A sea area aquaculture activity dynamic monitoring and identification method based on multi-source remote sensing data fusion

Abstract

The application provides a kind of sea area aquaculture activity dynamic monitoring identification method based on multi-source remote sensing data fusion, belongs to remote sensing data analysis technical field, the application constructs discriminant feature set by extracting polarimetric decomposition parameter, multi-scale texture statistics and spectral index, calculates multi-scale morphological complexity quantification feature to distinguish regular aquaculture structure and irregular sea state noise, builds multi-granularity cascade discriminant model to fuse the above-mentioned features and output aquaculture activity spatial distribution identification result, constructs artificial intelligence time series event perception model to process long time series backscattering intensity sequence and meteorological data and outputs aquaculture activity state change detection result and confidence interval estimation, finally generates sea area aquaculture activity dynamic monitoring identification report, solves the technical problem that multi-source remote sensing data fusion identification method is difficult to stably distinguish artificial aquaculture facilities and natural sea state noise and realize long time series dynamic change detection under the condition of complex sea state and cloud and fog shielding.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing data analysis technology, and specifically relates to a method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion. Background Technology

[0002] Dynamic monitoring of marine aquaculture activities is an important research direction in the field of marine remote sensing and target recognition. In existing technologies, researchers typically use single synthetic aperture radar (SAR) images or optical remote sensing images, combined with traditional machine learning classification methods, to identify aquaculture areas. Some schemes introduce multilayer perceptrons or convolutional neural networks to perform semantic segmentation of remote sensing images to extract the spatial distribution range of aquaculture facilities. These methods have been applied to the static identification of nearshore regular aquaculture areas. However, in actual marine monitoring scenarios, backscattering noise generated by natural factors such as waves, tides, and ship wakes is highly superimposed on the radar echo characteristics of aquaculture facilities. Single image features are insufficient to effectively distinguish between regular aquaculture structures and irregular sea state noise, leading to a high misjudgment rate. Simultaneously, frequent cloud and fog obscure optical images, causing irregular gaps in the observation sequence. Traditional time-series analysis methods output deterministic results under data gap conditions, lacking a quantitative description of detection confidence, further exacerbating the instability of detecting changes in aquaculture activity status. In other words, existing technologies have technical problems in which multi-source remote sensing data fusion and identification methods cannot robustly distinguish between artificial aquaculture facilities and natural sea noise and achieve long-term dynamic change detection under complex sea conditions and cloud cover. Summary of the Invention

[0003] In view of this, the present invention provides a dynamic monitoring and identification method for marine aquaculture activities based on multi-source remote sensing data fusion, which can solve the technical problem in the prior art that multi-source remote sensing data fusion identification methods are difficult to robustly distinguish between artificial aquaculture facilities and natural sea noise and achieve long-term dynamic change detection under complex sea conditions and cloud cover.

[0004] This invention is implemented as follows: This invention provides a method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion, comprising the following steps:

[0005] Acquire multi-source remote sensing data, including synthetic aperture radar data, optical remote sensing data, and meteorological reanalysis data, and perform registration and radiometric correction.

[0006] Extract polarization decomposition parameters, multi-scale texture statistics from synthetic aperture radar data, and spectral index features from optical remote sensing data to construct a discriminative feature set.

[0007] Calculate the multi-scale morphological complexity quantification features of the target area boundary to distinguish between regular aquaculture structures and irregular sea state noise;

[0008] A multi-source feature cascade fusion model is constructed to fuse the discriminative feature set with multi-scale morphologically complex quantitative features, and output the spatial distribution identification results of aquaculture activities.

[0009] A pulse neural network time-series detection framework is constructed to process long-time synthetic aperture radar backscatter intensity sequences and meteorological reanalysis data, and output the detection results of changes in aquaculture activity status and confidence interval estimates.

[0010] Based on the spatial distribution identification results of aquaculture activities and the detection results of changes in the status of aquaculture activities, a dynamic monitoring and identification report of marine aquaculture activities is generated.

[0011] The polarization decomposition parameter refers to the scattering mechanism component obtained after polarization decomposition of the synthetic aperture radar echo signal, which is used to characterize the relative contributions of target surface scattering, volume scattering and dihedral scattering.

[0012] The multi-scale texture statistics refer to the gray-level co-occurrence matrix parameters and local binary mode parameters calculated at multiple window scales, and the spectral indices include the normalized water index and the raft index.

[0013] The multi-scale morphological complexity quantification feature is calculated by a multi-scale morphological complexity quantification algorithm based on fractal dimension features. The box counting method is used to count the number of boxes that the target region boundary crosses under multiple scale windows. Linear regression is performed on the scale parameter and the number of boxes in a double logarithmic coordinate system, and the absolute value of the regression slope is used as the fractal dimension estimate.

[0014] The multi-scale morphological complexity quantification algorithm further calculates the generalized fractal dimension spectrum under different moment orders to characterize the local singularity distribution characteristics of the target region. The calculation process of the generalized fractal dimension spectrum is based on convolution operation. The fractal dimension estimate of the artificial aquaculture facility falls within a stable interval. The interval boundary is obtained by iterative statistical analysis after box counting statistical experiments on known aquaculture facility samples and known sea state noise samples.

[0015] The multi-source feature cascade fusion model is named the multi-granularity cascaded discrimination model. It is built on a deep forest architecture and the input layer is associated with polarization decomposition parameters, multi-scale texture statistics, spectral indices and multi-scale morphological complexity quantitative features. The multi-granularity scanning module performs sliding window scanning on the input feature map, which is fully based on the integrated voting of random forest and extreme random tree.

[0016] In this multi-granularity cascaded discriminant model, each layer of the cascaded structure consists of multiple random forests and multiple extreme random trees in parallel. The class probability vector output by each layer is concatenated with the original features and used as the input for the next layer. The number of cascaded layers is automatically determined based on the cross-validation accuracy.

[0017] Among them, the multi-granularity cascaded discrimination model sets up a dynamic inter-layer feedback mechanism. When the confidence of the output of a certain layer is lower than the confidence threshold, the structure of that layer is triggered to adaptively increase the number of decision trees for local retraining, and the refined features are fed back to the previous layer for re-evaluation.

[0018] The spiking neural network time-series detection framework is named the Time-Series Event Perception Model. The input layer converts long-time synthetic aperture radar backscatter intensity sequences and meteorological reanalysis data into spiking sequences through time coding. The intermediate layer adopts a leaky integral firing neuron model, combined with synaptic plasticity mechanism to achieve dynamic memory of long-time dependencies.

[0019] Among them, the time-series event perception model embeds a Bayesian uncertainty estimation module, introduces variational inference on the pulse firing threshold parameter, and enables the network output to have the ability to estimate confidence intervals; the output layer outputs the category of changes in the state of aquaculture activities and the corresponding confidence intervals.

[0020] Among them, the time-series event-aware model sets up an event-triggered iterative mechanism. When the amplitude of a sudden change in the pulse firing mode detected in the input sequence exceeds the threshold of the sudden change amplitude, the local sub-network is triggered to recalculate the features of the adjacent time window.

