Disease detection method fusing water surface and underwater information fish school dynamic space-time distribution characteristics

Abstract

The invention provides a disease detection method fusing water surface and underwater information fish school dynamic space-time distribution characteristics, visual image data of a water surface data source and acoustic data of an underwater data source are combined to judge the fish school health degree, and depth distribution of a fish school in water is measured through a water pressure sensor. A distribution evolution model of a fish school along with time is established through dynamic space-time modeling, gathering and dispersing states of the fish school and change rules of the fish school in different time periods are collected according to the position relation of the fish school in a three-dimensional space of an aquaculture water area, and a dynamic space-time model of the fish school is obtained. Identifying the density abnormal change and the space-time dynamic rule of the activity of the fish school through a dynamic space-time model so as to detect diseases; according to the invention, disease detection is realized by identifying the density abnormal change and the space-time dynamic law of the activity of the fish school, and the disease diffusion risk can be perceived in advance.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, and in particular to a disease detection method that integrates surface and underwater information to identify the dynamic spatiotemporal distribution characteristics of fish populations. Background Technology

[0003] Traditional methods for detecting fish diseases mainly rely on manual observation. Fish farmers need to spend a lot of time and energy regularly inspecting fish ponds or aquaculture areas. Moreover, this method is highly subjective and easily affected by factors such as the observer's experience, weather, and lighting, making it difficult to accurately detect problems in the early stages of disease and often missing the best time for prevention and control. Summary of the Invention

[0004] This invention proposes a disease detection method that integrates surface and underwater information to identify the dynamic spatiotemporal distribution characteristics of fish populations. By identifying abnormal changes in fish density and the spatiotemporal dynamic patterns of their activities, disease detection can be achieved, allowing for early detection of the risk of disease spread.

[0005] The present invention adopts the following technical solution.

[0006] A disease detection method integrating surface and underwater information on the dynamic spatiotemporal distribution characteristics of fish schools is proposed. This method combines visual image data from surface data sources and acoustic data from underwater data sources to determine the health of fish schools. It also measures the depth distribution of fish schools in the water using a water pressure sensor. A dynamic spatiotemporal model is established to create a model of the fish school's distribution evolution over time. The method collects data on the fish school's aggregation and dispersion states in the three-dimensional space of the aquaculture area, as well as their variation patterns at different times (such as day and night, and different growth stages), to obtain a dynamic spatiotemporal model of the fish school. This model is then used to identify abnormal changes in fish school density and the spatiotemporal dynamic patterns of their activities, thereby detecting diseases and recognizing the risk of disease spread in advance.

[0007] The water surface data source uses an image data acquisition device positioned above the water surface. This device includes a camera with high-definition resolution (4K) and telephoto zoom (20x). The camera is mounted on a fixed bracket above the fish farming area, preferably at an adjustable height of 3-5 meters. The height should ensure sufficient coverage of the water surface and clear observation of the fish's movements in the upper layers. The camera must be waterproof, dustproof, sunproof, and have automatic image stabilization to adapt to complex outdoor environments. It is also equipped with an adjustable gimbal to allow for easy adjustment of the shooting angle, capturing images of the fish's activities near the water surface from all angles.

[0008] The data acquisition process is pre-set with dynamic variations in acquisition frequency and duration. The preferred video recording parameters are set to record once every 15 minutes for a duration of 10 seconds. During the acquisition process, supplementary lighting, such as four 50W LED lights, can be turned on to ensure data acquisition during low-light periods such as early morning, evening, or cloudy days.

[0009] The underwater data source utilizes an underwater acoustic data acquisition array, which comprises multiple underwater acoustic sensors arranged at fixed intervals (uniformly distributed at 5-10 meters, with vertical spacing of 3-5 meters, adjustable according to area size, but horizontal spacing should be less than 20 meters and vertical spacing less than 10 meters). These sensors are fixed to specially designed underwater supports to ensure stability and prevent displacement or damage from water flow. The sampling frequency range is set between 20kHz and 50kHz to accurately capture the sound signals generated by fish swimming, feeding, and breathing. By adjusting the gain parameters, the sensors adapt to different intensities of sound input, enabling the detection of weak, normal swimming sounds, as well as abnormal sounds from fish under stress or illness, ensuring the acquired sound data has rich information content.

