish, abalone, standard cage-sea cucumber, and hanging cage-sea cucumber), achieving an overall classification accuracy of 89.6% and a Kappa coefficient of 0.85. Among these, the abalone category performed best, with recall and precision reaching 92.2% and 95.7% respectively, and an F1 score as high as 0.93. The hanging cage-sea cucumber category performed second best, showing a good balance between precision (92.7%) and recall (90.0%), with an F1 score of 0.90.

[0115] The Multi-Rule Constraint Aquaculture Type Discrimination Model (MRC-ATDM) achieved the best classification performance, with an overall accuracy of 93.5% and a Kappa coefficient of 0.89 compared to the aforementioned XGBoost model without multi-rule constraints. This model significantly improved classification accuracy across four target categories, particularly for standard net cage sea cucumbers, where recall increased to 89.1% and precision to 90.2%. This demonstrates that while the XGBoost model retains complete features, it lacks domain knowledge guidance. In contrast, MRC-ATDM, by systematically integrating data-driven methods and domain knowledge rules to construct a multi-index weighted scoring system, achieves automated discrimination of aquaculture types, overcoming the domain adaptability limitations of single data-driven models.

[0116] Figure 4 The final results of the cage aquaculture type determination are shown, such as... Figure 4 As shown in (d) and (e), the final proportion of various types of aquaculture area, the number of various types of aquaculture patches in different sub-regions, and the comparison of patch areas can be calculated based on the results of cage aquaculture type identification.

[0117] In summary, this embodiment proposes a multi-rule constraint for nearshore cage aquaculture type discrimination, which further enables the discrimination of aquaculture types in the interpreted cage aquaculture patches. It connects the fine spatial location information of remote sensing interpretation data with the rich attribute information of statistical data, providing data methodological support for the intelligent management of nearshore cage aquaculture in my country.

[0118] Based on the facility type classification results, sample point information and key environmental factors such as water depth are introduced to construct an aquaculture type classifier (i.e., XGBoost). This classifier supplements traditional remote sensing features by utilizing the correspondence between facility types and the aquaculture habits of different organisms, enhancing the ability to distinguish which organisms are being aquacultured underwater. The Multi-Rule Constraint Aquaculture Type Discrimination Model (MRC-ATDM) simultaneously incorporates statistical distribution rules, township distribution rules, and water depth ecological suitability rules. This allows for consistency verification between the individual patches and discrete inference results obtained from XGBoost and the statistical patterns, aquaculture structures, and aquaculture experience of the actual area. The combined effect of these three types of rules ensures that the discrimination result of each patch not only depends on its own characteristics but also remains consistent with the statistical and ecological constraints of the regional aquaculture pattern, further reducing the probability of misclassification and improving the ecological rationality and spatial consistency of the overall classification results.

[0119] Based on the same concept, this embodiment also provides a nearshore cage aquaculture type discrimination system with multiple rule constraints. This system is used to execute the steps of the nearshore cage aquaculture type discrimination method with multiple rule constraints provided in any of the above embodiments, including: The facility classification unit is configured to classify cage aquaculture patches in remote sensing images based on spectral, texture and shape features using a cage aquaculture facility classifier to obtain cage aquaculture facility type classification results. The preliminary classification unit for biological categories is configured to use an aquaculture type classifier to preliminarily identify the aquaculture biological types of each patch based on the facility type classification results, combined with the characteristics of the sample points and water depth environment, and output the cage aquaculture type classification results. The multi-rule constraint unit is configured to construct a multi-rule constraint aquaculture type discrimination model. The multi-rule constraint aquaculture type discrimination model is used to introduce multi-rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules to adjust the cage aquaculture type of each patch and form the final cage aquaculture type discrimination result.

[0120] The nearshore cage aquaculture type discrimination system with multiple rules constraints provided in this embodiment can realize the steps and processes of the nearshore cage aquaculture type discrimination method with multiple rules constraints provided in any of the above embodiments, and achieve the same technical effect, which will not be described in detail here.

