Early disease warning method for prawn by coupling environmental factors with body surface spectral characteristics

Abstract

The application discloses a kind of coupling environmental factor and body surface spectral characteristic's early disease warning method of prawn, belong to the water product cultivation monitoring technical field based on machine learning;First, the group hyperspectral image of prawn group is obtained in actual cultivation environment, and multi-dimensional environmental factors are collected synchronously, and the basic data set for model training is constructed.Afterwards, a small amount of labeled group hyperspectral sample is used to fine-tune the designed generative enhancement network, so that it is adapted to the distribution of underwater hyperspectral imaging features, and then generates diversified virtual group spectral data.Subsequently, the original data and the generated expansion data are jointly input into the designed group-level disease warning model for training, to realize the discrimination of early disease risk of prawn group.After model training is completed, it is deployed in the online monitoring system of cultivation pond, realizes the real-time input of group spectrum and environmental factors, dynamic inference and risk warning output, to provide continuous, low-interference early disease warning support for cultivation management.

Description

Technical Field

[0001] This invention belongs to the field of aquaculture monitoring technology based on machine learning, and particularly relates to an early disease warning method for shrimp that couples environmental factors with surface spectral characteristics. Background Technology

[0002] Shrimp farming is a crucial component of the aquaculture industry in my country and globally, accounting for a significant proportion of the total output and economic value among marine aquaculture species. With the continuous development of intensive and high-density farming models, shrimp farming systems are becoming increasingly dependent on and sensitive to environmental conditions, leading to a rising frequency of disease outbreaks. Early-stage shrimp diseases are characterized by short incubation periods, rapid spread, and strong population-based activity. Failure to identify and intervene in a timely manner can easily trigger large-scale mortality and cause severe economic losses. Therefore, developing a shrimp health monitoring and disease early warning technology tailored to aquaculture ponds, possessing early detection capabilities and population-level warning functions, is of significant practical importance.

[0003] Current methods for monitoring shrimp diseases mainly rely on manual pond observation, physiological and biochemical indicator detection, or pathogen analysis. For example, health status is assessed by observing shrimp feeding behavior, swimming patterns, and changes in body color, or by periodically sampling for microscopic examination, molecular testing, and histopathological analysis. These methods typically require professional personnel, have inherent time lags, and struggle to achieve continuous, real-time monitoring of the entire shrimp population. Furthermore, frequent manual sampling and manipulation can cause stress in shrimp, increasing disease risk and failing to meet the demands of modern intelligent aquaculture for high timeliness and low interference. With the development of sensor and information technologies, some studies have begun to introduce environmental parameters such as water temperature, dissolved oxygen, pH, and salinity as reference indicators for shrimp health assessment, using threshold judgments or statistical models for risk warnings. However, the impact of environmental parameters on disease occurrence is often non-linear, coupled, and time-delayed; relying on a single or limited number of environmental indicators is insufficient to accurately reflect the true physiological and health status of shrimp. In addition, shrimp respond significantly differently to environmental changes in different farming areas, species and management models. Traditional early warning methods based on experience thresholds have limited generalization ability and a high false alarm rate.

[0004] In recent years, hyperspectral imaging technology has received widespread attention in the fields of agricultural and aquatic health monitoring due to its ability to simultaneously acquire spatial and continuous spectral information of targets. The pigment distribution, tissue structure, and metabolic products of shrimp surface tissues change under different health states, resulting in differentiated spectral characteristics in the visible and near-infrared bands. Compared to ordinary images or single-point spectral detection, hyperspectral imaging can provide a holistic perception of a group of targets without contact, offering a new technical approach for shrimp health status analysis. However, in actual aquaculture pond environments, factors such as water turbidity, lighting conditions, and changes in shrimp posture significantly affect spectral acquisition results, leading to large fluctuations in spectral characteristics and increasing the difficulty of feature extraction and modeling. On the other hand, early-stage shrimp disease samples are inherently difficult to obtain and have high annotation costs. In real aquaculture environments, the proportion of shrimp in the early disease stage is low, and their surface symptoms are not obvious. Accurate annotation using traditional detection methods often requires complex experimental procedures. This results in data-driven models generally facing problems of insufficient sample size and imbalanced class distribution during the training phase, severely restricting model performance and generalization ability.

[0005] In summary, existing technologies for early identification of shrimp diseases generally suffer from insufficient real-time performance, limited population perception capabilities, inadequate utilization of environmental factors, and a lack of training samples. There is still a lack of an early warning method for shrimp diseases that can integrate population surface spectral characteristics with multidimensional environmental factors in the actual environment of aquaculture ponds, and possess high robustness and generalization ability. Summary of the Invention

[0006] To address the above problems, this invention discloses a method for early disease warning in shrimp that couples environmental factors with surface spectral characteristics, comprising the following steps: S1, acquire hyperspectral images of shrimp populations in actual aquaculture environments and simultaneously collect multidimensional environmental factors; S2, preprocess the hyperspectral images of the population to form a training dataset containing population spectral features, environmental factor vectors and risk labels; S3 reconstructs the population spectral features in the training dataset into a population hyperspectral tensor, which is then used as input to the generative augmentation network to augment the dataset. First, the hyperspectral encoder processes the data to obtain latent feature representations. Then, a deterministic perturbation is applied to the latent feature representations in the latent feature space vector. The perturbed latent representations are input into the hyperspectral generative decoding network, and the population hyperspectral tensor is recovered through a layer-by-layer deconvolution structure. The population hyperspectral tensor is mapped to spectral-environment coupled features, which are then assigned the corresponding population disease risk labels of the original samples through label inheritance rules. S4 trains the population hyperspectral disease early warning deep network using the expanded dataset; real-time acquisition of population spectral features and environmental factor vectors in the shrimp farming environment, after preprocessing in S2, is input into the trained population hyperspectral disease early warning deep network to obtain population disease risk prediction results.

