Early enteromorpha identification method, medium and system

Abstract

The invention provides an early-stage enteromorpha identification method, medium and system, and belongs to the technical field of enteromorpha identification. According to the early-stage enteromorpha identification method and system, a dynamic feature fusion module based on inter-frame difference is used for extracting time sequence features of continuous frames, cross-modal information integration is achieved through a double-attention multi-modal feature fusion module, an improved YOLOV8 network model is constructed for target detection, and the identification accuracy of the early-stage enteromorpha is improved. Meanwhile, a liquid neural network model is combined with a multiband ratio method and a floating algae exponential function to carry out spectral feature recognition, enteromorpha components in mixed pixels are processed through a sub-pixel decomposition technology, and finally, multi-dimensional information such as the detection confidence coefficient, the spectrum matching degree, the enteromorpha abundance value and the target number is comprehensively analyzed and detected through a decision fusion algorithm. Accurate recognition and positioning of the early enteromorpha are achieved, and the technical problem that accurate recognition and distinguishing of the early enteromorpha are difficult to achieve at present is solved.

Description

Technical Field

[0001] This invention belongs to the field of seaweed identification technology, specifically, it relates to an early seaweed identification method, medium and system. Background Technology

[0002] As a large seaweed, *Ulva prolifera* can cause severe damage to the marine ecosystem due to its explosive growth. Therefore, early identification technology is of great significance in the field of marine environmental monitoring. Traditional *Ulva prolifera* identification methods mainly rely on remote sensing image data, analyzing the spectral reflectance characteristics of different bands to distinguish *Ulva prolifera* from the seawater background. These methods are widely used in marine environmental monitoring, algae disaster early warning, and marine ecological assessment, achieving target identification by calculating spectral parameters such as the normalized difference vegetation index (NDVI) and phytoplankton index. However, traditional spectral identification methods exhibit significant limitations when dealing with complex marine environments, especially when *Ulva prolifera* is in its early growth stage. Its chlorophyll content is low, and its spectral reflectance characteristics overlap significantly with those of other marine plants, plankton, and seabed sediments, leading to severe spectral confusion and a substantial decrease in identification accuracy. In current marine remote sensing monitoring technologies, because the spectral signal of early *Ulva prolifera* is weak and similar to background noise, existing single-band or multi-band analysis methods struggle to establish stable and reliable identification models. Particularly in mixed pixel environments, the spectral characteristics of different marine elements interfere with each other, making it impossible for traditional methods to effectively distinguish early *Ulva prolifera* from other similar targets. In other words, existing technologies present a technical problem: early-stage Ulva prolifera is highly similar to other marine plants and sediments in spectral characteristics, making it difficult to accurately identify and distinguish them. Summary of the Invention

[0003] In view of this, the present invention provides an early identification method, medium and system for Ulva prolifera, which can solve the technical problem in the prior art that early Ulva prolifera is highly similar to other marine plants and sediments in spectral characteristics, making it difficult to achieve accurate identification and differentiation.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides an early method for identifying *Ulva prolifera*, comprising: collecting marine monitoring video data and corresponding multispectral remote sensing image data of the area; simultaneously acquiring environmental parameter data of water temperature, salinity, dissolved oxygen, and pH at the corresponding spatiotemporal locations to construct a multimodal data source; preprocessing continuous monitoring video frames to extract a video sequence of 9 frames, including the target frame and the four frames before and after it, and extracting feature maps for each frame; processing the extracted feature maps using a dynamic feature fusion module based on inter-frame differences, fusing global and local features between continuous frames through global channel information and global spatial information to generate adaptive fusion features; inputting the environmental parameter data of water temperature, salinity, dissolved oxygen, and pH into a dual-attention multimodal feature fusion module, and combining it with the adaptive fusion features. Cross-attention calculation is performed, and cross-modal feature fusion is achieved through a multilayer perceptron to output multimodal fused features. An improved YOLOv8 network model is constructed, and the multimodal fused features are input for early detection of *Ulva prolifera* targets, outputting detection confidence, bounding box coordinates, and target quantity. Multispectral remote sensing image data is input into a liquid neural network model, and the multi-band ratio method combined with the phytoplankton exponential function is used to identify the spectral features of *Ulva prolifera*. Sub-pixel decomposition technology and endmember extraction algorithm are used to process the *Ulva prolifera* components in mixed pixels, outputting spectral matching degree and *Ulva prolifera* abundance value. A decision fusion algorithm is used to comprehensively analyze the detection confidence, spectral matching degree, *Ulva prolifera* abundance value, and target quantity, and the comprehensive identification confidence is calculated through a weighted fusion function to output the early *Ulva prolifera* identification results.

[0005] Specifically, the steps of the dynamic feature fusion module based on inter-frame differences are as follows: the feature maps of the first four frames are added together to obtain the preceding feature map, and the feature maps of the last four frames are added together to obtain the subsequent feature map. The global channel information is used to guide the channel reduction mechanism, and the depthwise separable convolution is used to obtain global spatial information to achieve adaptive fusion of temporal features.

[0006] Specifically, the global channel information is obtained by performing concatenated operations, average pooling operations, convolution operations, and sigmoid activation operations on the preceding feature map, the target frame feature map, and the subsequent feature map.

[0007] Specifically, the global spatial information is obtained by cascading operations on the preceding feature map, the target frame feature map, and the subsequent feature map after performing depthwise separable convolution and sigmoid activation operations respectively.

[0008] Specifically, the adaptive fusion feature is obtained by performing an element-wise multiplication operation between global spatial information and a calibration feature map. The calibration feature map is obtained by performing an element-wise multiplication operation between global channel information and a cascaded feature map, followed by a convolution operation.

[0009] Specifically, the dual-attention multimodal feature fusion module includes two independent dual-path feature encoders that process environmental modal data and visual modal data respectively. By projecting adaptive fusion features into visual query matrices and visual key matrices, and environmental parameter data into environmental query matrices and environmental key matrices, cross-modal correlation calculation and softmax normalization are performed to generate environmental enhancement features and visual enhancement features.

[0010] Specifically, the improved YOLOV8 network model includes a backbone network that integrates a dynamic feature fusion module based on inter-frame differences, a neck network that integrates a dual-attention multimodal feature fusion module, and a head detection network.

[0011] Specifically, the liquid neural network model includes a reservoir layer, an input connection layer, and an output connection layer. The number of neurons in the reservoir layer is calculated by multiplying the number of channels in the target frame feature map by 64 and then adding the square root of the pixel size of the monitoring video frame, and a sparse random connection method is used.

[0012] The extracted feature maps include: feature maps of the first 4 frames, feature maps of the target frame, and feature maps of the last 4 frames.

[0013] Specifically, the multimodal fusion features are obtained by cascading environmental enhancement features and visual enhancement features after processing by a multilayer perceptron.

[0014] Specifically, the connection probability of the liquid neural network model is calculated using a bounding box density function based on the density of the bounding box coordinates.

[0015] Specifically, the phytoplankton index function is used to distinguish the spectral characteristics of Ulva prolifera from other marine plants. The inputs include near-infrared reflectance, red light reflectance, green light reflectance, blue light reflectance, and short-wave infrared reflectance, and the output is the phytoplankton index value.

[0016] Specifically, the sub-pixel decomposition technology separates the *Ulva prolifera* components in the mixed pixels through a linear spectral mixing model, extracts the endmember spectra of the pure pixels, and the linear spectral mixing model solves for the abundance values ​​of each endmember based on the non-negative least squares method and satisfies the constraint that the sum of abundances is 1.

[0017] Specifically, the endmember extraction algorithm uses vertex component analysis and the N-FINDR algorithm to determine the endmember spectra in the mixed pixels. Vertex component analysis obtains the endmembers by finding the convex hull vertices of the data cloud after dimensionality reduction through principal component analysis.

[0018] Specifically, the spectral matching degree is the similarity coefficient between the phytoplankton index value and the standard Ulva spectral index, which is calculated using the Pearson correlation coefficient algorithm.

[0019] Specifically, the abundance value of *Ulva prolifera* is the proportion of *Ulva prolifera* components in the mixed pixels output by the sub-pixel decomposition technology, which is obtained by calculating the endmember abundance using a linear spectral mixing model.