[0121] The nearshore cage aquaculture type discrimination method with multiple rule constraints provided in this application embodiment can be applied to... Figure 5 The electronic devices shown can be cloud devices or edge devices. Specifically, they can be, but are not limited to, mobile terminals such as mobile phones, tablets, handheld computers, and personal digital assistants (PDAs), smart home devices such as smart TVs and smart cameras, wearable devices such as smart bracelets, smartwatches, and smart glasses, or other computer devices such as desktop, laptop, notebook, ultra-mobile personal computer (UMPC), netbook, and smart screen.

[0122] like Figure 5 As shown, the electronic device 200 may include one or more of the following components: a processor 201, a memory 203, a communication interface 202, and a communication bus 204. The memory 203 can be connected to the processor 201 via the bus 204. The bus can transfer data between the processor 201 and the memory 203. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0123] Processor 201 may include one or more processing cores. Processor 201 can connect to various parts within the electronic device 200 using various interfaces and lines. It performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 203, and by calling data stored in memory 203. For example, processor 201 may include an application processor (AP), a modem processor, a CPU, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic array (PLA), and / or a neural network processing unit (NPU). The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed; the NPU implements artificial intelligence (AI) functions; and the modem handles wireless communication. Different processing units can be independent devices or integrated into one or more processors. For example, the multiple processing units shown above are all integrated into a single SoC, or the AP is a separate semiconductor chip, while other processing units are integrated into a single SoC. This application does not limit this to any particular type.

[0124] The memory 203 may include random access memory (RAM), read-only memory (ROM), or non-transitory computer-readable storage medium. The memory 203 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 203 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system or instructions for at least one function, such as a method for identifying nearshore cage aquaculture types with multiple rule constraints. The data storage area may store data created based on the use of the electronic device 200, such as remote sensing images of the study area and statistical yearbooks.

[0125] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for identifying nearshore cage aquaculture types under multiple rule constraints, characterized in that, include: S1. For the interpreted cage aquaculture patches in the remote sensing image, based on spectral features, texture features and shape features, a cage aquaculture facility classifier is used to classify the cage aquaculture patches by facility type, and the cage aquaculture facility type classification results are obtained. S2, based on the facility type classification results, combined with the characteristics of sample points and water depth environment, the aquaculture type classifier is used to preliminarily identify the aquaculture organism type of each patch, and output the cage aquaculture type classification results; S3, Construct a multi-rule constraint aquaculture type discrimination model. The multi-rule constraint aquaculture type discrimination model is used to introduce multiple rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules to adjust the cage aquaculture type of each patch and form the final cage aquaculture type discrimination result. The classification results of the cage aquaculture facilities include standard cages and hanging cages; In step S3, multi-rule constraints are introduced, including statistical distribution rules, township distribution rules, and water depth ecological suitability rules, to adjust the cage aquaculture type for each patch, forming the final cage aquaculture type discrimination result, including: Calculate the scores for statistical area index, township distribution index, and water depth condition index according to the statistical distribution rules, township distribution rules, and water depth ecological suitability rules, respectively. The probability values ​​of different aquaculture types in the cage aquaculture type classification results output in step S2 are weighted and summed with the scores of statistical area index, township distribution index, and water depth index to obtain a comprehensive score. The aquaculture type with the highest comprehensive score is taken as the final cage aquaculture type discrimination result.

2. The method according to claim 1, characterized in that, The types of cage aquaculture include: fish, standard cage-sea cucumber, hanging cage-sea cucumber, and abalone.