[0010] The underwater data acquisition from the underwater data source is synchronized with the data acquisition from the surface data source.

[0011] The distribution evolution model includes data on the distribution evolution of fish schools over time, which is obtained through data preprocessing of visual image data from surface data sources and acoustic data from underwater data sources. Data preprocessing includes surface image denoising, underwater sound denoising, and data fusion. The specific method for denoising water images is as follows: Due to the complex outdoor environment, water images are easily affected by factors such as light reflection and water surface ripples, resulting in salt-and-pepper noise and Gaussian noise. For salt-and-pepper noise in water images, a median filtering algorithm (the verified filtering operator is 5×5) is used. This algorithm selects the median value within the neighborhood of a pixel to replace that pixel value, effectively removing isolated noise points. When dealing with blurring caused by subtle ripples on the water surface, Gaussian filtering is used through convolution kernel operations to smooth the image based on a Gaussian function to reduce the impact of noise (Gaussian filter operator size is 3×3 or 5×5). The underwater sound noise reduction method is as follows: The main background noise in aquaculture areas consists of water flow noise and equipment operation noise (such as aerators and water pumps), which interferes with the acquisition of fish sound signals. The method involves first acquiring the sound information of water flow and equipment, and then obtaining the frequency domain characteristics of the water flow and equipment through frequency domain transformation. After frequency domain transformation of the acquired data, the sound signal is decomposed into sub-bands of different frequencies. Low-frequency water flow noise and high-frequency equipment noise are removed to extract the pure fish sound signal. Fourier frequency domain transformation and Laplace transform are used for frequency domain feature transformation. The specific data fusion method is as follows: the surface images and underwater fixed-point sound data are fused at a unified time; when there are differences in the frequency and start time of image acquisition and sound acquisition, time synchronization algorithms using interpolation and extrapolation are used to make the two accurately aligned in the time dimension to ensure the continuity of the data.

[0012] The detection method includes the use of acoustic signature sensor matrix data inside the water body; The modeling process of the dynamic spatiotemporal model includes the extraction of the dynamic spatiotemporal characteristics of the fish school. The extracted spatiotemporal features include individual features, surface group features, and water body fish school features. The specific method for extracting individual features is as follows: In the preprocessed images of the data acquired from the water surface, target detection and recognition algorithms are used to identify the bounding box of each fish, and then the morphological features of the fish's body length, body width, and body shape coefficient are extracted; through continuous analysis of multiple frames of images using a tracking algorithm, the instantaneous swimming speed, acceleration, and turning angle of the fish are obtained; combined with the acoustic data of the uppermost layer near the water surface, the physiological characteristics of the fish's movement frequency and feeding activity are determined. The specific method for extracting the characteristics of the upper surface population is as follows: based on individual characteristic statistics, the distribution density of the activity level of the fish population is calculated, that is, the distribution of fish of different activity levels (preferably 5 levels: inactive, generally inactive, moderate, generally active, and active) within a unit water volume; cluster analysis and density peak algorithm are used to identify the aggregation center, subgroup division, and overall shape and size changes of the population; through vector analysis of the swimming direction of the fish population, the migration trend, diffusion or contraction rate of the population is determined to extract the dynamic evolution characteristics of the fish population in the upper water surface space. The specific method for extracting the characteristics of fish schools in water is as follows: Based on the characteristics of the fish population on the upper surface, the data collected by each sensor is compared and integrated by combining the data of the acoustic sensor matrix inside the water body. The triangulation method and hyperbolic positioning algorithm are used to establish the relationship between acoustic intensity, acoustic density, time interval and distance to the fish population, number of fish population and activity level. Based on the sound frequency change caused by the Doppler effect, the swimming speed of the fish school is calculated. The magnitude of the sound signal and the distance to the fish school are established through the inverse proportional relationship model. By combining the characteristics of fish schools on the surface and in the water body, the density, distribution, change cycle and activity probability distribution characteristics of the overall fish school can be obtained.