[0021] Among them, the time-series event-aware model sets up a resource regulation function between the intermediate layer and the event-triggered iteration mechanism. , ,in The average frequency of the input pulse. The amplitude of the sudden change in the pulse firing mode. Calculate resource utilization for historical time windows. , , The weighted coefficients are and satisfy the following conditions: .

[0022] Specifically, when the value calculated by the resource regulation function belongs to the stable state range, the local subnetwork maintains the basic memory allocation ratio; when it belongs to the transition state range, the memory allocation ratio increases linearly, and the synaptic weight update step size decreases proportionally; when it belongs to the event-triggered state range, the event-triggered iteration mechanism is triggered.

[0023] The steps for establishing the training dataset for the multi-granularity cascaded discriminant model include: collecting multi-source remote sensing data and manually labeling the boundaries between aquaculture and non-aquaculture areas; extracting features from the labeled samples to form training sample feature vectors; and dividing the dataset into training, validation, and test sets according to proportions.

[0024] The confidence threshold was obtained through experimental statistical analysis of the confidence distribution of the validation set output; the mutation amplitude threshold was obtained through statistical analysis of the pulse firing pattern of the input sequence; and the interval boundary values ​​between the stationary state interval and the transition state interval were determined. The boundary values ​​of the event triggering state interval It was determined through statistical analysis of multiple sets of historical operating data and verification through multiple iterative experiments.

[0025] This invention constructs a multi-granularity cascaded discriminant model and a temporal event perception model, cascading and fusing polarization decomposition parameters, multi-scale texture statistics, spectral indices, and multi-scale morphological complexity quantitative features. It then uses a pulse neural network for temporal detection of long-term synthetic aperture radar backscattering intensity sequences and meteorological reanalysis data. This solves the technical problem of robustly distinguishing between artificial aquaculture facilities and natural sea state noise, and achieving long-term dynamic change detection, under complex sea conditions and cloud cover. The multi-granularity cascaded discriminant model utilizes fractal dimension features to capture the geometric self-similarity of the boundaries of artificial aquaculture facilities, forming separable regions with natural sea state noise in the feature space, thus achieving robust boundary extraction and noise suppression. The temporal event perception model uses Bayesian probability modeling of the pulse firing threshold parameter, enabling the system to output detection results with confidence intervals even when observation data is missing, avoiding misjudgments by deterministic output in data-definitive scenarios. In summary, this invention solves the technical problem mentioned in the background art that multi-source remote sensing data fusion and identification methods are difficult to robustly distinguish between artificial aquaculture facilities and natural sea noise and achieve long-term dynamic change detection under complex sea conditions and cloud cover. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 Distribution of fractal dimension estimates for different categories of samples.

[0028] Figure 3 The image shows the spatial distribution identification results of aquaculture activities output by the multi-granularity cascaded discriminant model.

[0029] Figure 4 The output of the time-series event perception model is a time-series curve of the detection results of changes in the state of aquaculture activities.

[0030] Figure 5 This is a time-series curve of the status change of each aquaculture patch in the dynamic monitoring and identification report of marine aquaculture activities.

[0031] Figure 6 This is a comparison chart showing the distribution of fractal dimension estimates across three types of samples: net cage arrays, raft aquaculture zones, and sea state noise.

[0032] Figure 7 This is a statistical chart showing the distribution of resource regulation function calculation results and the frequency of event-triggered iteration mechanisms. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0034] like Figure 1 The diagram shown is a flowchart of a method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion provided by the present invention. This method includes the following steps:

[0035] S01. Acquire multi-source remote sensing data, including synthetic aperture radar data, optical remote sensing data and meteorological reanalysis data, and perform registration and radiometric correction.

[0036] S02. Extract polarization decomposition parameters, multi-scale texture statistics and spectral index features from synthetic aperture radar data and optical remote sensing data to construct a discriminative feature set;

[0037] S03. Calculate the multi-scale morphological complexity quantification features of the target area boundary to distinguish between regular aquaculture structures and irregular sea state noise;

[0038] S04. Construct a multi-source feature cascade fusion model to fuse the discriminative feature set with multi-scale morphologically complex quantitative features and output the spatial distribution identification results of aquaculture activities.

[0039] S05. Construct a pulse neural network time-series detection framework to process long-time synthetic aperture radar backscatter intensity sequences and meteorological reanalysis data, and output detection results and confidence interval estimates of changes in aquaculture activity status.

[0040] S06. Based on the spatial distribution identification results of aquaculture activities and the detection results of changes in the status of aquaculture activities, generate a dynamic monitoring and identification report of marine aquaculture activities.

[0041] The following explains the non-public concepts and technical terms involved in the plan.

[0042] The polarization decomposition parameters refer to the scattering mechanism components obtained after polarization decomposition of the synthetic aperture radar echo signal, used to characterize the relative contributions of target surface scattering, volume scattering, and dihedral scattering. The multi-scale texture statistics refer to the gray-level co-occurrence matrix parameters and local binary mode parameters calculated at multiple window scales, used to characterize the spatial arrangement of local gray-level distributions in the image. The spectral indices include the normalized water body index and the raft index. The normalized water body index is used to enhance the spectral differences between water and non-water areas, while the raft index is used to enhance the spectral response differences between raft-type aquaculture facilities and open water areas.

[0043] The multi-scale morphological complexity quantification features are calculated using a multi-scale morphological complexity quantification algorithm based on fractal dimension features. This algorithm draws on fractal geometry theory and, considering the self-similar geometric structure features of net cage arrays and raft aquaculture zones, employs box counting to count the number of boxes that the target area boundary traverses across multiple scale windows. A linear regression is then performed on the scale parameters and the number of boxes in a double logarithmic coordinate system, and the absolute value of the regression slope is used as the fractal dimension estimate. Due to the regular arrangement of artificial aquaculture facilities, the fractal dimension estimate usually falls within a stable interval, denoted as . The boundary values ​​were obtained through box counting statistical experiments on multiple sets of known aquaculture facility samples and known sea state noise samples, and the experimental results were analyzed iteratively multiple times. The algorithm further calculates the generalized fractal dimension spectrum under different moment orders to characterize the local singularity distribution features of the target area and capture the spatial pattern changes of uneven density within the cage array. The generalized fractal dimension spectrum calculation process is based on convolution operations, which facilitates parallel processing of large-scale marine images. The introduction of this algorithm allows fractal dimension features to be used as a supplementary discriminant and weighted fusion with the semantic features extracted by the cascaded fusion model, improving the robustness of aquaculture facility boundary extraction and sea state noise suppression under complex sea conditions. The principle is that the boundary morphology of natural sea state noise lacks scale consistency, while the boundary morphology of artificial aquaculture facilities maintains geometric self-similarity at multiple scales, and the two are separable in the fractal dimension feature space.