[0020] Specifically, the comprehensive identification confidence level is calculated by summing the detection confidence level multiplied by a weighting coefficient, the spectral matching degree multiplied by a weighting coefficient, the abundance value of *Ulva prolifera* multiplied by a weighting coefficient, and the normalized value of the target quantity multiplied by a weighting coefficient.

[0021] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, perform the aforementioned early method for identifying seaweed.

[0022] A third aspect of the present invention provides an early identification system for seaweed, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

[0023] This invention constructs a multimodal data fusion method for early-stage *Ulva prolifera* identification, organically combining surveillance video data, multispectral remote sensing image data, and environmental parameter data. Utilizing temporal features and multidimensional information for collaborative analysis, it significantly enhances the ability to distinguish early-stage *Ulva prolifera* from other marine elements, solving the technical problem of insufficient accuracy in similar marine environments using traditional spectral identification methods. This method extracts the temporal variation patterns of *Ulva prolifera* through a dynamic feature fusion module based on inter-frame differences, and integrates complementary information from different data sources using a dual-attention multimodal feature fusion module. Even with similar spectral features, it can accurately distinguish *Ulva prolifera* by leveraging its unique growth dynamics and environmental response characteristics, effectively overcoming the limitations of single spectral data sources. Through the collaborative work of an improved YOLOv8 network model and a liquid neural network model, combined with sub-pixel decomposition technology and endmember extraction algorithms, it achieves precise separation and quantitative analysis of *Ulva prolifera* components in mixed pixels, significantly improving the accuracy and reliability of early-stage *Ulva prolifera* identification in complex marine environments. In summary, this invention solves the technical problem mentioned in the background art where early-stage *Ulva prolifera* has high similarity in spectral features to other marine plants and sediments, making accurate identification and differentiation difficult. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention.

[0025] Figure 2 This is a schematic diagram of the TDFF module architecture.

[0026] Figure 3 This is a schematic diagram of a dual-path DTFE architecture.

[0027] Figure 4 This is a schematic diagram of the interaction architecture between YOLOV8 and the liquid neural network. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0029] like Figure 1 The diagram shown is a flowchart of an early method for identifying *Ulva prolifera* provided in the first aspect of this invention. This method includes the following steps: S01. Collect marine monitoring video data and corresponding multispectral remote sensing image data, and simultaneously acquire environmental parameter data of water temperature, salinity, dissolved oxygen and pH value at corresponding spatiotemporal locations to construct a multimodal data source. The monitoring video data includes the spatial characteristics of early-stage Ulva prolifera in a star-shaped distribution. S02. Preprocess the continuous monitoring video frames to extract the target frame and the four frames before and after it, for a total of nine video frames. Extract the feature map of each frame through 3×3 convolution operation, batch normalization operation and ReLU activation function to obtain the feature map of the first four frames, the feature map of the target frame and the feature map of the last four frames. The number of channels of the target frame feature map is 512. S03. The first four frame feature maps, the target frame feature map and the last four frame feature maps are processed by a dynamic feature fusion module based on inter-frame differences. Global and local features between consecutive frames are fused through global channel information and global spatial information to generate adaptive fusion features. S04. Input the environmental parameter data of water temperature, salinity, dissolved oxygen and pH value into the dual attention multimodal feature fusion module, perform cross-attention calculation with the adaptive fusion feature, realize cross-modal feature fusion through multilayer perceptron, and output multimodal fusion feature; S05. Construct an improved YOLOV8 network model, which includes a backbone network that integrates the dynamic feature fusion module based on inter-frame differences, a neck network that integrates the dual-attention multimodal feature fusion module, and a head detection network. Input the multimodal fusion features to perform early detection of seaweed targets, and output the detection confidence, bounding box coordinates, and number of targets. S06. Input the multispectral remote sensing image data into the liquid neural network model. The number of reserve pool neurons in the liquid neural network model is set to the number of channels of the target frame feature map multiplied by 64 and then added to the square root of the pixel size of the monitoring video frame. The connection probability is calculated by the bounding box density function based on the density of the bounding box coordinates. The multi-band ratio method combined with the phytoplankton index function is used to identify the spectral features of Ulva prolifera. The Ulva prolifera components in the mixed pixels are processed by sub-pixel decomposition technology and endmember extraction algorithm. The spectral matching degree and Ulva prolifera abundance value are output. S07. The design decision fusion algorithm comprehensively analyzes the detection confidence, spectral matching degree, Ulva abundance value, and target quantity. It calculates the comprehensive identification confidence through a weighted fusion function. When the comprehensive identification confidence is greater than 0.85 and the Ulva abundance value is greater than 0.6, the early Ulva identification result is output as positive. When the comprehensive identification confidence is between 0.7 and 0.85 and the target quantity is greater than 3, the identification result is output as suspected positive. When the comprehensive identification confidence is less than 0.7, the identification result is output as negative. At the same time, the location coordinates, coverage area, and confidence score of the Ulva target are output as the final identification result.

[0030] The dynamic feature fusion module based on inter-frame differences obtains a preceding feature map by adding the feature maps of the first four frames and obtaining a subsequent feature map by adding the feature maps of the last four frames. It utilizes global channel information to guide a channel reduction mechanism and employs depthwise separable convolution to acquire global spatial information, achieving adaptive fusion of temporal features. The global channel information is calculated by performing concatenating, average pooling, convolution, and sigmoid activation operations on the preceding, target, and subsequent feature maps. The global spatial information is obtained by performing depthwise separable convolution and sigmoid activation operations on the preceding, target, and subsequent feature maps, followed by concatenation. The adaptive fused feature is obtained by element-wise multiplication of the global spatial information with the calibration feature map. The calibration feature map is obtained by element-wise multiplication of the global channel information with the concatenated feature map followed by a 1×1 convolution.

[0031] The dual-attention multimodal feature fusion module comprises two independent dual-path feature encoders, which process environmental modality data and visual modality data respectively. It generates environmental enhancement features and visual enhancement features by projecting the adaptive fusion features into visual query matrices and visual key matrices, and the environmental parameter data into environmental query matrices and environmental key matrices, followed by cross-modal correlation calculation and softmax normalization. The multimodal fusion features are obtained by cascading the environmental enhancement features and visual enhancement features through a multilayer perceptron.

[0032] The steps for establishing the training dataset for the improved YOLOV8 network model include: collecting monitoring videos of *Ulva prolifera* from different sea areas and seasons, labeling the early target locations and bounding box information of *Ulva prolifera*, recording corresponding environmental parameter data such as water temperature, salinity, dissolved oxygen, and pH, and dividing the dataset into training, validation, and test sets in a 7:2:1 ratio. The improved YOLOV8 network model is trained end-to-end using a focus loss function and a bounding box regression loss function. The initial learning rate is set to 0.001, the batch size is 16, and the training epochs are 300. The network parameters are optimized using backpropagation and a stochastic gradient descent optimizer. The detection confidence is the target detection probability value output by the improved YOLOV8 network model. The bounding box coordinates include the target center point coordinates, width, and height information. The target count is the number of *Ulva prolifera* targets detected in a single frame image.

[0033] The liquid neural network model structure includes a reservoir layer, an input connection layer, and an output connection layer. The number of neurons in the reservoir layer is calculated by multiplying the number of channels in the target frame feature map by 64 and then adding the square root of the pixel size of the monitoring video frame. A sparse random connection method is used, and the connection weights within the reservoir are initialized with a uniform distribution and a weight range of -0.5 to 0.5. The pixel size of the monitoring video frame is the product of the width and height of the input monitoring video frame. The steps for establishing the training dataset for the liquid neural network model include: collecting multispectral remote sensing image data, extracting spectral features of *Ulva prolifera*, seawater, other algae, and sediments, constructing a spectral feature library, and establishing a standard dataset for phytoplankton index. The training method of the liquid neural network model adopts a supervised learning method, training only the linear connection weights from the reservoir to the output layer, and optimizing the connection weights by minimizing the mean squared error loss function and ridge regression regularization.