[0007] Preferably, the preprocessing of the hyperspectral image of the population in step S2 specifically includes: Original hyperspectral image Dark current correction and white reference correction were performed; then, spectral suppression of the water background region was applied to obtain the corrected population hyperspectral image. ; For the corrected hyperspectral image Principal component analysis (PCA) is performed along the spectral dimension to extract principal component images. Then, adaptive thresholding segmentation is applied to these images to obtain initial segmentation results. Morphological opening and closing operations are then performed on these initial segmentation results to obtain a refined target mask. Finally, the mask is reconstructed into the hyperspectral image space to obtain the target pixel set corresponding to the shrimp population. ; In obtaining the target pixel set Then, the mean value is calculated for each band dimension to obtain the population spectral feature vector. .

[0008] Preferably, the risk label specifically refers to: Health status labels for all sampled individuals Summing and averaging yields the time. Corresponding group-level disease risk labels ,in This indicates that the individual is at risk of early-stage disease. This indicates that the individual is in a healthy state; After processing according to the preset risk mapping rules, group-level disease risk labels are obtained. , This includes health status, early disease risk status, and significant disease risk status; The deterministic mapping function that maps proportional risk quantities to group disease risk levels is defined as:

[0009] in, This is the threshold for the early risk ratio; The threshold for the proportion of significant risks; the threshold parameter is determined through statistical analysis of historical aquaculture data: (1) When the proportion of abnormal individuals is lower than the threshold for significant risks. (1) When the overall physiological state of the group is stable; (2) When the abnormal proportion reaches When the disease begins to spread within the population; (3) when the abnormal proportion reaches By then, the health of the population had already been significantly compromised.

[0010] Preferably, the S3 process specifically includes: First, the population spectral characteristics Reconstructed as the initial spectral filling tensor ,right A spatially based perturbation term is introduced to recover the spatial non-uniformity of the population distribution and obtain the population hyperspectral tensor. Secondly, the population hyperspectral tensor The data is sequentially input into a cascaded three-layer 3D convolutional hyperspectral encoder network to perform joint feature extraction of the spatial and spectral dimensions, and output a latent feature vector. Secondly, for the latent feature vector Apply controlled deterministic perturbations to construct the perturbed latent representation. Construct the perturbation vector ,in, The direction vector is the standard normal distribution. Let be the disturbance amplitude coefficient; then the potential representation after disturbance is:

[0011] The perturbated latent representation The input is fed into a hyperspectral generation and decoding network consisting of fully connected mapping layers and skip connection structures for feature fusion and decoding, reconstructing a virtual population hyperspectral tensor that is consistent in spatial morphology and spectral continuity. Finally, the virtual population hyperspectral tensor is extracted using target region mask constraints. The corresponding virtual population spectral feature vector ,use Will Remapped to spectral-environment coupling features ,Will The expanded training dataset was obtained by pairing it with the original population disease risk labels. .

[0012] Preferably, the population hyperspectral disease early warning depth network is composed of a three-level structure consisting of a hyperspectral feature encoding module, an environmental factor coupling module, and a risk assessment and prediction module. First, the spectral feature encoding module reconstructs the group spectral feature vector into a tensor form with an explicit spectral channel structure by performing a dimension reconstruction operation. Then, the reconstructed spectral tensor... By performing multi-scale spectral convolution operations to simultaneously capture spectral variation patterns at different scales, a spectral feature representation that integrates multi-scale information is obtained. Introducing a spectral self-attention mechanism to study spectral features Global correlation modeling is performed to obtain spectral enhancement feature representation. ; The environmental factor coupling module achieves deep coupling between environmental factors and spectral features through explicit feature modulation, resulting in a coupled feature representation. ; The risk assessment and prediction module will integrate the environmental features. The mapping is output as an interpretable population disease risk level.

[0013] Preferably, the reconstructed spectral tensor Multi-scale spectral convolution operations are performed to simultaneously capture spectral variation patterns at different scales, specifically including three parallel one-dimensional convolution branches: First Branch A one-dimensional convolutional layer Conv1D with a kernel size of k=3 is used to extract fine-grained spectral variation features between adjacent bands; Second branch A one-dimensional convolutional layer Conv1D with a kernel size of k=5 is used to model the spectral response modes formed by the combination of medium-range bands; Third Branch A one-dimensional convolutional layer Conv1D with a kernel size of k=7 is used to capture overall spectral trend changes across a wide spectral range; The spectral features output from the three branches are concatenated along the channel dimension to obtain a spectral feature representation that integrates multi-scale information. .