[0044] The multi-source feature cascade fusion model, named the multi-granularity cascaded discriminant model, is built on a deep forest architecture and belongs to the machine learning model based on decision tree ensemble. This model abandons the traditional backpropagation training method of neural networks and adopts a cascaded structure of decision tree ensemble to achieve deep feature representation learning. The model structure specifically consists of: input layer correlation polarization decomposition parameters, multi-scale texture statistics, spectral indices, and multi-scale morphological complexity quantification features; a multi-granularity scanning module performs a sliding window scan on the input feature map, with the window scale set denoted as... At each scale, a set of feature vectors corresponding to the receptive field is generated, simulating the local receptive field mechanism of a convolutional neural network, but entirely based on the ensemble voting of random forests and extreme random trees to avoid the gradient vanishing problem. Each layer of the cascaded structure consists of multiple random forests and multiple extreme random trees in parallel. The class probability vector output by each layer is concatenated with the original features and used as the input to the next layer, passing layer by layer to achieve deep feature evolution. The number of cascaded layers is automatically determined based on the cross-validation accuracy, avoiding manual setting of network depth hyperparameters. A dynamic inter-layer feedback mechanism is set inside the model, and when the confidence of the output of a certain layer falls below a threshold... When this occurs, the layer is triggered to adaptively increase the number of decision trees for local retraining, and the refined features are fed back to the previous layer for re-evaluation, forming a local iterative optimization closed loop, with a threshold. The specific values ​​were obtained through experimental statistical analysis of the confidence distribution of the validation set output.

[0045] The specific steps for establishing the training dataset of the multi-granularity cascaded discriminant model include: collecting multi-source remote sensing data and manually labeling the boundaries between aquaculture and non-aquaculture areas; classifying and organizing the labeled samples according to the type of aquaculture facility; extracting features from the labeled samples to obtain polarization decomposition parameters, multi-scale texture statistics, spectral indices, and multi-scale morphological complexity quantitative features, which constitute the training sample feature vector; and dividing the training set, validation set, and test set according to a certain ratio, which is determined after comparing the model convergence effect through multiple experiments.

[0046] The specific steps for training the multi-granularity cascaded discriminant model include: inputting the feature vectors of the training samples into the multi-granularity scanning module to obtain a multi-scale feature vector set; inputting the multi-scale feature vector set into the first layer of the cascaded structure, training the random forest and extreme random tree layer by layer, calculating the output confidence on the validation set after each layer is trained, and determining whether the dynamic inter-layer feedback mechanism is triggered; iterating layer by layer until the cross-validation accuracy no longer improves, determining the final number of cascaded layers, and completing the model training.

[0047] Compared to traditional deep neural networks, this model has the advantages of lower training data requirements, fewer hyperparameters, and suitability for small-sample aquaculture labeled data. Its principle lies in the cascaded structure of decision tree ensembles, which replaces the gradient backpropagation process with layer-by-layer feature evolution, reducing dependence on large-scale labeled data. Simultaneously, the model inherently possesses feature importance measurement capabilities, facilitating the identification of key discriminant variables affecting aquaculture identification results. These key discriminant variables refer to the feature components that contribute most to classification accuracy during the training process of each layer of the cascaded structure; the contribution ranking is calculated through the model's built-in feature importance evaluation mechanism. The introduction of this model improves the training efficiency and interpretability of aquaculture identification models in remote marine monitoring scenarios.

[0048] The proposed spiking neural network time-series detection framework, named the Time-Series Event Perception Model, is built upon a spiking neural network architecture and belongs to the machine learning model that simulates the spiking mechanism of biological neurons. The model structure is as follows: The input layer converts long-term synthetic aperture radar backscattering intensity sequences and meteorological reanalysis data, including wind speed, tides, and sea surface temperature, into spiking sequences through time encoding; the intermediate layer employs a leaky integral firing neuron model, where the neuron membrane potential accumulates with the input spiking, outputting a spiking signal and resetting upon reaching a firing threshold, thus achieving dynamic memorization of long-term time-series dependencies through synaptic plasticity; the network embeds a Bayesian uncertainty estimation module, introducing variational inference into the spiking threshold parameter to enable the network output to estimate confidence intervals; the network is configured with an event-triggered iterative mechanism, where a sudden change in the spiking pattern detected in the input sequence exceeds the threshold. At that time, the local subnetwork is triggered to recalculate the features of the adjacent time window, realizing the dynamic tilting allocation of computing resources to the abnormal change area; the output layer outputs the category of changes in the state of aquaculture activities and the corresponding confidence interval.

[0049] The steps for establishing the training dataset of the time-series event perception model specifically include: collecting historical long-term synthetic aperture radar observation data and corresponding meteorological reanalysis data, and aligning them with the time axis to form time-series samples; labeling the time nodes in the time-series samples where changes in the state of aquaculture activities such as cage placement and harvesting have occurred, and labeling time periods where no changes in state have occurred as stable states; and dividing the labeled time-series samples into training sets and validation sets.

[0050] The specific steps for training the time-series event perception model include: converting the time-series samples in the training set into pulse sequences and inputting them into the network using time encoding; using a leaky integral firing neuron model to transmit pulse signals layer by layer, and updating synaptic weights in conjunction with the synaptic plasticity mechanism; performing variational inference optimization on the pulse firing threshold parameter to ensure that the proportion of the network output confidence interval covering the labeled state change nodes reaches a preset requirement, the preset requirement value being obtained through statistical analysis of multiple experiments on the validation set; repeating the training process until the network's detection accuracy and confidence interval coverage on the validation set no longer improve, thus completing the model training.

[0051] The introduction of the time-series event perception model enables the solution to address the problem of missing observations caused by cloud cover. The principle is that the Bayesian uncertainty estimation module, through probabilistic modeling of the pulse firing threshold parameter, allows the network to still output judgment results with confidence intervals when observation data is missing or of reduced quality, avoiding the risk of misjudgment in scenarios with missing data due to a single deterministic output. The event-triggered iteration mechanism, through the detection of abrupt changes in pulse firing patterns, realizes the dynamic allocation of computing resources in the time dimension, improving the detection sensitivity and response speed of aquaculture activity mutation events.

[0052] The following explains the relationship between resource allocation and the calculation of various parameters in the time-series event-aware model. The time-series event-aware model sets a resource regulation function, denoted as [function name missing], between the intermediate layer and the event-triggered iterative mechanism. This is used to adjust the pulse firing threshold parameter and the memory allocation ratio of local sub-networks. The resource regulation function is calculated based on three data points: the average input pulse firing frequency, the amplitude of pulse firing mode mutations, and the resource utilization rate calculated using historical time windows. The calculation formula is as follows:

[0053] ;

[0054] In the formula, The average frequency of the input pulses, in units of Here, represents the amplitude of the pulse firing mode abrupt change, and represents the dimensionless normalized value. Calculate resource utilization rate for historical time windows, using dimensionless normalized values. , , The weighted coefficients are obtained through regression analysis of multiple sets of historical operating data, and satisfy the following conditions: .

[0055] When the value calculated by the resource regulation function When the data falls into different intervals, different adjustment methods are used to adjust the memory allocation ratio, the number of CUDA streams allocated, and the synaptic weight update step size. The intervals are divided as follows:

[0056] when When the local subnetwork is in a stable state, it is determined to be in a stable state. At this time, the local subnetwork maintains the basic memory allocation ratio, the corresponding basic CUDA stream allocation, and the synaptic weight update step size maintains the basic value.

[0057] when When the local subnetwork is in a transitional state, the memory allocation ratio of the local subnetwork is increased linearly, the number of CUDA streams allocated increases accordingly, and the synaptic weight update step size is reduced proportionally to improve the accuracy of local fine-grained calculations.

