A method for extracting a cultivation area based on time series filtering

Abstract

The application discloses a kind of based on time series filtering aquaculture area extraction method.The method includes: obtaining multi-temporal optical image data to target area and pre-processing, constructs time series image dataset;Sample dataset for aquaculture area extraction model training is constructed;A kind of aquaculture area extraction model with hierarchical progressive characteristics and spatial adaptive capacity is constructed;Model training learns the spatial structure characteristics and time evolution characteristics of aquaculture area, obtains aquaculture area initial mask;Time series filtering processing is executed, through constructing time sequence stability discrimination mechanism, transient noise is inhibited, and the extraction result of time series consistent aquaculture area is obtained;Change analysis module of aquaculture area is constructed, and the spatial distribution of aquaculture area in different time phases is automatically compared and analyzed, and the boundary extraction and spatial area quantitative representation of each time phase aquaculture area are obtained.The present application can effectively alleviate the problem of spectral confusion and space-time change interference, improve the stability and reliability of aquaculture area extraction result in time dimension.

Description

Technical Field

[0001] This invention relates to the field of aquaculture area extraction technology, and specifically to an aquaculture area extraction method based on time series filtering. Background Technology

[0002] Optical remote sensing technology can acquire the reflectance or emission spectral information of the same target in different bands at the same time through photography or scanning. Multispectral images typically contain 4-15 discrete bands (such as visible light, near-infrared, shortwave infrared, etc.), with a wide bandwidth (e.g., 50-200 nanometers) and relatively low spectral resolution. They are suitable for large-scale surface monitoring tasks, such as land use classification, vegetation cover analysis, water surface target identification, and water quality monitoring.

[0003] Mussel farming is a vital marine economic industry in my country's coastal areas, and its automated monitoring is crucial for marine ranch management and ecological environmental protection. Traditional monitoring methods primarily rely on manual field surveys, which are costly, inefficient, and lack timeliness. Existing remote sensing monitoring methods are mostly based on threshold segmentation or traditional machine learning algorithms, but they still have the following limitations:

[0004] 1. Spectral confusion problem: Aquaculture facilities have similar spectral characteristics to natural reefs and islands, which can easily lead to misclassification. Spatial feature information needs to be combined to further improve the accuracy of automatic identification.

[0005] 2. Spatiotemporal variation interference: The spectral characteristics of different breeding areas will change dynamically with the season and year, making traditional methods difficult to apply. It is necessary to introduce effective spatiotemporal analysis methods to reduce the impact of spatiotemporal interference on the identification results.

[0006] 3. Insufficient utilization of time series information: Most methods do not fully consider the optimization results of multi-temporal data, and new models need to be developed to efficiently utilize limited time series information. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for extracting aquaculture areas based on time series filtering.

[0008] To achieve the above objectives, this invention provides a method for extracting aquaculture areas based on time series filtering, the method comprising the following steps:

[0009] Step 1: Acquire multi-temporal optical remote sensing image data of the target area, and perform spatial registration and regional cropping preprocessing on the multi-temporal optical remote sensing image data to construct a time-series image dataset of the target area;

[0010] Step 2: Based on the time series image dataset, label the aquaculture area and the non-aquaculture area, and extract spatial domain information for each labeled object according to a preset spatial scale to form a sample dataset containing spatial context relationships;

[0011] Step 3: Construct an aquaculture area extraction model and train the model using a sample dataset to enable the model to learn the joint representation of the spatial structure features and temporal evolution features of the aquaculture area.

[0012] Step 4: Input the time series image dataset into the trained aquaculture area extraction model to obtain the aquaculture area extraction results corresponding to each time period;

[0013] Step 5: Jointly analyze the extraction results of multiple breeding areas corresponding to the same spatial location, construct a temporal stability discrimination mechanism to suppress instantaneous noise and enhance the consistency of breeding areas in the time dimension, and obtain breeding area extraction results with consistent temporal sequence.

[0014] Step 6: Compare and analyze the spatial distribution of aquaculture areas in different time phases, extract the boundaries between the expansion and recession areas of aquaculture areas; calculate the spatial area and its changes in aquaculture areas in each time phase, and realize the quantitative characterization of the spatiotemporal dynamic evolution of aquaculture areas.