[0014] Preferably, environmental factor vector For the input vector, The input is fed into a single fully connected layer for feature upscaling; the number of input neurons in the fully connected layer is 5, corresponding to 5 environmental factors, including water temperature parameters. Dissolved oxygen concentration parameters pH parameters of water body Salinity parameters and water flow velocity parameters The output layer has 32 neurons, and the output of the fully connected layer is activated by the ReLU function to obtain high-dimensional environmental embedding features. Subsequently, environmental embedding features Two parallel, independent fully connected layers are used for channel mapping to obtain the scale modulation parameters. and bias modulation parameters Ultimately, the parameters and Used to represent spectral enhancement features Perform channel-by-channel recalibration to obtain the feature representation after environment coupling. ; This is an element-wise multiplication operation, which involves multiplying elements at corresponding positions. The channel-by-channel recalibration operation represents the spectral enhancement features. Each channel is subjected to a linear transformation independently; the scale modulation parameters are... Expand to Same-dimensional scale modulation tensor; bias modulation parameters Expand to the corresponding bias tensor.

[0015] Preferably, it also includes a dynamic early warning process: The risk prediction results of the population disease at the corresponding time node obtained by the deep network prediction of the population hyperspectral disease early warning are calculated by risk aggregation in time and space dimensions to form a comprehensive risk assessment index at the pond level or regional level; when the comprehensive risk level exceeds the preset threshold, the system automatically triggers a graded early warning strategy.

[0016] Preferably, the dynamic early warning process specifically includes: Perform risk mapping calculations on the prediction results:

[0017] in, This is a risk level number, with values ​​of 1, 2, and 3; the higher the level number, the higher the risk level. Results of population disease risk prediction; Introducing a time series smoothing mechanism for risk scoring at consecutive time points. After averaging over a sliding time window, a smoothing risk index is obtained. :

[0018] in, Indicates the length of the time window. For time intervals; Multiple risk thresholds were set based on aquaculture experience and historical data. , Smoothing risk indicators Compare sequentially with the threshold, when It is determined to be in a normal state; when This is determined to be an early warning state of disease; when It has been determined to be in a high-risk warning state.

[0019] Compared with the prior art, the present invention has the following innovative features and beneficial effects: (1) This invention uses the hyperspectral characteristics of the shrimp population surface as the core information source for early disease perception in shrimp. Unlike the existing technology that mainly relies on individual observation or local sampling detection, this invention directly models the overall distribution of shrimp populations in aquaculture ponds, which can reflect the potential evolution trend of diseases at the population level and is more in line with the actual occurrence and spread of shrimp diseases.

[0020] (2) This invention proposes a modeling framework that deeply couples environmental factors with population spectral characteristics. Through the environmental factor coupling module, this invention realizes interactive modeling between environmental parameters and body surface spectral characteristics, effectively improving the accuracy and stability of early disease risk identification in shrimp.

[0021] (3) This invention introduces a generative model-driven population hyperspectral data augmentation mechanism. Addressing the issues of scarce early-stage disease samples and high annotation costs, it generates virtual population samples that conform to underwater hyperspectral distribution characteristics through targeted fine-tuning of a large image model, achieving data expansion across multiple scenarios and states. This generative augmentation approach, oriented towards population distribution and environmental changes, differs from traditional geometric or noise augmentation methods and significantly improves the model's generalization ability.

[0022] (4) This invention constructs an online reasoning and dynamic early warning system for actual aquaculture scenarios. It can achieve continuous monitoring, real-time early warning and strategy triggering without interfering with the normal growth of shrimp, providing technical support for intelligent and refined aquaculture management, and has obvious engineering application innovation. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall implementation process of the present invention.

[0024] Figure 2 This is a diagram illustrating the overall design structure of a generative augmentation network.

[0025] Figure 3 A deep network diagram for early warning of hyperspectral diseases in populations that senses environmental factors.

[0026] Figure 4 This is a comparison chart of the original population spectral statistical characteristics and the population spectral statistical characteristics after generative enhancement in an embodiment of the present invention.

[0027] Figure 5 This is a visualization of how the risk prediction model in this embodiment of the invention distinguishes the states of different groups.

[0028] Figure 6 This is a response curve diagram of the risk of group disease as a function of environmental disturbance in an embodiment of the present invention. Detailed Implementation

[0029] This invention proposes a method for early disease warning in shrimp that couples environmental factors with surface spectral characteristics. The method includes: S1. Population Spectrum and Environmental Coupling Data Acquisition and Risk Label Construction: First, underwater hyperspectral imaging of shrimp populations was performed in the actual environment of the aquaculture pond, and multi-dimensional environmental parameters such as water temperature, dissolved oxygen, pH, salinity, and flow velocity were collected simultaneously to construct time-aligned population spectral and environmental factor basic data. Subsequently, by collecting surface spectra of a small number of randomly sampled shrimp individuals and combining them with pathological test results, individual health status information was obtained and mapped into population-level early disease risk labels. Finally, the population spectral features were coupled with corresponding environmental factors to form a training dataset containing spectral information, environmental information, and risk labels.