[0058] when When the event is determined to be in the event-triggered state interval, the local sub-network obtains the highest memory allocation ratio, corresponding to the highest number of CUDA streams allocated, and triggers the event-triggered iterative mechanism to recalculate the features of the adjacent time window. The synaptic weight update step size is reduced to the minimum value to ensure the detection accuracy of sudden changes in the state of aquaculture activities.

[0059] Interval boundary values , The specific values ​​were determined through statistical analysis of the resource regulation function calculation results and corresponding detection accuracy and calculation time in multiple sets of historical operating data, and verified through multiple iterative experiments. The basic values ​​of the basic memory allocation ratio, the basic number of CUDA stream allocations, and the synaptic weight update step size were determined through multiple experimental tests on hardware resource capacity and model computational complexity.

[0060] Optionally, the present invention also provides a method for implementing the above-mentioned method through a computer program, constituting a dynamic monitoring and identification system for marine aquaculture activities. The system is equipped with a readable storage medium storing program instructions. When the program instructions are executed in the computer, they perform the following steps: acquiring multi-source remote sensing data, constructing a discriminative feature set, calculating multi-scale morphologically complex quantitative features, running a multi-granularity cascaded discriminative model and a temporal event perception model, and generating a dynamic monitoring and identification report for marine aquaculture activities.

[0061] The specific implementation of step S01 is as follows: Synthetic Aperture Radar (SAR) data, optical remote sensing data, and meteorological reanalysis data covering the target sea area are acquired. All three types of data are registered using a unified geographic coordinate system to eliminate spatial misalignment caused by differences in the observation perspectives of different sensors. SAR data undergoes radiometric calibration, converting the original digital quantization values ​​into backscattering coefficients and applying topographic radiometric correction to eliminate the influence of topographic undulations on scattering intensity. Optical remote sensing data undergoes atmospheric correction, converting apparent reflectance into surface reflectance to eliminate atmospheric scattering and absorption effects. Meteorological reanalysis data is aligned with the remote sensing images according to the observation time, forming a spatiotemporally matched multi-source dataset.

[0062] The specific implementation of step S02 is as follows: For synthetic aperture radar data, polarization decomposition processing is performed on its fully polarized or dual-polarized echo signals. The decomposition method is based on the principle of target scattering matrix decomposition, separating the echo signal into surface scattering components, volume scattering components, and dihedral scattering components, and outputting the power ratio of each component as polarization decomposition parameters. For multi-scale texture statistics, multiple sliding window scales are set, and gray-level co-occurrence matrices are calculated at each scale. Statistical parameters such as contrast, correlation, energy, and homogeneity are extracted from the matrix. Simultaneously, local binary mode descriptors are calculated to characterize the spatial arrangement of gray-level distribution within the neighborhood of each pixel. For optical remote sensing data, normalized water indexes are calculated using near-infrared and short-wave infrared bands, and floating raft indexes are calculated using specific band combinations to enhance the spectral response differences between water areas and floating raft aquaculture facilities, respectively. The polarization decomposition parameters, multi-scale texture statistics, and spectral indices are concatenated pixel by pixel to construct a complete discriminative feature set.

[0063] The specific implementation of step S03 is as follows: For the candidate aquaculture patches extracted within the target area, the box counting method based on fractal geometry theory is used to quantify the boundary morphological complexity. Specifically, the candidate patch boundaries are binarized, multiple scale windows are set, the number of boxes traversed by the boundary at each scale is counted, and a linear regression is performed on the scale parameters and the box number sequence in a double logarithmic coordinate system. The absolute value of the regression slope is the fractal dimension estimate. Because artificial aquaculture facilities have regular geometric self-similar structures in their net cage arrays and raft aquaculture zones, their fractal dimension estimates usually fall within a stable range. The interval boundaries were determined through iterative statistical analysis of box counting experiments on multiple sets of known aquaculture samples and sea state noise samples. The reference range is as follows: Approximately 1.2, Approximately 1.6. Based on this, the generalized fractal dimension spectrum under different moment orders is further calculated to capture the spatial distribution differences of local singularities within candidate patches. The generalized fractal dimension spectrum calculation is implemented in parallel based on convolution operations to support efficient processing of large-scale marine imagery. The output fractal dimension estimate and the generalized fractal dimension spectrum together constitute a multi-scale morphological complexity quantification feature.

[0064] The specific implementation of step S04 is as follows: The input to the multi-granularity cascaded discriminant model is a concatenated vector of discriminant feature set and multi-scale morphologically complex quantifiable features. The multi-granularity scanning module sets a window scale set for the input feature vector. A sliding window scanning process is performed, constructing random forests and extreme random trees at each scale, outputting class probability vectors respectively. These vectors are then concatenated to form a multi-scale feature vector set. Each layer in the cascaded structure consists of multiple random forests and multiple extreme random trees operating in parallel. The class probability vectors output by each ensembler are concatenated with the original features and fed into the next layer, progressively increasing the feature representation depth. After training each layer, the output confidence is evaluated on the validation set. When the confidence falls below a threshold, a dynamic inter-layer feedback mechanism is triggered, adaptively increasing the number of decision trees in the current layer for local retraining. The refined features are then fed back to the previous layer for re-evaluation. The confidence threshold is approximately 0.75. The number of cascaded layers is automatically determined based on the stopping condition that cross-validation accuracy no longer improves, eliminating the need for manual setting of network depth hyperparameters. Finally, the probability of aquaculture activity categories for each candidate patch is output, forming the spatial distribution identification result of aquaculture activities.

[0065] The specific implementation of step S05 is as follows: The input to the time-series event perception model is a long-term synthetic aperture radar backscattering intensity sequence and meteorological reanalysis data such as wind speed, tides, and sea surface temperature for the corresponding time period. The above input data is converted into a pulse sequence through time coding, with the pulse frequency proportional to the input signal strength. The intermediate layer leak-integration firing neuron model simulates the membrane potential integration and firing mechanism of biological neurons. The neuron membrane potential accumulates with the input pulses, and after reaching the firing threshold, it outputs a pulse and resets. The synaptic plasticity mechanism dynamically adjusts the synaptic weights according to the pulse firing sequence relationship, achieving dynamic memorization of long-term dependencies. Resource regulation function. Based on the average frequency of input pulses , amplitude of sudden changes in pulse firing mode Calculate resource utilization using historical time windows The weighted summation is obtained, and the weighting coefficients are... , , The reference value was determined through regression experiments using historical operational data. Approximately 0.3, Approximately 0.5, Approximately 0.2. When When the system is in a stable state, basic resource allocation is maintained; when the system is in a transitional state, the memory allocation ratio is increased linearly and the synaptic weight update step size is reduced; when the system is in an event-triggered state, the local subnetwork is triggered to recalculate the adjacent time window, with reference values ​​for the interval boundaries. Approximately 0.35, The value is approximately 0.65. The Bayesian uncertainty estimation module uses variational inference on the pulse firing threshold parameter, enabling the network to output detection results with confidence intervals even when observations are missing. The output layer outputs the category of changes in aquaculture activity status and its corresponding confidence interval, forming the detection results of changes in aquaculture activity status.