[0034] The phytoplankton index function is used to distinguish *Ulva prolifera* from other marine plants based on their spectral characteristics. Inputs include near-infrared reflectance, red reflectance, green reflectance, blue reflectance, and short-wave infrared reflectance; output is the phytoplankton index value. The bounding box density function is used to calculate the density of bounding boxes within a unit image region. Inputs include the bounding box coordinates, the width and height of the monitoring video frame, and the number of targets; output is the connectivity probability value. The weighted fusion function is used to integrate multiple recognition information. Inputs include the detection confidence score, spectral matching degree, *Ulva prolifera* abundance value, and the number of targets; output is the overall recognition confidence score.

[0035] The spectral matching degree is the similarity coefficient between the phytoplankton index value and the standard *Ulva prolifera* spectral index, calculated using the Pearson correlation coefficient algorithm, with a value ranging from 0 to 1. The *Ulva prolifera* abundance value is the proportion of *Ulva prolifera* components in the mixed pixels output by the sub-pixel decomposition technology, calculated using the endmember abundance of the linear spectral mixing model, with a value ranging from 0 to 1. The comprehensive identification confidence is calculated by summing the detection confidence multiplied by 0.4, the spectral matching degree multiplied by 0.3, the *Ulva prolifera* abundance value multiplied by 0.2, and the target quantity normalized value multiplied by 0.1. The target quantity normalized value is the ratio of the target quantity divided by a preset maximum target quantity of 10.

[0036] The coverage area is obtained by summing the areas of each *Ulva prolifera* target calculated from the bounding box coordinates, in square meters. The confidence score is a percentage conversion of the overall identification confidence. The location coordinates are the corresponding positions of the bounding box coordinates in the geographic coordinate system, calculated using the geographic location information of the monitoring device and an image coordinate transformation algorithm. The connectivity probability value is calculated by multiplying the ratio of the number of targets to the pixel size of the monitoring video frame by an adjustment coefficient of 0.01, with a value range limited to between 0.05 and 0.3.

[0037] The sub-pixel decomposition technique separates the *Ulva prolifera* components in mixed pixels using a linear spectral mixture model, extracting the endmember spectra of pure pixels. The linear spectral mixture model solves for the abundance values ​​of each endmember using non-negative least squares and satisfies the constraint that the sum of abundances is 1. The endmember extraction algorithm uses vertex component analysis and the N-FINDR algorithm to determine the endmember spectra in mixed pixels. The vertex component analysis method obtains endmembers by finding the convex hull vertices of the data cloud after dimensionality reduction using principal component analysis. The multi-band ratio method enhances the spectral contrast between *Ulva prolifera* and the background by calculating the normalized vegetation index of reflectance in the near-infrared and red bands.

[0038] The specific implementation methods of the above steps are described in detail below.

[0039] The specific implementation of step S01 is as follows: Continuous monitoring video data is collected using high-definition cameras deployed at marine monitoring stations. The video resolution is set to 1920×1080 pixels, the frame rate is 30 frames per second, and the acquisition interval is 10 minutes of continuous recording per hour. Simultaneously, multispectral remote sensing image data of the corresponding sea area is acquired using satellite sensors with multiple bands including visible light, near-infrared, and shortwave infrared, with a spatial resolution of no less than 10m and a temporal resolution of once per day. Environmental parameter data such as water temperature, salinity, dissolved oxygen, and pH are collected in real time using water quality sensors installed at the same monitoring point, with a sampling frequency of once per minute and data accuracies of ±0.1℃, ±0.01, ±0.1mg / L, and ±0.01, respectively. A spatiotemporal synchronization mechanism is established to ensure the consistency of various data in terms of timestamps and geographic coordinates, constructing a multimodal data source containing visual, spectral, and environmental information. The purpose of this step is to provide a comprehensive data foundation for subsequent algorithms and to improve the accuracy of early *Ulva prolifera* identification by utilizing the complementarity of multimodal data.

[0040] The specific implementation of step S02 is as follows: Target detection frames are extracted from the continuous monitoring video stream. Four frames before and after the target frame are selected as the center to form a 9-frame temporal video sequence, ensuring the integrity of the temporal information. Preprocessing operations are performed on each frame, including image size standardization to 640×640 pixels, pixel value normalization to the 0-1 range, and necessary noise reduction. A basic feature extraction module of a convolutional neural network is used, employing a 3×3 convolution kernel to extract local image features. The convolution stride is set to 1, and zero-padding is used. Batch normalization is applied to standardize the convolution output, accelerating network convergence and improving training stability. A modified linear unit activation function is used to introduce a nonlinear transformation, enhancing the network's expressive power. After multiple convolutions, batch normalization, and activation operations, feature maps of the first four frames, the target frame, and the last four frames are obtained. The target frame feature map has 512 channels and a spatial size of 20×20. This step converts the original video data into a high-dimensional feature representation suitable for subsequent processing, laying the foundation for temporal feature fusion.

[0041] The specific implementation of step S03 is as follows: The extracted feature map sequence is processed using a dynamic feature fusion strategy based on inter-frame differences. The feature maps of the first four frames are fused into a preceding feature map through element-wise addition, and the feature maps of the last four frames are similarly fused into a subsequent feature map through addition, preserving temporal context information. A global channel information extraction mechanism is constructed, performing channel-dimensional concatenation operations on the preceding feature map, the target frame feature map, and the subsequent feature map to obtain a fused feature map with 1536 channels. Global average pooling is applied to compress the spatial dimension to 1×1, and then channel attention weights are generated through 1×1 convolution and the Sigmoid activation function, with weight values ​​ranging from 0 to 1. A global spatial information extraction mechanism is constructed, applying depthwise separable convolution operations to the three feature maps respectively. This operation includes two steps: depthwise convolution and pointwise convolution, effectively reducing computational parameters while maintaining feature extraction capabilities. After Sigmoid activation, channel dimensions are concatenated to generate spatial attention weights. The concatenated feature maps are calibrated through element-wise multiplication using global channel information, and then calibrated through 1×1 convolution to obtain a calibrated feature map. Finally, adaptive fusion features are generated through element-wise multiplication of global spatial information and calibration feature maps. The purpose of this step is to fully utilize the motion patterns and trends in temporal information to enhance the ability to identify the early star-shaped distribution characteristics of Ulva prolifera.

[0042] The specific implementation of step S04 is as follows: A dual-attention multimodal feature fusion architecture is constructed to process visual features and environmental parameter data. Environmental parameter data includes water temperature, salinity, dissolved oxygen, and pH. First, data standardization is performed, mapping each parameter value to the range of 0 to 1. Two independent dual-path feature encoders are designed to process environmental modality and visual modality data respectively. The environmental modality encoder adopts a multilayer perceptron structure, including an input layer, two hidden layers, and an output layer. The number of neurons in the hidden layers are 64 and 32, respectively, using modified linear units as the activation function. The visual modality encoder adopts a convolutional neural network structure to further encode the adaptive fusion features. The adaptive fusion features are projected into a visual query matrix and a visual key matrix through linear transformation, and the environmental parameter data is projected into an environmental query matrix and an environmental key matrix. The dot product of the visual query matrix and the environmental key matrix is ​​calculated to obtain the cross-modal correlation score, and the attention weights are obtained by applying the softmax function for normalization. The environmental features are weighted and summed using the attention weights to generate environmental enhancement features, and visual enhancement features are generated in the same way. Environmental enhancement features and visual enhancement features are input into a multilayer perceptron for nonlinear transformation, and finally, multimodal fusion features are obtained through cascaded operations. The purpose of this step is to achieve effective fusion of information from different modalities, enhance visual features using environmental parameters, and improve the robustness of *Ulva prolifera* recognition.