[0015] Furthermore, the aquaculture area extraction model includes:

[0016] The tensor construction module is used to convert multi-temporal optical image samples acquired in the study area into multi-dimensional tensor forms suitable for model input, so as to characterize the spatial structure and temporal evolution features of the samples.

[0017] The feature extraction module is used to encode the features of the tensor-form samples and extract intermediate feature representations that can characterize the spatial structure and temporal evolution of the aquaculture area.

[0018] The first center perception module is used to explicitly construct a feature organization method with the center position as a reference in the spatial dimension, take the core response of the central region as the structural anchor point, and combine it with the context information of the surrounding region for directional reconstruction, thereby forming a feature expression mechanism dominated by the center to highlight the core structural features of the target region.

[0019] The second center perception module is used to repeatedly execute the center guidance and context reorganization process at multiple feature levels, so that the discrimination information of the center region is propagated and accumulated and expressed step by step. It also combines the relationship between the center and each spatial location for adaptive fusion, thereby forming a center perception enhanced feature representation with hierarchical progressive characteristics and spatial adaptive capabilities.

[0020] The spatial relationship module is used to construct global spatial dependencies based on the center-aware enhanced feature representation and generate an enhanced global center vector containing spatial context association information.

[0021] The classification module is used to generate initial classification results for aquaculture areas and non-aquaculture areas based on the global center vector.

[0022] Furthermore, both the first center perception module and the second center perception module include a center perception unit, which includes a center path and a circumferential path. The center path is used to extract center information features from the features output by the third convolution module based on center projection, and the circumferential path is used to extract circumferential information features from the features output by the third convolution module.

[0023] Furthermore, the specific method for extracting center information features from the central path is as follows:

[0024] Extracting intermediate features The spatial location of the center is determined, and then 1×1 convolution and upsampling operations are performed on the extracted spatial locations to obtain the center information features.

[0025] Furthermore, the circumferential path includes m circumferential convolutional kernels, and its method for extracting circumferential information features is as follows:

[0026] ;

[0027] in, Extract circumferential information features for the circumferential path. This indicates that the calculation is performed by concatenating data along the channel dimension. This represents the output of the k-th circular convolution kernel, where k = 1, 2, ..., m.

[0028] Furthermore, the spatial relationship module includes a fourth convolution module, which is used to project the channels in the feature matrix after center enhancement to the relationship dimension between the center projection and the circumferential projection.

[0029] Furthermore, the spatial relationship module also includes a fifth convolution module, which is used to generate a weight distribution of spatial positions based on the inner product of the central projection and the circumferential projection in the relationship dimension, as follows:

[0030] Flatten the circumferential projection along the spatial dimension to obtain the flattened tensor. ;

[0031] Computing Center Projection With circumferential projection Similarity score matrix for:

[0032] ;

[0033] Calculate the weight distribution of spatial location for:

[0034]

[0035] in, This represents the activation function.

[0036] Furthermore, the classification module includes a two-layer fully connected network. The first fully connected network is used to project the enhanced global center vector onto the hidden dimension, and the second fully connected network is used to map the output of the first fully connected network to the probability of the aquaculture area or non-aquaculture area category, so as to obtain a preliminary binary map of the aquaculture area and non-aquaculture area.

[0037] Furthermore, the generation method of the extraction results of the aquaculture area in step 5 is as follows:

[0038] The extraction results of breeding areas corresponding to multiple times are converted into binary masks;

[0039] For consecutive frame sequences, use a length of Sampling is performed within a sliding time window to form a window tensor. ;

[0040] For window tensors Accumulate and count at each pixel position to obtain an accumulation matrix. This is used to reflect the stability of local pixels;

[0041] Based on the minimum density threshold D, the cumulative counting matrix Perform binarization to generate a local stability mask. To eliminate occasional noise;

[0042] Local stability mask Construct a cumulative mask based on this. , To represent taking the union of sets, This is the cumulative mask from the previous moment;

[0043] By combining a sliding time window with local stability assessment, transient noise and local fluctuations in remote sensing sequences are suppressed, resulting in a time-consistent aquaculture area mask.

[0044] Beneficial effects: 1. This invention constructs a local-global coupling mechanism of "center-guided + multi-scale periphery" by explicitly extracting and propagating the central information features and the surrounding information features at the whole map level, which is beneficial to compensate for the surrounding context while maintaining the local central information.