[0066] The specific implementation of step S06 is as follows: Based on the spatial distribution identification results of aquaculture activities output in step S04 and the state change detection results of aquaculture activities output in step S05, the two are correlated and integrated in the spatiotemporal dimension. Specifically, the geographical range of aquaculture facilities is determined by the spatial distribution identification results, and the temporal change nodes of the activity state of each aquaculture patch and the corresponding confidence intervals are marked by the state change detection results. Finally, a dynamic monitoring and identification report of marine aquaculture activities is generated, which includes a spatial distribution map of aquaculture facilities, a temporal curve of state change, and confidence interval estimates, providing a technical basis for subsequent marine aquaculture management.

[0067] It should be noted that the key technologies of this invention include: First, the introduction of multi-scale morphological complexity quantification features. Fractal geometry theory is used to distinguish the geometric self-similarity of the regular boundaries of artificial aquaculture facilities from the irregular boundaries of natural sea noise in the feature space. This enables the cascaded fusion model to have stronger noise suppression capabilities under complex sea conditions, compensating for the shortcomings of pure semantic features in morphological discrimination. Second, the multi-granularity cascaded discrimination model replaces the gradient backpropagation training method with a deep forest architecture. Deep feature representation is achieved through layer-by-layer feature evolution and dynamic inter-layer feedback mechanisms, reducing the dependence on large-scale labeled samples. Simultaneously, the built-in feature importance measurement capability improves the training efficiency and interpretability of the model in remote sea areas with small sample sizes. Third, the temporal event perception model uses a Bayesian uncertainty estimation module to probabilistically model the pulse firing threshold parameter, enabling the system to output detection results with confidence intervals even when cloud cover leads to missing observation data. The event-triggered iterative mechanism dynamically concentrates computational resources on pulse pattern mutation regions, improving the detection sensitivity to abrupt changes in aquaculture activities. The three key technologies work together to complement each other, addressing the core challenges of feature discrimination, fusion modeling, and temporal detection. This enables the overall solution to operate robustly under complex sea conditions and cloud cover, achieving a complete technical loop from static spatial recognition to dynamic state detection.

[0068] It should be noted that in the monitoring of marine aquaculture activities, when there are frequent short-cycle state transitions such as cage placement and harvesting in the aquaculture area, the time interval between two adjacent synthetic aperture radar (SAR) passes may exceed the duration of the state change. This causes key change events to appear as single-point abrupt changes in the observation sequence, and these changes are highly overlapping with backscattering fluctuations caused by meteorological factors such as tidal surges and wave disturbances, making it difficult to distinguish them based solely on the amplitude differences of a single image. The reason for this technical problem is that the backscattering intensity of SAR targets on the sea surface is simultaneously modulated by the geometry of the aquaculture facilities, sea surface roughness, and meteorological boundary layer conditions. In short-cycle state change scenarios, the amplitude of backscattering changes caused by the state of the aquaculture facilities is close in magnitude to the background fluctuations caused by meteorological factors. Traditional threshold detection methods cannot effectively separate the signal source, while deep learning methods such as convolutional neural networks typically rely on deterministic outputs in time series modeling, lacking a quantitative description of detection uncertainty, which can easily lead to missed or false detections when the observation sequence is discontinuous. The usual solutions to the aforementioned technical problems are to increase the transit frequency of synthetic aperture radar to compress time intervals or to introduce multi-source satellite data to fill observation gaps. However, due to the limitations of satellite orbital periods, there is a physical upper limit to increasing the transit frequency; multi-source data fusion requires complex cross-sensor calibration processing under conditions of inconsistent sensor parameters, and when multi-source data are also affected by cloud and fog obstruction, the data loss problem is not fundamentally eliminated, and the output reliability of the deterministic fusion model under conditions of missing data cannot be guaranteed. This invention effectively solves this technical problem. The time-series event perception model synchronously encodes the backscatter intensity sequence and meteorological reanalysis data such as wind speed, tides, and sea surface temperature into pulse sequences. The leaked integral firing neurons distinguish meteorological background fluctuations and aquaculture activity state switching in the pulse firing time sequence pattern through the membrane potential integration mechanism. The synaptic plasticity mechanism retains the dynamic memory of historical time-series dependencies, enabling the model to continue judging state trends under sparse observation conditions. The Bayesian uncertainty estimation module transforms the model's uncertainty regarding missing data into confidence intervals by performing variational inference on the pulse firing threshold parameter, avoiding the forced output of deterministic conclusions during observation gaps. The event-triggered iterative mechanism automatically calls a local subnetwork to recalculate adjacent time windows when a pulse pattern mutation exceeds a threshold, concentrating computational resources on short-cycle state-switching events and improving the temporal resolution of rapidly changing events. The synergistic effect of these mechanisms enables the invention to robustly output confidence interval-based detection results of aquaculture activity state changes even under conditions of discontinuous observation sequences and significant meteorological background interference.

[0069] Specifically, the principle of this invention is as follows: Firstly, at the feature construction level, synthetic aperture radar polarization decomposition parameters describe the relative contribution of the target scattering mechanism, multi-scale texture statistics characterize the spatial arrangement of local grayscale values ​​in the image, and optical remote sensing spectral indices enhance the spectral response differences between water bodies and floating raft aquaculture facilities. These three types of features describe aquaculture facilities from three dimensions: physical mechanism, spatial structure, and spectral response, exhibiting stronger discriminative coverage compared to single features. Secondly, at the morphological quantification level, the multi-scale morphological complexity quantification algorithm is based on fractal geometry theory. It uses box counting to regress in a double logarithmic coordinate system to obtain fractal dimension estimates, and further calculates the generalized fractal dimension spectrum to characterize the local singularity distribution. Due to their regular arrangement, the boundary morphology of artificial aquaculture facilities maintains geometric self-similarity at multiple scales, and the fractal dimension estimates fall within a stable range; the boundary morphology of natural sea state noise lacks scale consistency, and the fractal dimension estimates deviate from this range. The two are separable in the fractal dimension feature space, allowing the cascaded fusion model to use this as a supplementary discriminant to suppress misjudgments due to sea state noise. Thirdly, at the feature fusion level, the multi-granularity cascaded discriminant model is based on a deep forest architecture. It simulates the local receptive field mechanism through a multi-granularity scanning module. The cascaded structure evolves deep features layer by layer. The dynamic inter-layer feedback mechanism triggers local retraining when the confidence level is insufficient, thus avoiding the gradient vanishing problem and reducing the dependence on large-scale labeled data. It is suitable for small-sample aquaculture labeling scenarios. Fourthly, at the temporal detection level, the temporal event perception model encodes the backscatter intensity sequence and meteorological reanalysis data into a pulse sequence. The leaked integral firing neuron, combined with the synaptic plasticity mechanism, realizes long-term temporal-dependent dynamic memory. The Bayesian uncertainty estimation module performs variational inference on the pulse firing threshold parameter, enabling the system to output detection results with confidence intervals even when cloud cover leads to observational gaps. The event-triggered iterative mechanism dynamically tilts computational resources to the pulse pattern mutation region, improving the detection sensitivity of aquaculture activity mutation events. The above-mentioned technical links work together from four dimensions: features, morphology, fusion, and time sequence, logically forming a complete closed loop from raw data to spatial distribution identification and then to state change detection. Each link provides complementary technical support for the core difficulties of complex sea conditions and data loss. Therefore, the technical solution of this invention can reliably realize the dynamic monitoring and identification of marine aquaculture activities under the above-mentioned working conditions.