[0043] The specific implementation of step S05 is as follows: An improved YOLOv8 target detection network model is constructed. This model integrates a dynamic feature fusion module and a multimodal feature fusion module on the basis of the original YOLOv8 architecture. The backbone network adopts the CSPDarknet53 architecture, integrating the dynamic feature fusion module based on inter-frame differences from step S03, introducing temporal information processing capabilities during feature extraction. The neck network adopts a path aggregation network structure, integrating the dual-attention multimodal feature fusion module from step S04 to achieve multi-scale feature fusion and multimodal information integration. The head detection network includes a classification branch and a regression branch. The classification branch outputs the detection confidence of the *Ulva prolifera* target, and the regression branch outputs the bounding box coordinate information, including the center point coordinates, width, and height. Model training adopts an end-to-end approach, using a focus loss function to address the sample imbalance problem. The adjustment parameters α and γ for the focus loss are set to 0.25 and 2.0, respectively. The bounding box regression loss function uses the complete intersection-union loss to improve localization accuracy. The initial learning rate was set to 0.001, a cosine annealing scheduling strategy was adopted, the batch size was set to 16, and the total number of training epochs was 300. A stochastic gradient descent optimizer was used, with a momentum parameter set to 0.9 and a weight decay parameter set to 0.0005. The purpose of this step was to establish a high-precision model for detecting *Ulva prolifera* targets, outputting detection results including detection confidence, bounding box coordinates, and the number of targets.

[0044] The specific implementation of step S06 is as follows: A liquid neural network model is constructed to process multispectral remote sensing image data. This model has a natural advantage in processing time-varying signals and spectral sequences. The number of neurons in the reservoir layer is calculated by multiplying the number of target frame feature map channels (512) by 64 and adding the square root of the monitoring video frame pixel size (1920×1080), resulting in approximately 34,838 neurons. The reservoir layer employs a sparse random connection method. The connection probability is calculated based on the bounding box density function, which takes the bounding box coordinates, the width and height of the monitoring video frame, and the number of targets as inputs and outputs a connection probability value. The connection probability is calculated by multiplying the ratio of the number of targets to the monitoring video frame pixel size by an adjustment coefficient of 0.01, with a value range limited to 0.05 to 0.3. The connection weights within the reservoir layer are initialized using a uniform distribution and randomized, with a weight range of -0.5 to 0.5. The input connection layer is responsible for inputting multispectral image data into the reservoir layer, and the output connection layer uses a linear connection method to map the reservoir state to the final output. A multi-band ratio method combined with a phytoplankton index function was used for spectral feature identification. The phytoplankton index function input included reflectance values ​​in the near-infrared, red, green, blue, and short-wave infrared bands. Normalized vegetation index (NWRI) was used to enhance the spectral contrast between *Ulva prolifera* and the background. Sub-pixel decomposition technology was applied to process mixed pixels. A linear spectral mixture model was used to solve for the abundance values ​​of each endmember using the non-negative least squares method, satisfying the constraint that the sum of abundances is 1. The endmember extraction algorithm employed vertex component analysis and the N-FINDR algorithm. Vertex component analysis used principal component analysis to reduce dimensionality and then find the convex hull vertices of the data cloud to determine the endmembers. This step identifies *Ulva prolifera* features from the spectral dimension and outputs spectral matching degree and *Ulva prolifera* abundance values ​​as supplementary information.

[0045] The specific implementation of step S07 is as follows: A decision fusion algorithm is designed to comprehensively analyze multi-source identification information. This algorithm uses detection confidence, spectral matching degree, *Ulva prolifera* abundance value, and target quantity as input parameters. Spectral matching degree is calculated using the Pearson correlation coefficient algorithm to determine the similarity between the phytoplankton index value and the standard *Ulva prolifera* spectral index, with a value range of 0 to 1. *Ulva prolifera* abundance value is the proportion of *Ulva prolifera* components in the mixed pixels output by the sub-pixel decomposition technology, obtained through endmember abundance calculation using a linear spectral mixing model, with a value range of 0 to 1. Target quantity normalization uses a preset maximum target quantity of 10 as the normalization base to calculate the normalized value. The weighted fusion function uses a linear weighting method, with weights allocated as follows: detection confidence 0.4, spectral matching degree 0.3, *Ulva prolifera* abundance value 0.2, and target quantity normalized value 0.1. The weight settings are based on the reliability and importance of each information source. The comprehensive identification confidence is obtained through the above weighted summation calculation. A tiered decision-making mechanism is established: a positive result is output when the overall identification confidence score is greater than 0.85 and the abundance value of *Ulva prolifera* is greater than 0.6; a suspected positive result is output when the overall identification confidence score is between 0.7 and 0.85 and the number of targets is greater than 3; and a negative result is output when the overall identification confidence score is less than 0.7. Simultaneously, the coverage area of ​​the *Ulva prolifera* targets is calculated. The area of ​​each target is calculated using bounding box coordinates and summed, then converted to square meters based on the ratio of pixels to actual distance. The output location coordinates are converted from pixel coordinates to geographic coordinates using the geolocation information of the monitoring equipment and an image coordinate transformation algorithm. The purpose of this step is to make a final identification decision by integrating multi-source information, providing a complete identification result including identification category, location information, coverage area, and confidence score.

[0046] Further details are needed regarding the detailed structure of the YOLOv8 network model, which comprises three main components: a backbone network, a neck network, and a head network. The backbone network uses the CSPDarknet53 architecture, including multiple CSP bottleneck blocks and spatial pyramid pooling modules. It integrates a dynamic feature fusion module based on inter-frame differences and introduces temporal feature fusion operations at the feature maps of layers 3, 4, and 5. The neck network employs a path aggregation network structure, including a top-down feature pyramid network and a bottom-up path aggregation structure. It integrates a dual-attention multimodal feature fusion module, incorporating environmental parameter information during feature fusion. The head network uses a decoupled head design, with the classification head and regression head processing independently. The classification head outputs the probability of target presence at each grid location, while the regression head outputs the position and size information of the bounding box. The anchor box mechanism uses an anchor-free design, directly predicting the offset of the target center point relative to the top-left corner of the grid.

[0047] The detailed steps for establishing the YOLOv8 network model training dataset include data collection, annotation, partitioning, and preprocessing. In the data collection phase, monitoring videos of *Ulva prolifera* were collected from different sea areas, including the Yellow Sea, East China Sea, and Bohai Sea, covering data from different seasons such as spring, summer, and autumn to ensure data diversity and representativeness. In the annotation phase, manual annotation was used to mark the bounding boxes of early-stage *Ulva prolifera* targets in the YOLO format, including category labels and normalized coordinate information. Environmental parameter data for each sample were recorded, including water temperature (5–35℃), salinity (25–40), dissolved oxygen (3–12 mg / L), and pH (7.5–8.5). Data was randomly partitioned into training, validation, and test sets in a 7:2:1 ratio to ensure consistent data distribution across subsets. In the preprocessing phase, image resizing and data augmentation were performed, including random flipping, rotation, scaling, and color dithering, to increase data diversity and the model's generalization ability.

[0048] Further details are needed regarding the detailed structure of the liquid neural network model, which consists of a reservoir layer, an input connection layer, and an output connection layer. The reservoir layer is the core component of the model, containing a large number of randomly connected recurrent neurons. These neurons employ a leaky integral firing model, exhibiting dynamic characteristics with adjustable time constants. The internal connections within the reservoir utilize a small-world network topology, ensuring both dense local connections and sparse global connections. The input connection layer maps multispectral image data across different bands to different neurons in the reservoir, with connection weights randomly initialized and kept fixed during training. The output connection layer employs a linear readout method, mapping the high-dimensional state space of the reservoir to spectral matching degree and *Ulva prolifera* abundance values ​​as outputs.

[0049] The detailed steps for establishing the training dataset for the liquid neural network model involve spectral data collection, feature extraction, and standardization. Multispectral remote sensing image data was collected, including reflectance information for the visible light band (400-780 nm), near-infrared band (780-1400 nm), and short-wave infrared band (1400-3000 nm). Standard spectral features were extracted from *Ulva prolifera*, seawater, other algae such as *Sargassum* and *Wakame*, and sediments to construct a spectral feature library containing the spectral features of various land cover species. A standard dataset of phytoplankton indices was established, including spectral index reference values ​​for different *Ulva prolifera* growth stages and under different environmental conditions. A supervised learning method was used to train the model, optimizing only the linear connection weights from the reservoir to the output layer, and overfitting was prevented by minimizing the mean squared error loss function and using ridge regression regularization.