[0013] The disease detection method is carried out through a dynamic spatiotemporal model, assuming that there are differences in the average amount of movement, cumulative amount of movement, instantaneous amount of movement, movement range and movement area between healthy fish groups and diseased fish groups inside and on the surface of the water.

[0014] The data and detection inside the water body are performed using acoustic sensors (vibration signals). However, due to the turbidity of the water inside the aquaculture body, the video acquisition range is limited to the upper surface layer of the water body (the activity of fish in the range of about 0~30CM depth).

[0015] The training method for the dynamic spatiotemporal model is as follows: Based on the collaborative data collection from surface optical and underwater acoustic sensors and the detection of fish school spatiotemporal features, the collected data is regularized into a data format that is easy to train a machine learning model. Preferably, this is a matrix of data points × time × feature dimensions. Fish populations mapped to disease severity levels The data format and mathematical formula are as follows: ; Where i represents the location of the i-th data point; j is the normalized time interval; A preferred example of data point arrangement is a clockwise arrangement from the water surface to the bottom, such as... Figure 2 As shown. A preferred example in this paper is to collect data every 15 minutes for a duration of 10 seconds, with a data frame rate of 10Hz.

[0016] k represents the data dimension, corresponding to the density, distribution, transformation period, and activity probability distribution characteristics. The value range of X is [0, 65535], corresponding to the signal strength; The value range of Y is [0,1,2,3,4,5], corresponding to the disease severity level; Mapping Machine learning models include Support Vector Machine (SVM), Linear Algorithm (LDA), Convolutional Neural Network (CNN), and Bayesian Model (BS). During the training of the model (mapping), the data collected in the early stage is divided into training set, validation set and test set (in a random ratio of 60%:20%:20%). The specific method for organizing the detection data into a matrix is ​​as follows: information on fish schools on and under the water surface is obtained through data collection; the data on the water surface is image information in the form of planar data; while the data under the water is the intensity data curve of the vibration sensor; the data form is a sequence of known data points. The matrixing of data points is achieved using spatial alignment; that is, the planar image data of the water surface is sliced ​​along the centerline of the underwater sensor location to ensure that all data have a consistent resolution, as expressed by the formula: Where I represents the image, M and N represent the image size, i represents the matrix-aligned sub-image, and ix and iy represent the underwater sensor arrangement. After the two forms of data are aligned in the geometric dimension of the data points through the result processing, the i-th dimension of the data point dimension of the above X_(i,j,k) matrix is ​​unified. The temporal resolution of image data acquisition is the video frame rate, while the data acquisition frame rate of the underwater vibration sensor is different. These two types of data have temporal differences and need to be aligned in the time dimension. The alignment process is as follows: by segmenting the vibration data points in time, the vibration curve data is aligned with the lower frame rate in terms of temporal resolution; the formula is expressed as follows. Where S represents vibration data and T represents the set of data points over time; This indicates an aligned segment, and j represents the j-th dimension of the time dimension of the X_(i,j,k) matrix; The feature dimensions of k correspond to the density, distribution, transformation period, and activity probability distribution features mentioned above. This part uses a method for extracting features of fish swarms in water.

[0017] The above processing forms a matrix of data points × time × feature dimensions, which facilitates the construction and use of subsequent disease detection models.

[0018] The disease detection method targets a single body of water containing only one group of fish, which is a mixed group of healthy and diseased fish. The value range of Y is [0,1,2,3,4,5], corresponding to disease levels 0 to 5, where 0 represents health; a value other than 0 indicates a severity of disease ranging from 1 to 5.