[0045] 2. The spatial relationship module is used to project high-dimensional channel information into low-dimensional relationship representation, and the spatial attention distribution is generated based on the correlation between the center and each spatial location, so that the final aggregated center vector contains both the center pixel information and the semantic context after global position weighting.

[0046] 3. In terms of implementation, it adopts a depth-separable form of peripheral convolution and lightweight 1×1 fusion, which takes into account both multi-scale receptive field and computational efficiency, thus adapting to high-resolution input scenarios such as patch level or remote sensing images, and solving the problems of spectral confusion and poor adaptability to spatiotemporal changes. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a method for extracting aquaculture areas based on time-series filtering according to an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the structure of the aquaculture area extraction model according to an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the feature extraction module according to an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the structure of the first central sensing module according to an embodiment of the present invention;

[0051] Figure 5 This is a schematic diagram of the spatial relationship module according to an embodiment of the present invention;

[0052] Figure 6 It is a binary image after color encoding. Detailed Implementation

[0053] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0054] like Figure 1 As shown, this embodiment of the invention provides a method for extracting aquaculture areas based on time series filtering, including:

[0055] Step 1: Acquire multi-temporal optical remote sensing image data of the target area, and perform spatial registration and region cropping preprocessing on the multi-temporal optical remote sensing image data to construct a time-series image dataset of the target area. Specifically, the aforementioned multi-temporal optical remote sensing image data can be obtained from the Landsat 8 satellite. The Landsat 8 satellite provides optical data in 11 bands, including visible light, near-infrared, and short-wave infrared bands, but is not limited to these. It can also be acquired using a UAV equipped with sensors for flight control. The aforementioned preprocessing method involves spatial registration and region cropping of the multi-temporal optical remote sensing image data.

[0056] Step 2: Based on the time series image dataset, typical aquaculture areas and non-aquaculture areas are labeled. Spatial domain information (patch) is extracted for each labeled object according to a preset spatial scale to construct a sample dataset containing spatial context relationships.

[0057] Step 3: Design and construct an aquaculture area extraction model, and train the aquaculture area extraction model using a sample dataset, so that the model learns the joint representation of the spatial structure features and temporal evolution features of the aquaculture area.

[0058] For details, see Figure 2 The aquaculture area extraction model includes a tensor extraction module 1, a feature extraction module (StemBlock) 2, a first center awareness module (Center Aware Block I) 3, a second center awareness module (Center Aware Block II) 4, a spatial relationship module (Spatial Relation Module) 5, and a classification module 6.

[0059] Tensor extraction module 1 is used to extract tensor information from the sample dataset, obtaining... ,in, Let B be a real number, B represent a single training batch, and C1 be a feature. The number of spectral channels, where H and W represent the height and width of the feature map, respectively.

[0060] See Figure 3 The feature extraction module 2 includes a first convolutional module 21 (kernel=3, padding=1), a second convolutional module 22 (kernel=3, padding=1), and a first downsampling module 23 (kernel=2). The first convolutional module 21 integrates a convolutional layer 211, a batch normalization layer 212, and an activation layer 213, which are used to process tensors respectively. The convolution, batch normalization, and activation operations are performed sequentially. The second convolution module 22 has the same structure as the first convolution module 21 and is used to sequentially perform convolution, batch normalization, and activation operations on the features output by the first convolution module 21 to extract depth features. The first downsampling module 23 is used to spatially reduce the dimensionality of the features output by the second convolution module 22 to enhance key features, reduce computational cost, and obtain intermediate features. C2 is an intermediate feature. The number of spectral channels is preferably 7.

[0061] The first center perception module 3 is a center-guided local structure modeling module proposed in this invention, used for center perception modeling of local response information in the intermediate features. By explicitly constructing a feature organization method with the center position as a reference in the spatial dimension, the core response of the central region is used as the structural anchor point, and the surrounding context information is combined for directional reconstruction, thereby forming a center-led feature expression mechanism to highlight the core structural features of the target region. The second center perception module 4 is a multi-level center enhancement structure built on the first center perception module 3, used for progressive aggregation and semantic enhancement of multi-level center information of the features output by the first center perception module 3. By repeatedly executing the center guidance and context reorganization process at multiple feature levels, the discrimination information of the central region is propagated and accumulated at each level, and adaptive fusion is performed by combining the relationship between the center and each spatial position, thereby forming a center perception enhanced feature representation with hierarchical progressive characteristics and spatial adaptive capabilities.