[0070] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0071] The specific implementation of step S01 is as follows: acquire synthetic aperture radar data, optical remote sensing data and meteorological reanalysis data covering the target sea area, perform georegistration on the multi-source data, perform geometric correction using a rational function model, perform radiometric calibration and speckle noise filtering on the synthetic aperture radar data, and perform atmospheric correction on the optical remote sensing data to make the spatial resolution and coordinate system of various data consistent.

[0072] The specific implementation of step S02 is as follows: polarization decomposition is performed on the synthetic aperture radar data to extract the surface scattering component. Volume scattering component With dihedral scattering components The three satisfy the normalization constraint. In the window scale set Next, the gray-level co-occurrence matrix is ​​calculated, and statistics such as contrast, correlation, energy, and homogeneity are extracted to construct a multi-scale texture statistics vector. ,in This represents the total dimension of texture features, defaulting to 12. Normalized water index. With raft index The calculation formula is as follows:

[0073] ;

[0074] ;

[0075] In the formula, For green band reflectivity, For near-infrared reflectivity, The reflectance is in the red band; all three values ​​are dimensionless normalized values. To prevent division by zero of the stability coefficient, the empirical value is... Dimensions and , Consistency ensures that the denominator is not zero. and All are dimensionless normalization exponents, with values ​​ranging from 1 to 10. to Discriminant feature set vector It is composed of the above features:

[0076] ;

[0077] In the formula, , To determine the total dimension of the feature set, .

[0078] The specific implementation of step S03 is as follows: The fractal dimension of the target region boundary is calculated using the box counting method, with the box side length as the basis. The scale parameter represents the number of boxes required to cover the boundary. In a double logarithmic coordinate system and Perform linear regression to estimate the fractal dimension. The calculation formula is as follows:

[0079] ;

[0080] In the formula, The total number of scale samples. For the first Box side length at each scale To correspond to the number of boxes covered, and They are respectively and exist to The mean of the above, This is a dimensionless estimate of the fractal dimension; its absolute value is the final fractal dimension. Generalized fractal dimension spectrum. Calculated using the following formula:

[0081] ;

[0082] In the formula, It is the order of the moment. For scale Next The proportion of boundary points within each box to the total number of boundary points satisfies the following condition. , This represents the dimensionless generalized fractal dimension. The calculation process for the generalized fractal dimension spectrum is based on convolution operations, utilizing a sliding window to statistically analyze the point density within local boxes. The estimated fractal dimension of artificial breeding facilities typically falls within a stable interval. Interval boundaries and Box counting statistical experiments were conducted on multiple sets of known aquaculture facility samples and known sea state noise samples, and the results were analyzed through multiple iterations. Multi-scale morphologically complex quantifiable feature vectors were then determined. From each scale and Sequence splicing structure:

[0083] ;

[0084] In the formula, , The total number of samples for the moment order. The selected sequence of moment orders has a default range of [range]. to .

[0085] The specific implementation of step S04 is as follows: a multi-granularity cascaded discriminant model is used... and spliced ​​vector As input, The multi-granularity scanning module operates at the window scale. Below A sliding scan is performed, and a random forest and an extreme random tree are trained separately within each window. The output class probability vectors are then concatenated to form an enhanced feature vector. , ,in To enhance the total dimension of the feature vectors, it is determined by the sum of the dimensions of the class probability vectors output by the random forest and the extreme random tree at each window scale. The cascaded structure... Layer output probability vector The calculation formula is as follows:

[0086] ;

[0087] In the formula, For cascade layer numbering, For the first The number of random forests and extremely random trees in each layer, For the first Layer The class probability vectors output by each random forest. For the first Layer The class probability vector output by each extreme random tree. This is a dimensionless class probability vector, with the sum of its components being 1. The layer input is:

[0088] ;

[0089] In the formula, Each layer concatenates the original features with the output probability vector of the previous layer and passes the result to the next layer, thus achieving layer-by-layer feature evolution. When a layer outputs confidence... Below the threshold When this occurs, the layer is triggered to adaptively increase the number of decision trees and perform local retraining. The confidence level is determined through experimental statistical analysis of the output confidence distribution of the validation set, and is typically set to 0.7. The number of cascaded layers is automatically determined based on the cross-validation accuracy, and the final output is the spatial distribution identification result of aquaculture activities.

[0090] The specific implementation of step S05 is as follows: The temporal event perception model uses a long-time synthetic aperture radar backscattering intensity sequence. With meteorological reanalysis data, including wind speed Tidal height Sea surface temperature As input, it is converted into a pulse sequence through time coding. The pulse firing time of the time coding is... satisfy:

[0091] ;

[0092] In the formula, The length of the encoding time window, in units of , The input values ​​are normalized and are dimensionless. The upper bound for normalization is dimensionless and is usually taken as 1. Units are The dimensions on both sides of the equal sign are Membrane potential in a leaky integral firing neuron model The dynamic equations are as follows:

[0093] ;

[0094] In the formula, The membrane time constant is expressed in units of 1000 m / s. , For the first At time 1 neuron The membrane potential, in units of , For the first The first presynaptic neuron to the first Synaptic connection weights of postsynaptic neurons, in units of , For the first A sequence of pulses from a presynaptic neuron, with values ​​of 0 or 1, dimensionless. For the first Bias term for each neuron, in units of The dimensions on both sides of the equal sign are .when Reaching the issuance threshold At that time, an output pulse is generated and the membrane potential is reset to the resting potential. , The default value is 0, and the unit is... , Units are Synaptic plasticity mechanisms use pulse-time-dependent plasticity rules to update weights:

[0095] ;

[0096] In the formula, Postsynaptic pulse firing time With the timing of presynaptic pulse firing The difference, in units of , and These are the synaptic enhancement amplitude coefficient and the synaptic inhibition amplitude coefficient, respectively, in units of 1. , and These are the enhancement time constant and the suppression time constant, respectively, with units of 1. , This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Units are The dimensions on both sides of the equal sign are The Bayesian uncertainty estimation module addresses the threshold parameter. Introducing variational inference, let... ,in for The variational posterior mean, for Variational posterior variance, in units of The optimization is achieved by minimizing the lower bound loss of evidence, ensuring that the confidence interval coverage meets a preset requirement. This preset requirement is obtained through statistical analysis of multiple experiments on the validation set. Resource regulation function. The calculation formula is as follows:

[0097] ;

[0098] In the formula, The average frequency of the input pulses, in units of , for The corresponding weighting coefficients, in units of ,make Dimensionless numerical values represents the amplitude of the pulse firing mode abrupt change, and is a dimensionless normalized value. for The corresponding weighting coefficients are dimensionless. Resource utilization is calculated for historical time windows, and the values ​​are dimensionless normalized values. for The corresponding weighting coefficients are dimensionless. , , satisfy ,in for The normalized reference value, in units of , The value is dimensionless. , , This was obtained through regression analysis of multiple sets of historical operating data. When the time interval is determined to be a stable state, the local subnetwork maintains the basic memory allocation ratio and the basic synaptic weight update step size; when When this is the transitional state interval, the memory allocation ratio increases linearly, and the synaptic weight update step size decreases proportionally; when When the interval is determined to be an event-triggered state interval, the event-triggered iterative mechanism is triggered to recalculate the features of the adjacent time window, and the synaptic weight update step size is reduced to the minimum value. Interval boundary and The event trigger threshold was determined through statistical analysis of multiple sets of historical operational data and verification through iterative experiments. Used to determine whether the amplitude of the sudden change in pulse firing mode exceeds the triggering condition, i.e., when This triggers a recalculation of the local subnetwork. This is determined through statistical analysis of historical operational data. The output layer outputs the category of changes in aquaculture activity status and the corresponding confidence interval.