3. The method according to claim 2, characterized in that, Before introducing multi-rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules, the multi-rule constrained aquaculture type discrimination model is also used to initially adjust the classification results of cage aquaculture types based on hard constraints. The specific steps are as follows: Based on the facility type classification results, a hard constraint is imposed on the cage aquaculture type classification results output in step S2: if the facility type of a certain cage aquaculture patch is a standard cage according to the facility type classification results, then only the aquaculture organism types of fish and standard cage-sea cucumber are allowed; if the facility type of the cage aquaculture patch is a hanging cage cage, then only the aquaculture organism types of hanging cage-sea cucumber and abalone are allowed. When the predicted category in the cage aquaculture type classification result does not meet the hard constraint with the facility type classification result, its predicted probability is removed from the probability matrix, and the category with the highest probability is selected from the remaining predicted probabilities as the cage aquaculture type prediction result for the cage aquaculture patch. At the same time, the remaining predicted probabilities are normalized to obtain the initial adjustment result that meets the hard constraint.

4. The method according to claim 1, characterized in that, The calculation steps for the statistical area index score are as follows: Based on statistical yearbooks and classification results of cage aquaculture types, the proportion of aquaculture area for each aquaculture type in each sub-region of the study area was calculated, and the target proportion and the current proportion were obtained accordingly. The relative difference between the target ratio and the existing ratio is quantified, and the statistical area index score is calculated using the following formula: ， In the formula, This refers to the serial number of the cage aquaculture patch. As a candidate category for cage aquaculture, To score the area indicator, , The first The target proportion and the existing proportion within the sub-region where each patch is located.

5. The method according to claim 1, characterized in that, The township distribution index score is calculated according to the following steps: Based on the spatial differences in the aquaculture distribution in each sub-region within the study area, the correspondence between different aquaculture organism types and each sub-region is determined; Based on the correspondence between different types of farmed organisms and each sub-region, a binary discrimination mechanism is used to calculate the township distribution index score.

6. The method according to claim 1, characterized in that, The score for the water depth condition index is calculated according to the following steps: Based on the constraints of biological habits, the water depth conditions corresponding to different types of cultured organisms were determined; A piecewise function is constructed to calculate the water depth condition index score; the expression of the piecewise function is as follows: ， In the formula, This refers to the serial number of the cage aquaculture patch. As a candidate category for cage aquaculture, The score is based on the water depth condition index. This represents the water depth.

7. The method according to claim 1, characterized in that, The classifier for the cage aquaculture facility is a random forest model, and the classifier for the aquaculture type is an XGBoost classification model.

8. A nearshore cage aquaculture type discrimination system with multiple rule constraints, characterized in that, The system is used to perform the method as described in any one of claims 1 to 7, including: The facility classification unit is configured to classify cage aquaculture patches in remote sensing images based on spectral, texture and shape features using a cage aquaculture facility classifier to obtain cage aquaculture facility type classification results. The preliminary classification unit for biological categories is configured to use an aquaculture type classifier to preliminarily identify the aquaculture biological types of each patch based on the facility type classification results, combined with the characteristics of the sample points and water depth environment, and output the cage aquaculture type classification results. The multi-rule constraint unit is configured to construct a multi-rule constraint aquaculture type discrimination model. The multi-rule constraint aquaculture type discrimination model is used to introduce multi-rule constraints such as statistical distribution rules, township distribution rules, and water depth ecological suitability rules to adjust the cage aquaculture type of each patch and form the final cage aquaculture type discrimination result. The classification results of the cage aquaculture facilities include standard cages and hanging cages; In the multi-rule constraint unit, statistical distribution rules, township distribution rules, and water depth ecological suitability rules are introduced to adjust the cage aquaculture type of each patch, forming the final cage aquaculture type discrimination result, including: Calculate the scores for statistical area index, township distribution index, and water depth condition index according to the statistical distribution rules, township distribution rules, and water depth ecological suitability rules, respectively. The probability values ​​of different aquaculture types in the preliminary classification results of cage aquaculture type output by the biological category are weighted and summed with the scores of statistical area index, township distribution index, and water depth index to obtain a comprehensive score. The aquaculture type with the highest comprehensive score is taken as the final cage aquaculture type discrimination result.

Images