[0030] S2, Population Spectral Feature Expansion Based on Generative Augmentation Network: First, based on the population spectral feature samples in the training dataset, the generative augmentation network is adapted to the hyperspectral domain structure and fine-tuned in terms of parameters. Then, controlled perturbations are introduced into the latent space of the fine-tuned generative network to reconstruct the original population spectral samples at multiple scales, generating virtual population hyperspectral samples covering different population densities, pose distributions, lighting conditions, and water flow states. Finally, the generated virtual samples inherit the corresponding population disease risk labels to construct the expanded training dataset for robust training of the subsequent population-level hyperspectral disease early warning model.

[0031] S3, Training of Population-Level Hyperspectral Disease Early Warning Network Model: First, multi-scale spatial-spectral joint encoding is performed on the input population spectral features; then, explicit interactive modeling is performed between the encoded spectral features and environmental factor vectors; finally, the early disease risk level of the shrimp population is output through the risk assessment and prediction module to achieve population-level health status prediction.

[0032] S4, Real-time Online Inference and Dynamic Early Warning of Population Spectra: First, hyperspectral data of the population and environmental factor data are collected in real time in the underwater environment of the pond; then, the population characteristics constructed in real time are input into the deployed disease early warning model for online inference to obtain the population disease risk prediction results at the corresponding time point; further, through risk aggregation calculation in the time and spatial dimensions, a comprehensive risk assessment index at the pond level or regional level is formed; finally, when the comprehensive risk level exceeds the preset threshold, the system automatically triggers a graded early warning strategy and performs online adaptive updates of the model according to the risk change trend, thereby realizing continuous health management and proactive intervention in the shrimp farming process.

[0033] The specific implementation process of the present invention will be further described below with reference to specific embodiments. The overall process is as follows: Figure 1 As shown.

[0034] S1. Collection of population spectral and environmental coupled data and construction of risk labels S1-1 Group Hyperspectral Image Acquisition and Standardization Processing. Underwater hyperspectral imaging devices were deployed at predetermined locations on the bottom and walls of the aquaculture pond to continuously scan the shrimp population within the aquaculture area at set times. The original hyperspectral images of the population acquired at each time point are denoted as follows: .

[0035] Next, the original hyperspectral image After hyperspectral normalization processing; specifically, this includes processing the original hyperspectral image... Dark current correction and white reference correction were performed to unify the illumination response at different acquisition times. Then, spectral suppression was applied to the water background region to obtain the corrected population hyperspectral image. .

[0036] S1-2 Population Target Region Extraction and Spectral Feature Construction. Corrected Population Hyperspectral Image. After target region mask constraint extraction, the pixel set corresponding to the shrimp population is obtained. . Specifically: For the corrected hyperspectral image Principal component analysis (PCA) is performed along the spectral dimension to extract principal component images. Then, adaptive thresholding segmentation is applied to these images to obtain initial segmentation results. Morphological opening and closing operations are then performed on these initial segmentation results to obtain a refined target mask. Finally, the mask is reconstructed into the hyperspectral image space to obtain the target pixel set corresponding to the shrimp population. .

[0037] In obtaining the target pixel set Then, the mean value is calculated for each band dimension to obtain the population spectral feature vector. :

[0038] in, Indicates time The spectral reflectance characteristics of the entire shrimp population at all times.

[0039] S1-3 Environmental factor collection and multidimensional environmental sequence construction. Simultaneously, a multi-parameter environmental sensing system was deployed to collect aquatic environmental information. The system... Environmental factors collected at different times constitute an environmental factor vector. .in, Indicates water temperature parameter, This represents the dissolved oxygen concentration parameter. Indicates the pH value of water. This represents the salinity parameter. This represents the water flow velocity parameter.

[0040] S1-4 Sampling Individual Testing and Health Status Quantification. (Time) Within the corresponding sampling period, the number of samples randomly drawn from the aquaculture ponds is The individual shrimp constitute the sampled set. Each sampled individual underwent surface spectral data collection and pathological testing to determine their health status. Therefore, a binary health status label is assigned to each sampled individual. .in This indicates that the individual is at risk of early-stage disease. This indicates that the individual is in a healthy state.

[0041] S1-5 Construction of Group-Level Disease Risk Labels. Health status labels for all sampled individuals. Summing and averaging yields the time. Corresponding group-level disease risk labels . After processing according to the preset risk mapping rules, group-level disease risk labels are obtained. .in, It includes three types: healthy status, early disease risk status, and significant disease risk status. The deterministic mapping function that maps proportional risk quantities to group disease risk levels is defined as:

[0042] in, This is the threshold for the early risk ratio; The threshold is the proportion of significant risks. This threshold parameter is determined through statistical analysis of historical aquaculture data: (1) When the proportion of abnormal individuals is lower than... (1) When the overall physiological state of the group is stable; (2) When the abnormal proportion reaches When the disease begins to spread within the population; (3) when the abnormal proportion reaches By then, the health of the population had already been significantly impaired. Ultimately, the population-level disease risk label was defined as... These correspond to the healthy state, the early disease risk state, and the significant disease risk state, respectively.

[0043] Ultimately, population spectral characteristics With environmental factor vector as input features , As the output of the dataset, it constitutes the training samples. The samples at all time points t together constitute the training dataset. .