[0099] The specific implementation of step S06 is as follows: The spatial distribution identification results of aquaculture activities output in step S04 are superimposed and integrated with the detection results of aquaculture activity status changes and confidence intervals output in step S05. The results are then classified and summarized according to the target sea area, time segment and aquaculture facility type to generate a dynamic monitoring and identification report of marine aquaculture activities. The report includes the spatial boundary of the aquaculture area, the time node of status change and the corresponding confidence interval estimate.

[0100] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0101] This embodiment uses a nearshore marine aquaculture monitoring task as a background to illustrate the complete implementation process of the method described in this invention. The monitoring targets are the net cage array aquaculture areas and raft aquaculture zones distributed in this marine area, with a monitoring period covering 12 months, involving three states: net cage placement, existence, and harvest.

[0102] In step S01, technicians collected multiple synthetic aperture radar (SAR) images covering the target sea area, with a spatial resolution of 10m and a time interval of 12 days; simultaneously, optical remote sensing images for the corresponding time periods were collected, with a spatial resolution of 20m; meteorological reanalysis data had a temporal resolution of 6 hours and included three elements: wind speed, tides, and sea surface temperature. After georegistration and radiometric correction of the three types of data, a spatiotemporally aligned multi-source dataset was formed. The registration error was controlled within 0.5 pixels, and the dynamic range of the backscattering coefficient after radiometric calibration was [not specified]. dB.

[0103] In step S02, technicians perform polarization decomposition on the dual-polarization synthetic aperture radar data, extracting three types of polarization decomposition parameters: surface scattering power ratio, volume scattering power ratio, and dihedral scattering power ratio. The window scale for the multi-scale texture statistics is set to three levels: 3×3, 7×7, and 15×15. For each level, four statistical parameters of the gray-level co-occurrence matrix and local binary mode descriptors are extracted, resulting in a total of 15-dimensional texture features. The normalized water index and raft index are calculated based on the near-infrared and short-wave infrared bands, respectively, and together they constitute the spectral index features. The discriminant feature set comprises a total of 20-dimensional feature vectors.

[0104] In step S03, for the candidate aquaculture patch boundaries, the scale sequence of the box counting method is set to 2, 4, 8, 16, and 32 pixels, and linear regression is performed in a double logarithmic coordinate system, such as... Figure 2 As shown, the fractal dimension estimates for the cage array samples are concentrated in the range of 1.25–1.55, while the fractal dimension estimates for the sea state noise samples are mainly distributed in the range of 1.65–1.90. The two exhibit clear separability in the feature space. The generalized fractal dimension spectrum is calculated using the moment order. from to There are 11 discrete values ​​in total, and the corresponding generalized fractal dimension sequence is output. The quantifiable feature dimension of the multi-scale morphological complexity measure is 12-dimensional, which is concatenated with the discriminative feature set and then input into the subsequent model.

[0105] In step S04, the sliding window scale set of the multi-granularity cascaded discriminant model is set as follows: At each scale, a random forest (500 decision trees) and an extreme random tree (500 decision trees) were constructed, and the output class probability vectors were concatenated to form a multi-scale feature vector set. The cascaded structure was automatically determined to have 4 layers through cross-validation, and the confidence threshold for the dynamic inter-layer feedback mechanism was set to 0.75. The training samples included 600 samples from aquaculture areas and 600 samples from non-aquaculture areas, divided into training, validation, and test sets in a 7:2:1 ratio. The final recognition results are as follows. Figure 3 As shown, the spatial distribution identification results of aquaculture activities clearly present the spatial range of the cage array and raft aquaculture zone within the target sea area. To further describe the model's identification performance on different categories, the identification accuracy on the validation set is summarized in Table 1:

[0106] Table 1. Recognition Accuracy of Multi-Granularity Cascaded Discriminant Model on Validation Set

[0107]

[0108] In step S05, the input to the time-series event perception model is a sequence of 26 synthetic aperture radar backscattering intensities over 12 months, along with corresponding meteorological reanalysis data. Time encoding uses rate encoding, and the backscattering coefficients are normalized and mapped to pulse firing frequencies ranging from 0 to 200 Hz. The membrane time constant of the leaky integral firing neuron is referenced at 20 ms, the firing threshold is initialized to 0.5, and optimization is performed through variational inference. The weighting coefficients of the resource regulation function are... , , The interval boundary is taken , In the observation sequence, there were three time points where optical images were missing due to cloud cover. The Bayesian uncertainty estimation module output expanded confidence intervals for these missing time points, with the confidence interval width increasing from approximately 0.08 when there were no missing images to approximately 0.19. Figure 4 As shown, the time-series curves of the aquaculture activity state change detection results visually present the probability of state changes and the confidence interval range at each time point. The time-series event perception model successfully detected 10 out of 12 known state change nodes, and the threshold of the event trigger state interval was [not specified]. The activation trigger was performed 8 times, and each trigger recalculated the features of the adjacent time window. The pulse firing frequency and state change probability corresponding to each image are shown in Table 2.

[0109] Table 2 Summary of main output parameters of the time-series event-aware model

[0110]

[0111] In step S06, the technicians integrate the spatial distribution identification results of aquaculture activities with the detection results of changes in the state of aquaculture activities in a spatiotemporal dimension, such as... Figure 5 As shown, the dynamic monitoring and identification report presents the geographical extent of each aquaculture patch and the corresponding time-series curves of its status changes, along with confidence interval estimates for each node. The distribution differences of the fractal dimension estimates across different sample categories are shown below. Figure 6 As shown, this intuitively verifies the effectiveness of multi-scale morphological complexity quantification features in distinguishing regular aquaculture structures from irregular sea state noise. Figure 7 The distribution of the calculation results of the resource regulation function throughout the entire time-series monitoring process, as well as the trigger frequency statistics of the three state intervals, are presented, reflecting the dynamic allocation law of computing resources by the event-triggered iterative mechanism.