[0062] See Figure 4 The first center perception module 3 includes a third convolutional module 31 (kernel=3, padding=1), a center perception unit 32 (Kernel List), and a second downsampling module 33 (kernel=2). The third convolutional module 31 sequentially performs convolution, batch normalization, and activation operations on the received features. The center perception unit 32 includes a center path and a circumferential path. The center path is used to extract center information features from the features output by the third convolutional module 31 based on center projection, as follows:

[0063] Extracting intermediate features The spatial location of the center is determined, and then 1×1 convolution and upsampling operations are performed sequentially on the extracted spatial locations to obtain the center information features. for:

[0064]

[0065] in, This represents the upsampling operation where the center response is bilinearly interpolated to the spatial dimensions (H, W). This represents a 1×1 convolution operation for channel mapping. , This represents the central single-pixel block obtained through direct indexing. These are the coordinates of the center pixel in the height and width directions, respectively. .

[0066] The circumferential path is used to extract circumferential information features from the features output by the third convolutional module 31 based on circumferential projection. Specifically, the circumferential path includes m circumferential convolutional kernels, and its method for extracting circumferential information features is as follows:

[0067]

[0068] in, Extract circumferential information features for the circumferential path. This indicates that the calculation is performed by concatenating data along the channel dimension. This represents the output of the k-th circular convolution kernel, where k = 1, 2, ..., m.

[0069] The second downsampling module 33 is used to downsample the center information features and peripheral information features extracted by the center perception unit 32 to obtain the feature matrix after center enhancement. The structure of the second center perception module 4 is the same as that of the first center perception module 3, except that the first center perception module 3 processes the intermediate features output by the feature extraction module 2. The second center perception module 4 processes the feature matrix after center enhancement output by the first center perception module 3.

[0070] See Figure 5 The spatial relationship module 5 includes a fourth convolutional module 51 (kernel=1, padding=0), a fifth convolutional module 52 (kernel=1, padding=0), and an aggregation module 53. The fourth convolutional module 51 projects the channels in the center-enhanced feature matrix to the relationship dimension between the center projection and the circumferential projection. The fifth convolutional module 52 generates the spatial position weight distribution based on the inner product of the center projection and the circumferential projection in the relationship dimension. The aggregation module 53 performs weighted aggregation on the center-enhanced feature matrix in the spatial dimension to obtain an enhanced feature matrix. Specifically, the relationship between the center projection and the circumferential projection is expressed as follows:

[0071]

[0072] in, An R-dimensional description of the central location. This is a mapping of the input to R in a dimensionality reduction manner across the channels.

[0073] The weight distribution of the above spatial locations is generated as follows:

[0074] Flatten the circumferential projection along the spatial dimension to obtain the flattened tensor. for:

[0075]

[0076] in, This indicates a flattening operation along spatial dimensions. Indicates circumferential projection. , which represents the total number of pixels in the space.

[0077] Calculate the similarity score matrix between the central projection and the circumferential projection. for:

[0078]

[0079] Where T is the transpose of a matrix.

[0080] Calculate the weight distribution of spatial location for:

[0081]

[0082] in, This represents the activation function.

[0083] The enhanced global center vector described above is calculated as follows:

[0084]

[0085]

[0086] in, The calculated enhanced global center vector, This indicates that element-wise multiplication is performed according to the broadcast. , These are the coordinate indices in the height and width directions, respectively.

[0087] Classification module 6 generates a preliminary binary classification map of aquaculture and non-aquaculture areas based on the enhanced global center vector output by the spatial relationship module. Specifically, classification module 6 includes two fully connected layers. The first fully connected layer projects the enhanced global center vector onto the hidden dimensions, and the second fully connected layer maps the output of the first fully connected layer to the probability of the aquaculture or non-aquaculture category, as detailed below:

[0088]

[0089] in, This represents the probability of mapping to either aquaculture or non-aquaculture zones. , These are the weights of the first fully connected network and the second fully connected network, respectively. and These are the bias terms for the first fully connected network and the second fully connected network, respectively. D is the hidden dimension. K is the number of categories. The ReLU activation function is used for probability. By using the softmax function, preliminary binary maps of the aquaculture and non-aquaculture areas can be obtained.

[0090] Step 4: Input the time series image dataset into the trained aquaculture area extraction model to obtain the aquaculture area extraction results for each time period.