[0050] Further explanation is needed regarding the multifaceted effects and advantages of the interaction between the two models. The YOLOv8 network model focuses on target detection in the spatial domain, accurately locating the spatial position and geometric features of *Ulva prolifera* targets, but it has limitations in spectral feature recognition. The liquid neural network model excels at processing time-varying signals in the spectral domain, accurately identifying the biological characteristics of *Ulva prolifera* from a spectral perspective, but lacks spatial localization capabilities. The interaction and fusion of the two models achieves complementary advantages between spatial and spectral domain information. The spatial location information provided by YOLOv8 guides the liquid neural network to focus on spectral analysis of the region of interest, while the spectral matching and abundance information provided by the liquid neural network validates and supplements the detection results of YOLOv8. This interaction mechanism significantly improves recognition accuracy and robustness, especially in the early stages when *Ulva prolifera* features are not significant and the similarity to the background is high; the introduction of spectral information effectively reduces the false recognition rate. Compared to using the YOLOv8 network model alone, the interactive fusion scheme can better handle interference factors in complex marine environments, reducing missed and false detections caused by changes in light intensity, wave effects, etc.

[0051] The dynamic feature fusion technique based on inter-frame differences is one of the core innovations of this invention. Traditional target detection methods mainly rely on single-frame image information, ignoring the motion patterns and trends implied in the temporal context. This technique constructs three temporal feature maps—preceding, target, and subsequent—and utilizes a dual attention mechanism of global channel information and global spatial information to achieve adaptive fusion of temporal features. Compared to existing technologies, this method can capture the dynamic evolution characteristics of the early star-shaped distribution of Ulva prolifera, effectively distinguish the motion patterns of Ulva prolifera from those of other floating objects, and significantly improve the recognition accuracy and robustness in complex marine environments.

[0052] The dual-attention multimodal feature fusion technique achieves effective fusion of heterogeneous data sources by constructing a cross-attention computation mechanism between environmental and visual modalities. Traditional methods typically rely solely on visual or environmental information, lacking the synergistic utilization of multimodal information. This technique introduces environmental parameters such as water temperature, salinity, dissolved oxygen, and pH as prior knowledge into the visual feature learning process, generating environment-enhanced visual features through a dual-path feature encoder and a cross-modal attention mechanism. Compared to single-modal methods, this technique maintains stable recognition performance under complex and changing environmental conditions, especially in the early stages when the similarity between seaweed and the seawater background is high; the introduction of environmental parameter information significantly improves feature discrimination ability.

[0053] A liquid neural network and YOLOv8 collaborative processing architecture achieves deep fusion of spatial and spectral domain information. Traditional target detection methods mainly focus on spatial geometric features, while spectral analysis methods mainly focus on band reflectance characteristics; the two lack an effective collaborative mechanism. This technique constructs a spectral analysis branch of the liquid neural network to process multispectral data, complementing the spatial detection branch of YOLOv8, and achieving unified integration of information from the two domains through a decision fusion algorithm. Compared to single-domain methods, this architecture can verify recognition results from multiple dimensions, significantly reducing the false recognition rate, especially when the early morphological features of *Ulva prolifera* are not obvious, where the supplementary verification role of spectral features is particularly important.

[0054] The synergistic effect of these technological approaches forms a complete multi-dimensional, multi-modal, and multi-temporal early-stage Ulva prolifera identification system. Dynamic feature fusion provides temporal-dimensional feature enhancement, multi-modal fusion enables the synergistic utilization of environmental and visual information, and spatial-spectral collaborative processing ensures multiple verifications of the identification results. These three technological approaches support and verify each other, constructing a robust identification decision-making mechanism. Compared to existing single-technology approaches, this collaborative system can adapt to complex and ever-changing marine environments, maintaining high-precision identification performance even under conditions of weak early-stage Ulva prolifera features and numerous interfering factors, providing reliable technical support for marine ecological monitoring.

[0055] It should be noted that the present invention also solves the following technical problems: Traditional methods for identifying *Ulva prolifera* lack effective utilization of temporal information, failing to capture the dynamic changes during its growth process, resulting in insufficient timeliness and continuity of identification results. This invention utilizes a dynamic feature fusion module based on inter-frame differences to process a continuous nine-frame video sequence. By fusing global channel information and global spatial information, it effectively extracts the temporal change patterns and spatial distribution characteristics of *Ulva prolifera*. Through adaptive fusion of feature maps from the first four frames, the target frame, and the last four frames, this module can capture the complete dynamic process of *Ulva prolifera* from its emergence to its spread, significantly improving the timeliness and accuracy of early-stage *Ulva prolifera* identification.

[0056] Existing remote sensing identification technologies have limited accuracy when dealing with mixed pixel problems, especially when *Ulva prolifera* is mixed with seawater and other algae, making it difficult for traditional methods to accurately separate and quantify the proportions of each component. This invention introduces a liquid neural network model combined with sub-pixel decomposition technology, employing a linear spectral mixing model and endmember extraction algorithm to accurately separate *Ulva prolifera* components in mixed pixels and calculate their abundance values. This technology determines endmember spectra using vertex component analysis and the N-FINDR algorithm, and solves for the abundance values ​​of each endmember using non-negative least squares, effectively solving the technical challenge of quantitative analysis of *Ulva prolifera* components in mixed pixel environments, providing reliable technical support for early and accurate monitoring of *Ulva prolifera*.

[0057] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, perform the aforementioned early method for identifying seaweed.

[0058] A third aspect of the present invention provides an early identification system for seaweed, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

[0059] Specifically, the principle of this invention is as follows: This invention can solve the problem of spectral feature similarity in early identification of *Ulva prolifera*, mainly based on the technical principles of multimodal data fusion and feature enhancement. First, by constructing a multimodal data source including video surveillance data, multispectral remote sensing images, and environmental parameters, the traditional single spectral information is expanded into a multidimensional feature space. When spectral features are similar, feature information from other modalities can provide additional discriminative basis, thereby improving the accuracy and robustness of identification. Second, the dynamic feature fusion module based on inter-frame differences captures the dynamic features and spatial diffusion patterns of *Ulva prolifera* during its growth process by analyzing the temporal changes of continuous video frames. These temporal features are not possessed by other static targets such as marine plants and sediments, and can effectively distinguish different target types with similar spectral features. Third, the dual-attention multimodal feature fusion module achieves deep fusion of different modal data through a cross-attention mechanism, enabling visual features, spectral features, and environmental parameters to mutually verify and complement each other. When feature confusion occurs in one modality, other modalities can provide error correction and enhancement information, significantly improving the stability of identification. Furthermore, the liquid neural network model, combined with the multi-band ratio method and the phytoplankton exponential function, amplifies the spectral differences between early-stage *Ulva prolifera* and other targets through nonlinear feature transformation and spectral enhancement, improving the detection capability of weak signals. Finally, sub-pixel decomposition technology and endmember extraction algorithms, through precise unmixing of the linear spectral mixing model, can accurately separate the spectral contributions of each component in a mixed-pixel environment, calculate the true abundance values ​​of *Ulva prolifera* components, and effectively solve the quantitative identification problem under conditions of similar spectral features.

[0060] 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.

[0061] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be repeated in detail here.

[0062] The specific implementation of step S02 is as follows: Extract the target detection frame from the continuous monitoring video stream, and select four frames before and after the target detection frame to form a nine-frame time-series video sequence. Perform preprocessing operations on each frame, including normalizing the image size to 640×640 pixels, and the pixel value normalization is as follows: ; In the formula, The normalized pixel value, ranging from 0 to 1; These are the original pixel values; The minimum pixel value is 0. The maximum pixel value is 255. The basic feature extraction module of a convolutional neural network is used, employing a 3×3 convolution kernel for convolution operations. The convolution output is calculated as follows: ; In the formula, To output feature map at location The value; For the convolution kernel at position The weights; This refers to the pixel value of the input image at the corresponding location; For bias terms; For the spatial location index of the feature map; This represents the spatial index of the convolution kernel, with values ​​ranging from 0 to 2. The batch normalization operation is calculated as follows: ; In the formula, This is the output after batch normalization; Input features; This is the batch average. This represents the batch variance. For scaling parameters; For offset parameters; To prevent small constants from being divided by zero, the value is taken as... The ReLU activation function is calculated as follows: ; In the formula, The output of the activation function; The input value is . After multiple convolutions, batch normalization, and activation operations, the feature maps of the first 4 frames, the target frame, and the last 4 frames are obtained. The number of channels in the target frame feature map is set to 512, and the feature map space size is 20×20.