[0112] Compared to traditional single-sensor threshold classification methods, this invention simultaneously improves monitoring robustness across three dimensions: morphological quantification, feature fusion, and temporal modeling. Traditional methods rely on fixed thresholds to segment single images, failing to distinguish between morphologically similar aquaculture facilities and sea noise. This invention uses fractal geometry theory to characterize the geometric self-similarity of the regular boundaries of artificial facilities in the feature space, enabling the model to possess supplementary discriminative capabilities at the morphological level. Traditional deep neural networks are prone to overfitting under conditions of small-sample aquaculture labeled data. This invention's multi-granularity cascaded discriminative model replaces gradient backpropagation with a layer-by-layer feature evolution mechanism integrated by decision trees, reducing the dependence on the number of labeled samples. Traditional temporal analysis is forced to interpolate or skip missing nodes when observation data is missing, making it impossible to quantify the reliability of the output results. This invention's temporal event perception model explicitly outputs uncertainty in the form of confidence intervals through Bayesian probabilistic modeling, ensuring that the detection results still have interpretable confidence under cloud and fog conditions, effectively addressing the problem of discontinuous observations from a fundamental perspective.

[0113] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0114] Table 3. Variable Explanation Table (Part 1)

[0115]

[0116] Table 4. Variable Explanation Table (Part Two)

[0117]

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring and identification of sea farming activities based on multi-source remote sensing data fusion, characterized in that, Includes the following steps: Acquire multi-source remote sensing data, including synthetic aperture radar data, optical remote sensing data, and meteorological reanalysis data, and perform registration and radiometric correction. Extract polarization decomposition parameters and multi-scale texture statistics from synthetic aperture radar data and spectral index features from optical remote sensing data to construct a discriminative feature set. Calculate the multi-scale morphological complexity quantification features of the target area boundary to distinguish between regular aquaculture structures and irregular sea state noise; A multi-source feature cascade fusion model is constructed to fuse the discriminative feature set with multi-scale morphologically complex quantitative features, and output the spatial distribution identification results of aquaculture activities. A pulse neural network time-series detection framework is constructed to process long-time synthetic aperture radar backscatter intensity sequences and meteorological reanalysis data, and output the detection results of changes in aquaculture activity status and confidence interval estimates. Based on the spatial distribution identification results of aquaculture activities and the detection results of changes in the status of aquaculture activities, a dynamic monitoring and identification report of marine aquaculture activities is generated; The multi-scale morphological complexity quantification feature is calculated by a multi-scale morphological complexity quantification algorithm based on fractal dimension features. The box counting method is used to count the number of boxes that the target region boundary crosses under multiple scale windows. Linear regression is performed on the scale parameter and the number of boxes in a double logarithmic coordinate system, and the absolute value of the regression slope is used as the fractal dimension estimate. The multi-scale morphological complexity quantification algorithm further calculates the generalized fractal dimension spectrum under different moment orders to characterize the local singularity distribution characteristics of the target region. The calculation process of the generalized fractal dimension spectrum is based on convolution operation. The fractal dimension estimate of the artificial aquaculture facility falls within a stable interval. The interval boundary is obtained by iterative statistical analysis after box counting statistical experiments on known aquaculture facility samples and known sea state noise samples. The spiking neural network time-series detection framework is named the Time-Series Event Perception Model. The Time-Series Event Perception Model embeds a Bayesian uncertainty estimation module to perform variational inference on the pulse firing threshold parameter, enabling the network output to have the ability to estimate confidence intervals. The output layer outputs the category of changes in the state of aquaculture activities and the corresponding confidence intervals. The timing event-aware model sets a resource regulation function between the middle layer and the event-triggered iteration mechanism A formula for adjusting the pulse emission threshold parameter and the local sub-network display memory allocation ratio is: ； In the formula, is the average of input pulse firing frequency, unit is is the mutation amplitude of pulse firing mode, is a dimensionless normalized value is the historical time window computing resource occupancy rate, is a dimensionless normalized value , , is the weighted coefficient, the values of the three are obtained by regression experiment analysis on multiple sets of historical operation data, and satisfy ; When the value calculated by the resource regulation function When the data falls into different intervals, different adjustment methods are used to adjust the memory allocation ratio, the number of CUDA streams allocated, and the synaptic weight update step size. The intervals are divided as follows: when When the local subnetwork is in a stable state, it is determined to be in a stable state. At this time, the local subnetwork maintains the basic memory allocation ratio, the corresponding basic CUDA stream allocation, and the synaptic weight update step size maintains the basic value. when When the local subnetwork is in a transitional state, the memory allocation ratio of the local subnetwork is increased linearly, the number of CUDA streams allocated increases accordingly, and the synaptic weight update step size is reduced proportionally to improve the accuracy of local fine-grained calculations. when When the event is determined to be in the event-triggered state interval, the local sub-network obtains the highest memory allocation ratio, corresponding to the highest number of CUDA streams allocated, and triggers the event-triggered iterative mechanism to recalculate the features of the adjacent time window. The synaptic weight update step size is reduced to the minimum value to ensure the detection accuracy of sudden changes in the state of aquaculture activities. Among them, the interval boundary and This was determined through statistical analysis of multiple sets of historical operating data and verification through multiple iterative experiments. The network is configured with an event-triggered iteration mechanism. When the amplitude of a sudden change in the pulse firing pattern is detected in the input sequence and exceeds the threshold of the sudden change amplitude, the local sub-network is triggered to recalculate the features of the adjacent time window, thereby realizing the dynamic tilting allocation of computing resources to the abnormal change region.

2. The method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion according to claim 1, characterized in that, The polarization decomposition parameter refers to the scattering mechanism component obtained after polarization decomposition of the synthetic aperture radar echo signal, which is used to characterize the relative contributions of target surface scattering, volume scattering and dihedral scattering.

3. The method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion according to claim 2, characterized in that, The multi-scale texture statistics refer to the gray-level co-occurrence matrix parameters and local binary mode parameters calculated at multiple window scales, and the spectral indices include the normalized water index and the raft index.

4. The method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion according to claim 3, characterized in that, The multi-source feature cascade fusion model is named the multi-granularity cascaded discriminant model. It is built on a deep forest architecture and the input layer is associated with polarization decomposition parameters, multi-scale texture statistics, spectral indices and multi-scale morphological complexity quantitative features. The multi-granularity scanning module performs a sliding window scan on the input feature map, which is based entirely on the integrated voting implementation of random forest and extreme random tree.

5. The method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion according to claim 4, characterized in that, The cascaded structure of the multi-granularity cascaded discriminant model consists of multiple random forests and multiple extreme random trees in parallel at each layer. The class probability vector output by each layer is concatenated with the original features and used as the input of the next layer. The number of cascaded layers is automatically determined based on the cross-validation accuracy.

6. The method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion according to claim 5, characterized in that, The multi-granularity cascaded discriminant model sets up a dynamic inter-layer feedback mechanism. When the confidence of the output of a certain layer is lower than the confidence threshold, the structure of that layer is triggered to adaptively increase the number of decision trees for local retraining, and the refined features are fed back to the previous layer for re-evaluation.

7. The method for dynamic monitoring and identification of marine aquaculture activities based on multi-source remote sensing data fusion according to claim 6, characterized in that, The input layer of the time-series event perception model converts long-time synthetic aperture radar backscatter intensity sequences and meteorological reanalysis data into pulse sequences through time coding; the intermediate layer adopts a leaky integral firing neuron model, combined with synaptic plasticity mechanism to realize dynamic memory of long-time dependencies.

Images