[0091] Step 5: Jointly analyze the extraction results of multiple time-corresponding aquaculture areas at the same spatial location, and construct a temporal stability discrimination mechanism to suppress instantaneous noise and enhance the consistency of aquaculture areas in the temporal dimension, so as to obtain temporally consistent aquaculture area extraction results, as follows:

[0092] The extraction results of aquaculture areas corresponding to multiple times are converted into binary masks. For continuous frame sequences, a mask of length [length missing] is used. Sampling is performed within a sliding time window to form a window tensor. For window tensors Accumulate and count at each pixel position to obtain an accumulation matrix. This is used to reflect the stability of local pixels. The accumulated counting matrix is ​​based on the minimum density threshold D. Perform binarization to generate a local stability mask. To eliminate sporadic noise. In the local stability mask Construct a cumulative mask based on this. , To represent taking the union of sets, This is the cumulative mask from the previous time step. By combining a sliding time window with local stability assessment, instantaneous noise and local fluctuations in the remote sensing sequence are suppressed, resulting in a time-consistent aquaculture area mask.

[0093] Step 6: Compare and analyze the spatial distribution of aquaculture areas at different time phases, extracting the boundaries between the expansion and recession areas of aquaculture areas; calculate the spatial area and its changes in aquaculture areas at each time phase, achieving a quantitative characterization of the spatiotemporal dynamic evolution of aquaculture areas. Different colors can be used to encode aquaculture areas and non-aquaculture areas, generating a visualized image of aquaculture area changes.

[0094] When training the model for extracting aquaculture areas, Landsat 8 multi-temporal optical images were acquired, and the latitude and longitude range of the study area was confirmed. Based on the latitude and longitude information, the multi-temporal optical images were batch-cropped to obtain RGB images of the area, in which the approximate extent of the aquaculture area was visible to humans. Then, sampling was performed using a combination of field sampling and visual interpretation methods, and a sliding window method was used to extract image patches (e.g., 25×25 pixels) as an aid to construct model training samples, including positive samples (aquaculture areas), negative samples (non-aquaculture areas), and interference samples. The cross-entropy loss function and the AdamW optimizer (learning rate 0.0001, weight decay 1e-5) were used for 200 training epochs, with the learning rate dynamically adjusted using ReduceLROnPlateau.

[0095] Let's take the extraction from a mussel farming area in a certain region as an example:

[0096] Landsat 8 optical imagery of the region from 2013 to 2025 was acquired. A well-trained aquaculture area extraction model was used for prediction, employing a 25×25 sliding window with a step size of 5 pixels and a batch size of 64 to obtain preliminary binary images. Time-series filtering was then applied to these preliminary binary images to obtain the final binary images. The 2025 binary image was then color-coded as follows... Figure 6 As shown, Figure 6 White represents the aquaculture area, and black represents the non-aquaculture area. By comparing the aquaculture areas at different times, the changes in the aquaculture areas can be studied.

[0097] The above description is merely a preferred embodiment of the present invention. It should be noted that for those skilled in the art, other parts not specifically described are existing technology or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting aquaculture areas based on time series filtering, characterized in that, The method includes the following steps: Step 1: Acquire multi-temporal optical remote sensing image data of the target area, and perform spatial registration and regional cropping preprocessing on the multi-temporal optical remote sensing image data to construct a time-series image dataset of the target area; Step 2: Based on the time series image dataset, label the aquaculture area and the non-aquaculture area, and extract spatial domain information for each labeled object according to a preset spatial scale to form a sample dataset containing spatial context relationships; Step 3: Construct an aquaculture area extraction model and train the model using a sample dataset to enable the model to learn the joint representation of the spatial structure features and temporal evolution features of the aquaculture area. Step 4: Input the time series image dataset into the trained aquaculture area extraction model to obtain the aquaculture area extraction results corresponding to each time period; Step 5: Jointly analyze the extraction results of multiple breeding areas corresponding to the same spatial location, construct a temporal stability discrimination mechanism to suppress instantaneous noise and enhance the consistency of breeding areas in the time dimension, and obtain breeding area extraction results with consistent temporal sequence. Step 6: Compare and analyze the spatial distribution of aquaculture areas at different time phases, and extract the boundaries between the expansion and recession areas of aquaculture areas; Calculate the spatial area and its changes in the aquaculture zone at each time phase to achieve a quantitative characterization of the spatiotemporal dynamic evolution of the aquaculture zone; The aquaculture area extraction model includes: The tensor construction module is used to convert multi-temporal optical image samples acquired in the study area into multi-dimensional tensor forms suitable for model input, so as to characterize the spatial structure and temporal evolution features of the samples. The feature extraction module is used to encode the features of the tensor-form samples and extract intermediate feature representations that can characterize the spatial structure and temporal evolution of the aquaculture area. The first center perception module is used to explicitly construct a feature organization method with the center position as a reference in the spatial dimension, take the core response of the central region as the structural anchor point, and combine it with the context information of the surrounding region for directional reconstruction, thereby forming a feature expression mechanism dominated by the center to highlight the core structural features of the target region. The second center perception module is used to repeatedly execute the center guidance and context reorganization process at multiple feature levels, so that the discrimination information of the center region is propagated and accumulated and expressed step by step. It also combines the relationship between the center and each spatial location for adaptive fusion, thereby forming a center perception enhanced feature representation with hierarchical progressive characteristics and spatial adaptive capabilities. The spatial relationship module is used to construct global spatial dependencies based on the center-aware enhanced feature representation and generate an enhanced global center vector containing spatial context association information. The classification module is used to generate initial classification results for aquaculture areas and non-aquaculture areas based on the global center vector.