[0063] The specific implementation of step S03 is as follows: The extracted feature map sequence is processed using a dynamic feature fusion strategy based on inter-frame differences. The feature maps of the first four frames are fused through element-wise addition, and the preceding feature map is calculated as follows: ; In the formula, This is the preceding feature map; For the first Frame feature map; This is the time frame index. The feature maps of the next four frames are also fused using an addition operation. The subsequent feature map calculation is as follows:

[0064] ;

[0065] In the formula, For subsequent feature maps, a global channel information extraction mechanism is constructed. This mechanism performs a channel-dimensional concatenation operation on the preceding feature map, the target frame feature map, and the subsequent feature maps. The global channel information is calculated and represented as follows: ; In the formula, This is global channel information; Use the Sigmoid activation function; This is a 1×1 convolution operation; This is a global average pooling operation; For cascading operations; The target frame feature map is generated. A global spatial information extraction mechanism is constructed, and the global spatial information is calculated and represented as follows: ; In the formula, For global spatial information; This is a depthwise separable convolution operation. The adaptive fusion feature computation is represented as follows: ; In the formula, For adaptive fusion features; This is an element-wise multiplication operation.

[0066] The specific implementation of step S04 is as follows: A dual-attention multimodal feature fusion architecture is constructed to process visual features and environmental parameter data. The environmental parameter data is standardized and represented as follows: ; In the formula, These are the standardized environmental parameters; These are the original environmental parameter values; The mean; The standard deviation is denoted as . The adaptive fusion features are projected into a visual query matrix and a visual key matrix through a linear transformation. The projection process is as follows: , ; In the formula, For visual query matrix; For visual key matrix; This is the projection weight matrix; This is the bias vector. The environmental parameter data is projected into an environmental query matrix and an environmental key matrix, represented as follows: , ; In the formula, For environment query matrix; For the environment key matrix; This is the environmental modal projection weight matrix; This represents the environment modality bias vector. The cross-modal correlation calculation and attention weight generation are represented as follows: , ; In the formula, The attention weights for visual perception of the environment; Attention weights from environment to vision; The dimension of the key matrix is ​​64. The environmental enhancement features and visual enhancement features are calculated and represented as follows: , ; In the formula, Enhance environmental characteristics; Visual enhancement features. Multimodal fusion features are obtained through cascading operations: ; In the formula, This is a multimodal fusion feature; It is a multilayer perceptron.

[0067] The specific implementation of step S05 is as follows: An improved YOLOv8 object detection network model is constructed. This model integrates a dynamic feature fusion module and a multimodal feature fusion module on the basis of the original YOLOv8 architecture. The focus loss function is calculated as follows: ; In the formula, Loss of focus; This is the balance factor, with a value of 0.25. This is the focusing parameter, with a value of 2.0. Let be the predicted probability of the target category. The perfect intersection-union loss function is calculated as follows: ; In the formula, For complete intersection and union, the loss is compared; For intersection, union, and comparison; Center point of the prediction box and the center point of the true frame The square of the Euclidean distance; The diagonal distance of the smallest bounding rectangle that can simultaneously contain both the predicted bounding box and the ground truth bounding box; These are weight parameters; The parameter for measuring aspect ratio consistency is calculated as follows: ,in For the width and height of the actual bounding box, This is for predicting the width and height of the bounding box.

[0068] The specific implementation of step S06 is as follows: A liquid neural network model is constructed to process multispectral remote sensing image data. The calculation of the number of neurons in the reservoir is expressed as follows: ; In the formula, The number of neurons in the reserve pool is approximately 34,838; 512 represents the number of channels in the target frame feature map; 64 is the multiplier; and 1920×1080 is the pixel size of the surveillance video frame. The bounding box density function is calculated as follows: ; In the formula, The bounding box density; For the target quantity; To monitor the area of ​​video frames, the calculation is as follows: ,in The frame width is 1920. The frame height is 1080. The connection probability calculation is as follows: ; In the formula, 0.01 represents the connection probability; 0.05 is the adjustment coefficient; and 0.05 and 0.3 are the minimum and maximum values ​​of the connection probability, respectively. The normalized vegetation index (NDI) calculated using the multi-band ratio method is expressed as follows: ; In the formula, Normalized Difference Vegetation Index; Reflectivity in the near-infrared band; Here is the reflectance in the red light band. The phytoplankton exponential function is calculated as follows: ; In the formula, The phytoplankton index; For shortwave infrared band reflectivity; These represent the center wavelengths of the near-infrared, red, and short-wave infrared bands, respectively, in nanometers, with typical values ​​of 850, 650, and 1610 nanometers. The linear spectral mixing model calculation is expressed as follows: ; In the formula, For mixed pixels at wavelength Reflectance at that location; This represents the number of endmembers, typically ranging from 3 to 6. For the first Abundance values ​​of each endmember; For the first Each endmember at wavelength Reflectance at that location; This is the error term; This is an endmember index. Endmember abundance calculation satisfies the following constraints: , ; In the formula, the constraints ensure that the abundance values ​​are non-negative and that their sum is 1.

[0069] The specific implementation of step S07 is as follows: A decision fusion algorithm is designed to comprehensively analyze multi-source identification information. The spectral matching degree is calculated using the Pearson correlation coefficient algorithm, and is expressed as follows: ; In the formula, The value represents the spectral matching degree, ranging from 0 to 1. For the first The phytoplankton index value per pixel; This represents the average phytoplankton index. For the first Standard spectral index of *Ulva prolifera* per pixel; The average of the standard spectral indices; This represents the number of pixels. The normalized value of the target number is calculated as follows: ; In the formula, The normalized value of the target quantity; The maximum number of targets is preset, with a value of 10. The overall identification confidence score is calculated using a weighted fusion function, as shown below: ; In the formula, To comprehensively identify confidence levels; To test the confidence level; Here, represents the abundance value of *Ulva prolifera*; weighting coefficients of 0.4, 0.3, 0.2, and 0.1 correspond to the importance of each information source, respectively. The coverage area is calculated as follows: ; In the formula, Total coverage area, unit: ; For the first The width of each target in pixels; For the first The height of each target in pixels; This is the ratio of pixels to actual distance, in units of... / pixel. The position coordinate transformation calculation is represented as follows: ; In the formula, These are geographic coordinates, in meters. The geographic location coordinates of the monitoring equipment are in meters. It is a rotation matrix, specifically in the form of: ; The equipment's orientation angle is expressed in radians. These are pixel coordinates.

[0070] The global channel information acquisition method of the dynamic feature fusion module based on inter-frame differences is as follows: First, the preceding feature map, the target frame feature map, and the subsequent feature map are concatenated along their channel dimensions. Then, a global average pooling operation is applied to compress the spatial dimension to 1×1. Finally, channel attention weights are generated through a 1×1 convolution operation and a sigmoid activation function. The global spatial information acquisition method is as follows: A depthwise separable convolution operation is applied to each of the three feature maps. This operation includes two steps: depthwise convolution and pointwise convolution. The depthwise convolution calculation is represented as follows: ; In the formula, Output is a depthwise convolution; For the first The depthwise convolutional kernel weights of the channels; The corresponding position value in the input feature map; Spatial location index for depthwise convolution; For channel indexing; This represents the spatial index of the depthwise convolution kernel. Pointwise convolution is represented as follows: ; In the formula, The output is a pointwise convolution; For pointwise convolution weights; Input the number of channels; For output channel index; For input channel index; This is the spatial location index for pointwise convolution.

[0071] The dual-attention multimodal feature fusion module employs a multilayer perceptron structure for its environment modal encoder, consisting of an input layer, two hidden layers, and an output layer. The hidden layer contains 64 and 32 neurons, respectively. The visual modal encoder uses a convolutional neural network structure to encode the adaptively fused features. Cross-modal correlation calculation is performed by calculating the dot product of the visual query matrix and the environment key matrix, followed by normalization using the softmax function to obtain the attention weights.