[0044] S2. Enlarging the Population Spectral Feature Dataset Based on Generative Augmentation Networks This step is based on the training dataset obtained in S1. To address the limited number of hyperspectral samples from shrimp populations, a generative enhancement and simulation method for underwater hyperspectral feature distribution was developed.

[0045] Overall structural design of S2-1 generative augmentation network.

[0046] Figure 2 The overall architecture diagram of the generative augmentation network is shown below. The generative augmentation network uses the training dataset output from stage S1. As input. To adapt to the modeling requirements of generative augmentation networks for spatial structure, population spectral features are used. Reconstructed into a population hyperspectral tensor . Specifically: First, according to the S1 stage, the target pixel set is obtained. Then, the spectral characteristics of the population were... The spectral fill tensor is copied to all spatial locations within the target region along the band dimension, and pixels in non-target regions are set to zero to obtain the initial spectral fill tensor. ; Next, regarding A spatially based perturbation term is introduced to recover the spatial non-uniformity of the population distribution, forming a population hyperspectral tensor. This is used as a direct input to the generative model.

[0047] The overall network of the generative augmentation network constructed in this invention is denoted as . .in, For hyperspectral encoder networks, This is a decoding network for a hyperspectral generator.

[0048] S2-2 hyperspectral encoder network. Population hyperspectral tensor First, it passes through a hyperspectral encoder. Processing yields latent feature representations. The process includes the following hierarchical structure: Population hyperspectral tensor The algorithm sequentially passes through three convolutional layers, each using a 3x3x3 kernel, to extract joint features from both the spatial and spectral dimensions. Each convolutional layer is followed by a batch normalization layer and a ReLU nonlinear activation layer. Finally, the hyperspectral encoder outputs a latent feature vector. .

[0049] S2-3 Latent Space Perturbation and Multi-Scenario Virtual Sample Simulation. To achieve multi-scenario simulation, the latent feature space vector is subjected to... Apply deterministic perturbation and construct perturbation vector .in, The direction vector is the standard normal distribution. Let be the disturbance amplitude coefficient. Then the potential representation after disturbance is: .

[0050] By hierarchically controlling the amplitude of disturbance, the generated samples exhibit systematic differences in population density, spatial coverage, spectral reflectance intensity, and local structural morphology, which can be used to simulate the population distribution characteristics under different farming densities, shrimp postures, lighting conditions, and water flow disturbances.

[0051] S2-4 generator decoding network. The perturbed latent representation is: Input to hyperspectral generation and decoding network In this study, the population hyperspectral tensor is recovered through a layer-by-layer deconvolution structure. The process includes: latent vectors The input is a fully connected mapping layer; subsequently, the spectral dimension is gradually recovered through three transposed convolutional layers, with each transposed convolutional layer being symmetrical to the corresponding layer in the encoder. Furthermore, a skip connection structure is introduced during decoding to fuse shallow spatial structure features with deep semantic features, maintaining consistency in the spatial morphology and spectral continuity of the generated samples. The specific operation of the skip connection is as follows: the encoder's first... The feature map output of layer and the corresponding scale in the decoder The layer input features are concatenated along the channel dimension, where .

[0052] S2-5 Generates Sample Risk Labels and Expands the Dataset. Hyperspectral Tensor for Each Virtual Population. Based on the label inheritance rules, the group disease risk labels corresponding to the original samples are determined, and then generated. Corresponding tags .

[0053] virtual population hyperspectral tensor After the target region mask constraint processing described in S1, the target region mask of the shrimp population is obtained, and the spatial pixel set corresponding to the shrimp population in the virtual sample is determined. And for... Statistical aggregation of spectral information within the target area is performed to construct a virtual population spectral feature vector. .

[0054] virtual population hyperspectral tensor Remapped to spectral-environment coupling features Finally, the expanded training dataset is obtained. This dataset serves as the unified input for the subsequent training module of the population-level hyperspectral disease early warning model.

[0055] S3, Construction of a Deep Network Model for Early Warning of Population Hyperspectral Diseases This step uses the expanded training dataset obtained in S2 to construct an environmental factor-aware, population-based hyperspectral disease early warning deep network, EAP-GHNet. Figure 3 This is a deep network diagram for early warning of population-based hyperspectral diseases based on environmental factor perception. The network consists of a three-level structure: a hyperspectral feature encoding module, an environmental factor coupling module, and a risk assessment and prediction module.

[0056] S3-1 Hyperspectral Feature Encoding Module. Used for deep representation learning of population-level spectral features. First, the population spectral feature vector... (or virtual population spectral feature vector) After dimensionality reconstruction, it is reconstructed into a tensor form with an explicit spectral channel structure. .

[0057] Secondly, the reconstructed spectral tensor Multi-scale spectral convolution operations are performed to simultaneously capture spectral variation patterns at different scales. Specifically, this includes three parallel one-dimensional convolution branches: First Branch A one-dimensional convolutional layer Conv1D with a kernel size of k=3 is used to extract fine-grained spectral variation features between adjacent bands; Second branch A one-dimensional convolutional layer Conv1D with a kernel size of k=5 is used to model the spectral response modes formed by the combination of medium-range bands; Third Branch A one-dimensional convolutional layer Conv1D with a kernel size of k=7 is used to capture overall spectral trend changes across a wide spectral range.