2. The method for extracting aquaculture areas based on time series filtering according to claim 1, characterized in that, Both the first center perception module and the second center perception module include a center perception unit. The center perception unit includes a center path and a circumferential path. The center path is used to extract center information features from the features output by the third convolution module based on center projection. The circumferential path is used to extract circumferential information features from the features output by the third convolution module.

3. The method for extracting aquaculture areas based on time series filtering according to claim 2, characterized in that, The specific method for extracting central information features from the central path is as follows: Extracting intermediate features The spatial location of the center is determined, and then 1×1 convolution and upsampling operations are performed on the extracted spatial locations to obtain the center information features.

4. The method for extracting aquaculture areas based on time series filtering according to claim 2, characterized in that, The circumferential path includes m circumferential convolutional kernels, and its method for extracting circumferential information features is as follows: ； in, As the central information feature, Extract circumferential information features for the circumferential path. This indicates that the calculation is performed by concatenating data along the channel dimension. This represents the output of the k-th circular convolution kernel, where k = 1, 2, ..., m.

5. The method for extracting aquaculture areas based on time series filtering according to claim 1, characterized in that, The spatial relationship module includes a fourth convolution module, which is used to project the channels in the feature matrix after center enhancement to the relationship dimension between the center projection and the circumferential projection.

6. The method for extracting aquaculture areas based on time series filtering according to claim 5, characterized in that, The spatial relationship module also includes a fifth convolution module, which is used to generate a weight distribution of spatial positions based on the inner product of the central projection and the circumferential projection in the relationship dimension, as follows: Flatten the circumferential projection along the spatial dimension to obtain the flattened tensor. ; Computing Center Projection With circumferential projection Similarity score matrix for: ； in, B is a real number, representing a single training batch, and N represents the total number of spatial pixels; Calculate the weight distribution of spatial location for: ； in, This represents the activation function.

7. The method for extracting aquaculture areas based on time series filtering according to claim 1, characterized in that, The classification module includes two fully connected layers. The first fully connected layer projects the enhanced global center vector onto the hidden dimension, and the second fully connected layer maps the output of the first fully connected layer to the probability of the aquaculture area or non-aquaculture area category, so as to obtain a preliminary binary map of the aquaculture area and non-aquaculture area.

8. The method for extracting aquaculture areas based on time series filtering according to claim 1, characterized in that, The method for generating the extraction results of the aquaculture area in step 5 is as follows: The extraction results of breeding areas corresponding to multiple times are converted into binary masks; For consecutive frame sequences, use a length of Sampling is performed within a sliding time window to form a window tensor. ; For window tensors Accumulate and count at each pixel position to obtain an accumulation matrix. This is used to reflect the stability of local pixels; Based on the minimum density threshold D, the cumulative counting matrix Perform binarization to generate a local stability mask. To eliminate occasional noise; Local stability mask Construct a cumulative mask based on this. , To represent taking the union of sets, This is the cumulative mask from the previous moment; By combining a sliding time window with local stability assessment, transient noise and local fluctuations in remote sensing sequences are suppressed, resulting in a time-consistent aquaculture area mask.

Images