[0072] In the liquid neural network model, the internal connection weights of the reservoir are initialized with a uniform distribution and a weight range of -0.5 to 0.5. The input connection layer is responsible for inputting multispectral image data into the reservoir, and the output connection layer uses a linear connection method to map the reservoir state to the final output. The sub-pixel decomposition technique uses a linear spectral mixture model, and solves for the abundance values ​​of each endmember based on the non-negative least squares method, satisfying the constraint that the sum of abundances is 1.

[0073] It needs to be explained that the pixel normalization formula The principle is to unify the image pixel values ​​under different lighting conditions to a standard range of 0 to 1, eliminating the influence of lighting changes on feature extraction. Compared with the traditional fixed threshold processing method, this normalization method can adaptively adjust the image contrast under different environmental conditions, significantly improving the robustness and accuracy of early seaweed identification.

[0074] Inter-frame fusion formula and The principle is to capture the motion patterns and morphological changes of seaweed in consecutive frames by aggregating temporal features. Compared with single-frame detection methods, this temporal fusion mechanism can effectively identify the dynamic characteristics of the early star-shaped distribution of seaweed, reduce the false detection rate caused by wave fluctuations and changes in light, and improve the stability of detection.

[0075] Adaptive fusion feature formula The principle is to achieve adaptive weight allocation of temporal features through a dual attention mechanism of global spatial information and global channel information. This mechanism can dynamically adjust the feature response according to the importance of different regions, which significantly improves the ability to identify weak seaweed signals in complex marine backgrounds compared with traditional uniform weight fusion methods.

[0076] Cross-modal attention formula The principle is to achieve cross-modal information fusion by calculating the correlation between visual features and environmental parameters. This mechanism can use environmental factors such as water temperature and salinity to guide the learning of visual features. Compared with single visual modality detection methods, multimodal fusion significantly improves the recognition accuracy and anti-interference ability under complex and changing environmental conditions.

[0077] Focus loss function The principle is to solve the problem of imbalance between positive and negative samples by dynamically adjusting the loss weights of different samples. This loss function enables the model to pay more attention to early-stage seaweed samples that are difficult to identify. Compared with the traditional cross-entropy loss, the focus loss significantly improves the model's learning effect on sparse positive samples and increases the detection rate of early-stage seaweed.

[0078] phytoplankton exponential function The principle is to use the difference in reflectance characteristics of seaweed in different bands for spectral identification. This index can effectively distinguish seaweed from background substances such as seawater and sediments. Compared with the traditional normalized vegetation index, the phytoplankton index has higher sensitivity and specificity for identifying seaweed in the marine environment.

[0079] Linear spectral mixing model The principle is based on the assumption that the reflectance of a mixed pixel is a linear combination of the reflectances of each endmember, where the abundance constraint condition is... and Ensuring the rationality of physical meaning, the model can accurately separate the Ulva prolifera components in the mixed pixels. Compared with the traditional hard classification method, the linear mixture model provides sub-pixel level Ulva prolifera abundance information, which significantly improves the accuracy of early Ulva prolifera identification.

[0080] Depthwise convolution formula in depthwise separable convolution and pointwise convolution formula The principle is to decompose the standard convolution into two independent steps: depthwise convolution to handle spatial correlation and pointwise convolution to handle inter-channel correlation. This decomposition method significantly reduces computational parameters while maintaining feature extraction capabilities. Compared with the standard convolution operation, depthwise separable convolution significantly reduces computational complexity and improves the model's running efficiency.

[0081] Comprehensive identification confidence formula The principle is to achieve the final decision by weighted fusion of multi-source information. This fusion strategy makes full use of the complementary advantages of spatial domain detection information, spectral domain matching information and abundance information. Compared with the decision method of single information source, multi-source fusion significantly reduces the false recognition rate and improves the reliability and confidence of the recognition results.

[0082] To better understand and implement this invention, a specific application scenario is provided below as Example 2: The technical team deployed an early-stage *Ulva prolifera* identification system in the sea area near Rizhao-Lianyungang to solve the technical problems of low efficiency and insufficient accuracy of traditional manual monitoring. This system, based on multimodal data fusion and deep learning-based early-stage *Ulva prolifera* identification methods, was practically applied during the peak *Ulva prolifera* green tide period from April 1st to May 30th, 2023.

[0083] The system first installed high-definition cameras on three buoys near Rizhao and Lianyungang, with three cameras on each buoy. The video resolution was set to 1920×1080 pixels, the frame rate to 30 frames per second, and ten minutes of video footage were collected every hour, totaling 14,400 video clips. These clips covered different weather conditions (sunny, cloudy, partly cloudy), different time periods (daytime, dusk, night), and different sea states (calm, light waves, large waves). A multispectral remote sensing image acquisition system was deployed simultaneously, using satellite sensors with eight bands including visible light, near-infrared, and shortwave infrared, a spatial resolution of 10m, and a temporal resolution of once per day. Water quality sensors were installed at the same monitoring points to collect environmental parameters such as water temperature, salinity, dissolved oxygen, and pH in real time. Data were collected within the ranges of temperature (8.5–25.3℃), salinity (28.2–35.7), dissolved oxygen (4.8–11.2 mg / L), and pH (7.8–8.4), with a sampling frequency of once per minute. Establish a GPS spatiotemporal synchronization mechanism to ensure the accurate correspondence of various types of data in timestamps and geographic coordinates, and construct a multimodal data source that includes visual information, spectral information and environmental information.

[0084] The system extracts target detection frames from a continuous monitoring video stream, and selects four frames before and after each frame to form a nine-frame time-series video sequence. Each frame undergoes standardization processing, including resizing to 640×640 pixels, normalizing pixel values ​​to the 0-1 range, and using Gaussian filtering for noise reduction. The video sequence is then processed by a dynamic feature fusion module based on inter-frame differences. Figure 2 The diagram shows the architecture of the TDFF module. This module uses a 3×3 convolutional kernel for feature extraction, with a stride of 1 and zero-padding. After batch normalization and ReLU activation, feature maps of the first four frames, the target frame, and the last four frames are obtained. The target frame feature map has 512 channels and a spatial size of 20×20.

[0085] The system employs a dynamic feature fusion strategy based on inter-frame differences. It fuses the feature maps of the first four frames into a pre-sequence feature map through element-wise addition, and similarly fuses the feature maps of the last four frames into a subsequent feature map. A global channel information extraction mechanism is constructed, performing channel-dimensional concatenation operations on the pre-sequence feature map, the target frame feature map, and the subsequent feature map to obtain a fused feature map with 1536 channels. Global average pooling is applied to compress the spatial dimension to 1×1, and channel attention weights are generated through 1×1 convolution operations and the Sigmoid activation function. Simultaneously, a global spatial information extraction mechanism is constructed, applying depthwise separable convolution operations to the three feature maps respectively. This operation includes two steps: depthwise convolution and pointwise convolution, effectively reducing computational parameters by 92.3% while maintaining feature extraction capabilities.

[0086] In the multimodal feature fusion stage, the system employs a dual-attention multimodal feature fusion module to process visual features and environmental parameter data, such as... Figure 3 The diagram illustrates the dual-channel DTFE architecture. Environmental parameter data includes water temperature, salinity, dissolved oxygen, and pH. First, the data is standardized, mapping each parameter value to the range of 0 to 1. Two independent dual-channel feature encoders are designed. The environmental modality encoder employs a multilayer perceptron structure, including an input layer, two hidden layers, and an output layer, with 64 and 32 neurons in the hidden layers, respectively. The visual modality encoder uses a convolutional neural network structure to encode adaptive fusion features. The adaptive fusion features are projected into a visual query matrix and a visual key matrix through linear transformation, while the environmental parameter data is projected into an environmental query matrix and an environmental key matrix. The dot product of the visual query matrix and the environmental key matrix is ​​calculated to obtain the cross-modal relevance score. The attention weights are obtained by normalization using the softmax function. The environmental enhancement features and visual enhancement features are then input into the multilayer perceptron for nonlinear transformation.