[0058] The spectral features output from the three branches are concatenated along the channel dimension to obtain a spectral feature representation that integrates multi-scale information. .

[0059] Finally, to enhance the model's ability to model key bands and their interrelationships, a spectral self-attention mechanism is introduced to apply spectral features. Global correlation modeling is performed to obtain spectral enhancement feature representation. .

[0060] S3-2 Environmental Factor Coupling Module. Considering that independently modeling spectral features alone is insufficient to accurately reflect the true risk of disease, this module achieves deep coupling between environmental factors and spectral features through explicit feature modulation.

[0061] The environmental factor vector By the system The composition of environmental factors collected in real time. This includes: Indicates water temperature parameter, This represents the dissolved oxygen concentration parameter. Indicates the pH value of water. This represents the salinity parameter. This represents the water flow velocity parameter.

[0062] The environmental factor vector For the input vector, The input is fed into a single fully connected layer for feature upscaling; specifically, the fully connected layer has 5 input neurons and 32 output neurons. The output of this fully connected layer is activated by the ReLU function to obtain high-dimensional environmental embedding features. Therefore, this feature contains a combined response expression resulting from the fusion and dimensionality enhancement of five environmental factors. Subsequently, the environmental embedding feature... Two parallel, independent fully connected layers are used for channel mapping to obtain the scale modulation parameters. and bias modulation parameters Ultimately, the parameters and Used to represent spectral enhancement features Perform channel-by-channel recalibration to obtain the feature representation after environment coupling. ; This is an element-wise multiplication operation, which involves multiplying elements at corresponding positions.

[0063] The channel-by-channel recalibration operation represents the spectral enhancement features. Each channel is subjected to a linear transformation independently. The scale modulation parameters are then... Expand to Same-dimensional scale modulation tensor; bias modulation parameters Expand to the corresponding bias tensor.

[0064] S3-3 Risk Assessment and Prediction Module. Used for the feature representation of fused spectral and environmental information. This is mapped to an interpretable population disease risk level output. Specifically, Environmental coupling characteristics Global average pooling is performed along the spectral dimension to obtain a fixed-length group-level comprehensive feature vector. .

[0065] Furthermore, synthesize the feature vectors The data is then processed through two fully connected layers for nonlinear mapping and activated by the softmax function, ultimately outputting the population disease risk prediction result. The result is taken as the risk level with the highest probability as the model output.

[0066] During the model training phase, the weighted cross-entropy loss function is used as the optimization objective, and conventional model training methods are employed, which will not be elaborated upon here.

[0067] S4, Real-time Online Inference and Dynamic Early Warning of Population Spectrum S4-1 Real-time Population Spectral Acquisition and Synchronous Processing. During the operation of the pond underwater monitoring system, the underwater hyperspectral acquisition device acquires data at set time intervals. Continuous spectral imaging was performed on shrimp populations in a pond to obtain real-time hyperspectral data of the populations. .

[0068] Subsequently, real-time population hyperspectral data After the target region extraction and statistical modeling operations are consistent with those in stage S1, the real-time population spectral feature vector is obtained. .

[0069] Furthermore, the pond environment sensing system collects environmental factor data at the same time point to obtain an environmental factor vector. .

[0070] The population spectral features are concatenated with environmental factor vectors to construct online inference input features. .

[0071] S4-2 Online Group Disease Risk Inference Calculation Process. During the online inference phase, features are input in real-time. The data were fed into the pre-trained Environmental Factor Sensing Population Hyperspectral Disease Early Warning Network (EAP-GHNet) to obtain population disease risk prediction results. It is divided into three categories: healthy, early disease, and high-risk disease.

[0072] S4-3 A Temporal Modeling and Continuous Risk Assessment Mechanism for Population Disease Risk. After predicting the probability of population disease risk at a single time point, directly relying on instantaneous prediction results for early warning is easily affected by factors such as water disturbance, light changes, and short-term shrimp aggregation behavior, leading to unstable fluctuations in risk assessment. Therefore, this invention introduces a temporal modeling and continuous assessment mechanism for population disease risk to further process the model output. It consists of a single-time-point risk mapping layer and a time-series risk smoothing layer.

[0073] To transform the probabilistic model output into a continuous risk indicator that can be used for threshold determination, the system performs risk mapping calculations on the online prediction results:

[0074] in, The risk level is assigned a number, with values ​​of 1, 2, and 3. A higher number indicates a higher level of risk. This risk mapping layer can be viewed as a fixed-weight linear mapping that transforms the classification results into a numerical risk score.

[0075] Secondly, a time series smoothing mechanism is introduced. Risk scoring at consecutive time points. After averaging over a sliding time window, a smoothing risk index is obtained. :

[0076] in, Indicates the length of the time window. For time intervals.

[0077] S4-4 Graded Early Warning Judgment and Response Module. This module sets multi-level risk thresholds based on aquaculture experience and historical data. , Smoothing risk indicators The warning level is determined by comparing the data with the threshold values ​​in sequence.

[0078] when It is determined to be in a normal state; when This is determined to be an early warning state of disease; when It has been determined to be in a high-risk warning state.