[0087] The system constructs an improved YOLOv8 network model for early detection of *Ulva prolifera* targets, such as... Figure 4 The diagram illustrates the interaction architecture of YOLOv8 and the Liquid Neural Network. This model integrates a dynamic feature fusion module and a multimodal feature fusion module on top of the original YOLOv8 architecture. The backbone network uses the CSPDarknet53 architecture, the neck network uses a path aggregation network structure, and the head detection network includes classification and regression branches. Model training is end-to-end, using a focus loss function to address the sample imbalance problem. The focus loss parameters α are set to 0.25, and γ is set to 2.0. The bounding box regression loss function uses the full intersection-over-union loss. The initial learning rate is set to 0.001, the batch size is set to 16, the total number of training epochs is 300, and a stochastic gradient descent optimizer is used.

[0088] A parallel liquid neural network model was constructed to process multispectral remote sensing image data. The reservoir layer contained 34,838 neurons, employing a sparse random connection method. The connection probability was calculated based on the bounding box density function, ranging from 0.05 to 0.3. The connection weights within the reservoir were initialized with a uniform distribution, ranging from -0.5 to 0.5. A multi-band ratio method combined with a phytoplankton index function was used for spectral feature identification. The phytoplankton index function input included reflectance values ​​for the near-infrared band (850 nm), red band (650 nm), green band (550 nm), blue band (450 nm), and short-wave infrared band (1610 nm). The spectral contrast between *Ulva prolifera* and the background was enhanced using a normalized vegetation index (NWRI). Sub-pixel decomposition technology was applied to process mixed pixels, and a linear spectral mixture model was used to solve for the abundance values ​​of each endmember based on the non-negative least squares method.

[0089] The system design decision fusion algorithm comprehensively analyzes multi-source identification information, using detection confidence, spectral matching degree, *Ulva prolifera* abundance value, and target quantity as input parameters. Spectral matching degree is calculated using the Pearson correlation coefficient algorithm, with a value ranging from 0 to 1. *Ulva prolifera* abundance value is the proportion of *Ulva prolifera* components in the mixed pixels output by sub-pixel decomposition technology, calculated using a linear spectral mixture model. A weighted fusion function is used for comprehensive analysis, with weights allocated as follows: detection confidence 0.4, spectral matching degree 0.3, *Ulva prolifera* abundance value 0.2, and target quantity normalized value 0.1. A tiered decision mechanism is established: a positive result is output when the comprehensive identification confidence is greater than 0.85 and the *Ulva prolifera* abundance value is greater than 0.6; a suspected positive result is output when the comprehensive identification confidence is between 0.7 and 0.85 and the target quantity is greater than 3; and a negative result is output when the comprehensive identification confidence is less than 0.7.

[0090] During actual operation, the system continuously monitored for 60 days, processing 20,000 valid video clips and corresponding multispectral remote sensing images, and 864,000 environmental parameter data points. The system was trained using the PyTorch deep learning framework, distributed across 5 GPUs with a batch size of 32. The Adam optimizer was used, with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001. The total training duration was 300 epochs. Detailed evaluation was performed on the validation set, and the resulting performance metrics are shown in Table 1. Table 1 Performance Indicators of Early-Stage Ulva Identification System

[0091] The recognition performance test results under different environmental conditions are shown in Table 2: Table 2 Comparison of recognition accuracy under different environmental conditions

[0092] The ablation experiment results of the system in terms of multimodal data fusion are shown in Table 3: Table 3 Results of Multimodal Data Fusion Ablation Experiment

[0093] The contribution analysis results of different algorithm modules are shown in Table 4: Table 4. Algorithm Module Contribution Analysis Table

[0094] After two months of continuous operation, the system successfully identified early-stage *Ulva prolifera* targets 1,847 times, including 1,023 positive identifications, 592 suspected positive identifications, and 232 negative identifications. Through manual verification, the system's overall identification accuracy reached 92.0%, with a false alarm rate below 7.6% and a false negative rate of 4.3%. The system's average response time is 3.5 seconds, meeting the requirements for real-time monitoring.

[0095] It should be noted that the variables involved in this invention are explained in detail in Tables 5 and 6.

[0096] Table 5. Variable Explanation Table (Part 1)

[0097] Table 6. Variable Explanation Table (Part Two)

[0098] 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. An early Enteromorpha identification method, characterized in that, The method comprises the following steps: Collecting sea area monitoring video data and corresponding multi-spectral remote sensing image data, synchronously acquiring water temperature, salinity, dissolved oxygen and pH value environmental parameter data of corresponding space-time position, and constructing a multi-modal data source; Preprocessing continuous monitoring video frames, extracting 9 frames of video sequence including the target frame and its previous and subsequent 4 frames, and extracting frame feature maps; a dynamic feature fusion module based on inter-frame difference is used to process the extracted feature maps, global and local features between continuous frames are fused through global channel information and global spatial information, and adaptive fusion features are generated; water temperature, salinity, dissolved oxygen and pH value environmental parameter data are input into a double-attention multi-modal feature fusion module, cross-attention calculation is performed with the adaptive fusion features, cross-modal feature fusion is realized through a multi-layer perception, and multi-modal fusion features are output; An improved YOLOV8 network model is constructed, multi-modal fusion features are input for early Enteromorpha prolifera target detection, detection confidence, bounding box coordinates and target number are output; multi-spectral remote sensing image data are input into a liquid neural network model, a multi-band ratio method is combined with a phytoplankton index function to identify Enteromorpha prolifera spectral features, a sub-pixel decomposition technology and an endmember extraction algorithm are used to process Enteromorpha prolifera components in mixed pixels, spectral matching degree and Enteromorpha prolifera abundance value are output; a decision fusion algorithm is used to comprehensively analyze detection confidence, spectral matching degree, Enteromorpha prolifera abundance value and target number, a weighted fusion function is used to calculate comprehensive recognition confidence, and early Enteromorpha prolifera recognition results are output.

2. The early Enteromorpha identification method according to claim 1, characterized in that, The steps of the dynamic feature fusion module based on inter-frame difference are as follows: performing addition operation on the first 4 frame feature maps to obtain a previous feature map, performing addition operation on the last 4 frame feature maps to obtain a subsequent feature map, using global channel information to guide channel reduction mechanism, using depth separable convolution to obtain global spatial information, and realizing adaptive fusion of time sequence features.

3. The early Enteromorpha identification method according to claim 2, characterized in that, The global channel information is obtained by performing cascade operation, average pooling operation, convolution operation and Sigmoid activation operation on the previous feature map, the target frame feature map and the subsequent feature map.

4. The early Enteromorpha identification method according to claim 3, characterized in that, The global spatial information is obtained by performing depth separable convolution operation and Sigmoid activation operation on the previous feature map, the target frame feature map and the subsequent feature map respectively, and then performing cascade operation.

5. The early Enteromorpha identification method according to claim 4, characterized in that, The adaptive fusion features are obtained by performing element-wise multiplication operation on the global spatial information and the calibration feature map, and the calibration feature map is obtained by performing element-wise multiplication operation on the global channel information and the cascade feature map, and then performing convolution operation.

6. The early Enteromorpha identification method according to claim 5, characterized in that, The double-attention multi-modal feature fusion module comprises two independent double-path feature encoders, which process environmental modal data and visual modal data respectively, projects the adaptive fusion features into a visual query matrix and a visual key matrix, projects the environmental parameter data into an environmental query matrix and an environmental key matrix, performs cross-modal correlation calculation and softmax normalization, and generates environmental enhanced features and visual enhanced features.

7. The early Enteromorpha identification method according to claim 6, characterized in that, The improved YOLOV8 network model specifically comprises a backbone network containing a dynamic feature fusion module based on inter-frame difference, a neck network integrating a double-attention multi-modal feature fusion module, and a head detection network.

8. The early Enteromorpha identification method according to claim 7, characterized in that, The liquid neural network model specifically comprises a reservoir layer, an input connection layer and an output connection layer, the number of neuron nodes in the reservoir layer is obtained by multiplying the number of channels of a target frame feature map by 64 and then adding the square root of the pixel size of a monitoring video frame, and a sparse random connection mode is adopted.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions, and the program instructions are used to execute the early Enteromorpha identification method in any one of claims 1-8 when running in the computer.

10. An early Enteromorpha identification system, characterized in that, The system is any one of a computer, a server and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

Images