[0079] Once the system enters an early warning state, it automatically outputs the corresponding pond-level early warning signal and marks and stores the population spectral data, environmental factor data, and risk prediction results for the current time point and the time periods before and after it, providing a basis for subsequent analysis and management decisions.

[0080] The S4-5 model online adaptive update module. When the system detects a high-risk state within multiple consecutive time windows, it constructs an online sample set from the online inference samples for the corresponding time periods. .

[0081] Furthermore, the loss function for online adaptive updates... The average of all samples is calculated:

[0082] in, Indicates the number of samples in the set. For an input sample in the set, This indicates the pseudo-label generated based on the warning level obtained for this sample. This indicates that the parameter can be updated. This represents a stability constraint, ensuring that the updates do not deviate excessively from the original training parameters.

[0083] Without affecting real-time inference, the system utilizes an online sample set. The weights of some network parameters in the online inference network of the disease early warning model are incrementally updated. Through constrained parameter fine-tuning, the model can gradually adapt to changes in the pond environment, achieving continuous and stable early warning of group diseases, while avoiding model performance degradation caused by insufficient online samples.

[0084] S5, Simulation Experiment Figure 4 The diagram shows a comparison between the statistical characteristics of the original population spectrum and the statistical characteristics of the population spectrum after generative enhancement. The mean distribution of the enhanced population spectrum in each band remains consistent with the original sample, while the standard deviation is significantly increased. This indicates that generative enhancement effectively expands the sample distribution space without destroying the original spectral structure characteristics, thereby improving the diversity and scene coverage of the training data.

[0085] Figure 5 The risk prediction model is visualized to distinguish the states of different groups. The feature space distribution shows that the disease risk levels of different groups form a clear clustering structure in the fused feature space. There is good separability among healthy groups, early-stage disease-risk groups, and high-risk disease groups, verifying that the spectral-environment coupled risk prediction model constructed in this invention has effective discriminative ability.

[0086] Figure 6 The results show the response curves of population disease risk to environmental disturbances. The experimental results indicate that as environmental factors gradually deviate from the suitable range, the population disease risk score exhibits a continuous upward trend, demonstrating a clear early response characteristic before reaching the dynamic threshold. These results demonstrate that the spectral-environment coupling model constructed in this invention can identify potential risks before disease outbreaks become apparent, achieving true early warning and significantly improving the foresight and reliability of aquaculture pond health management.

[0087] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0088] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for early disease warning in shrimp by coupling environmental factors and body surface spectral characteristics, characterized in that, Includes the following processes: S1, acquire hyperspectral images of shrimp populations in actual aquaculture environments and simultaneously collect multidimensional environmental factors; S2, preprocess the hyperspectral images of the population to form a training dataset containing population spectral features, environmental factor vectors and risk labels; S3 reconstructs the population spectral features in the training dataset into a population hyperspectral tensor, which is then used as input to the generative augmentation network to augment the dataset. First, the hyperspectral encoder processes the data to obtain latent feature representations. Then, a deterministic perturbation is applied to the latent feature representations in the latent feature space vector. The perturbed latent representations are input into the hyperspectral generative decoding network, and the population hyperspectral tensor is recovered through a layer-by-layer deconvolution structure. The population hyperspectral tensor is mapped to spectral-environment coupled features, which are then assigned the corresponding population disease risk labels of the original samples through label inheritance rules. S4 trains the population hyperspectral disease early warning deep network using the expanded dataset; real-time acquisition of population spectral features and environmental factor vectors in the shrimp farming environment, after preprocessing in S2, is input into the trained population hyperspectral disease early warning deep network to obtain population disease risk prediction results.

2. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 1, characterized in that: The preprocessing of the hyperspectral image of the population in S2 is specifically as follows: Original hyperspectral image Dark current correction and white reference correction were performed; then, spectral suppression of the water background region was applied to obtain the corrected population hyperspectral image. ; For the corrected hyperspectral image Principal component analysis (PCA) is performed along the spectral dimension to extract principal component images. Then, adaptive thresholding segmentation is applied to these images to obtain initial segmentation results. Morphological opening and closing operations are then performed on these initial segmentation results to obtain a refined target mask. Finally, the mask is reconstructed into the hyperspectral image space to obtain the target pixel set corresponding to the shrimp population. ; In obtaining the target pixel set Then, the mean value is calculated for each band dimension to obtain the population spectral feature vector. .

3. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 1, characterized in that: The risk label specifically refers to: Health status labels for all sampled individuals Summing and averaging yields the time. Corresponding group-level disease risk labels ,in This indicates that the individual is at risk of early-stage disease. This indicates that the individual is in a healthy state; After processing according to the preset risk mapping rules, group-level disease risk labels are obtained. , This includes health status, early disease risk status, and significant disease risk status; The deterministic mapping function that maps proportional risk quantities to group disease risk levels is defined as: in, This is the threshold for the early risk ratio; The threshold for the proportion of significant risks; the threshold parameter is determined through statistical analysis of historical aquaculture data: (1) When the proportion of abnormal individuals is lower than the threshold for significant risks. (1) When the overall physiological state of the group is stable; (2) When the abnormal proportion reaches When the disease begins to spread within the population; (3) when the abnormal proportion reaches By then, the health of the population had already been significantly compromised.

4. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 1, characterized in that: The specific process of S3 includes: First, the population spectral characteristics Reconstructed as the initial spectral filling tensor ,right A spatially location-based perturbation term is introduced to recover the spatial non-uniformity of the population distribution and obtain the population hyperspectral tensor. Secondly, the population hyperspectral tensor The data is sequentially input into a cascaded three-layer 3D convolutional hyperspectral encoder network to perform joint feature extraction of the spatial and spectral dimensions, and output a latent feature vector. Secondly, for the latent feature vector Apply controlled deterministic perturbations to construct the perturbed latent representation. Construct the perturbation vector ,in, The direction vector is the standard normal distribution. Let be the disturbance amplitude coefficient; then the potential representation after disturbance is: The perturbated latent representation The input is fed into a hyperspectral generation and decoding network consisting of fully connected mapping layers and skip connection structures for feature fusion and decoding, reconstructing a virtual population hyperspectral tensor that is consistent in spatial morphology and spectral continuity. Finally, the virtual population hyperspectral tensor is extracted using target region mask constraints. The corresponding virtual population spectral feature vector ,use Will Remapped to spectral-environment coupling features ,Will The expanded training dataset was obtained by pairing it with the original population disease risk labels. .

5. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 1, characterized in that: The hyperspectral disease early warning deep network consists of a three-level structure consisting of a hyperspectral feature encoding module, an environmental factor coupling module, and a risk assessment and prediction module. First, the spectral feature encoding module reconstructs the group spectral feature vector into a tensor form with an explicit spectral channel structure by performing a dimension reconstruction operation. Then, the reconstructed spectral tensor... By performing multi-scale spectral convolution operations to simultaneously capture spectral variation patterns at different scales, a spectral feature representation that integrates multi-scale information is obtained. Introducing a spectral self-attention mechanism to study spectral features Global correlation modeling is performed to obtain spectral enhancement feature representation. ; The environmental factor coupling module achieves deep coupling between environmental factors and spectral features through explicit feature modulation, resulting in a coupled feature representation. ; The risk assessment and prediction module will integrate the environmental features. The mapping is output as an interpretable population disease risk level.

6. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 5, characterized in that: The reconstructed spectral tensor Multi-scale spectral convolution operations are performed to simultaneously capture spectral variation patterns at different scales, specifically including three parallel one-dimensional convolution branches: First Branch A one-dimensional convolutional layer Conv1D with a kernel size of k=3 is used to extract fine-grained spectral variation features between adjacent bands; Second branch A one-dimensional convolutional layer Conv1D with a kernel size of k=5 is used to model the spectral response modes formed by the combination of medium-range bands; Third Branch A one-dimensional convolutional layer Conv1D with a kernel size of k=7 is used to capture overall spectral trend changes across a wide spectral range; The spectral features output from the three branches are concatenated along the channel dimension to obtain a spectral feature representation that integrates multi-scale information. .

7. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 5, characterized in that: The environmental factor coupling module is specifically as follows: Environmental factor vector For the input vector, The input is fed into a single fully connected layer for feature upsizing. The fully connected layer has 5 input neurons, corresponding to 5 environmental factors, including water temperature parameters. Dissolved oxygen concentration parameters pH parameters of water body Salinity parameters and water flow velocity parameters ; The output layer has 32 neurons, and the output of the fully connected layer is activated by the ReLU function to obtain high-dimensional context embedding features. ; Subsequently, environmental embedding features Two parallel, independent fully connected layers are used for channel mapping to obtain the scale modulation parameters. and bias modulation parameters Ultimately, the parameters and Used to represent spectral enhancement features Perform channel-by-channel recalibration to obtain the feature representation after environment coupling. ; This is an element-wise multiplication operation, which involves multiplying elements at corresponding positions. The channel-by-channel recalibration operation represents the spectral enhancement features. Each channel is subjected to a linear transformation independently; scale modulation parameters Expand to Same-dimensional scale modulation tensor; bias modulation parameters Expand to the corresponding bias tensor.

8. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 1, characterized in that: It also includes a dynamic early warning process: The risk prediction results of the population disease at the corresponding time node obtained by the population hyperspectral disease early warning deep network prediction are used to form a pond-level or regional-level comprehensive risk assessment index by risk aggregation calculation in the time and spatial dimensions. When the overall risk level exceeds a preset threshold, the system automatically triggers a tiered early warning strategy.

9. The method for early disease warning of shrimp by coupling environmental factors and body surface spectral characteristics as described in claim 8, characterized in that: The dynamic early warning process is as follows: Perform risk mapping calculations on the prediction results: in, This is a risk level number, with values ​​of 1, 2, and 3; the higher the level number, the higher the risk level. Results of population disease risk prediction; Introducing a time series smoothing mechanism for risk scoring at consecutive time points. After averaging over a sliding time window, a smoothing risk index is obtained. : in, Indicates the length of the time window. For time intervals; Multiple risk thresholds were set based on aquaculture experience and historical data. , Smoothing risk indicators Compare sequentially with the threshold, when It is determined to be in a normal state; when This is determined to be an early warning state of disease; when It has been determined to be in a high-risk warning state.

Images