[0019] The present invention has the following advantages: 1. This invention utilizes a method combining surface and underwater data sources, comprehensively applying underwater visual images with data from multiple sensors such as acoustic and water pressure sensors to provide a comprehensive understanding of the fish's condition. The visual images acquired by the underwater camera directly present the fish's surface characteristics (such as the presence of white spots, ulcers, or other signs of disease) and swimming patterns (normal schooling, scattered swimming, or abnormally agitated or lethargic swimming). The acoustic sensor simultaneously monitors the sound of the fish swimming; healthy fish swim with a stable rhythm and relatively stable sound frequency and intensity, while sick fish may emit noticeably different sounds due to weakness or difficulty breathing. This invention, by combining these two methods, can more accurately determine the health of the fish. The water pressure sensor used in this invention can assist in measuring the depth distribution of the fish in the water, further enriching the information dimensions and overcoming the limitations of single data modalities.

[0020] 2. This invention employs a dynamic spatiotemporal modeling method. By establishing a model of the fish population's distribution and evolution over time, it analyzes the current aggregation and dispersion states of the fish population in the three-dimensional space of the aquaculture area, as well as their changing patterns at different times (e.g., day and night, different growth stages), further establishing a dynamic spatiotemporal model of the fish population. Through this dynamic model, abnormal changes in fish density and the spatiotemporal dynamic patterns of fish activity can be identified to detect diseases and identify the risk of disease spread in advance. Attached Figure Description

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Appendix Figure 1 This is a flowchart illustrating the present invention; Appendix Figure 2 This is a schematic diagram illustrating an example of data point arrangement in an embodiment of the present invention; Appendix Figure 3 This is an example of an embodiment of the invention where data is analyzed using an established SVM model to obtain a schematic diagram of the activity levels of diseased and healthy fish in a calm state (duration 15 minutes). Detailed Implementation

[0022] like Figure 1 As shown, a disease detection method integrating surface and underwater information on the dynamic spatiotemporal distribution characteristics of fish schools is proposed. This method combines visual image data from surface data sources and acoustic data from underwater data sources to determine the health of the fish school. It also measures the depth distribution of the fish school in the water using a water pressure sensor. A dynamic spatiotemporal model is established to create a model of the fish school's distribution evolution over time. The method collects data on the fish school's aggregation and dispersion states in the three-dimensional space of the aquaculture area, as well as their variation patterns at different times (such as day and night, and different growth stages), to obtain a dynamic spatiotemporal model of the fish school. This model is then used to identify abnormal changes in fish school density and the spatiotemporal dynamic patterns of their activities, thereby detecting diseases and recognizing the risk of disease spread in advance.

[0023] The water surface data source uses an image data acquisition device positioned above the water surface. This device includes a camera with high-definition resolution (4K) and telephoto zoom (20x). The camera is mounted on a fixed bracket above the fish farming area, preferably at an adjustable height of 3-5 meters. The height should ensure sufficient coverage of the water surface and clear observation of the fish's movements in the upper layers. The camera must be waterproof, dustproof, sunproof, and have automatic image stabilization to adapt to complex outdoor environments. It is also equipped with an adjustable gimbal to allow for easy adjustment of the shooting angle, capturing images of the fish's activities near the water surface from all angles.

[0024] The data acquisition process is pre-set with dynamic variations in acquisition frequency and duration. The preferred video recording parameters are set to record once every 15 minutes for a duration of 10 seconds. During the acquisition process, supplementary lighting, such as four 50W LED lights, can be turned on to ensure data acquisition during low-light periods such as early morning, evening, or cloudy days.

[0025] The underwater data source utilizes an underwater acoustic data acquisition array, which comprises multiple underwater acoustic sensors arranged at fixed intervals (uniformly distributed at 5-10 meters, with vertical spacing of 3-5 meters, adjustable according to area size, but horizontal spacing should be less than 20 meters and vertical spacing less than 10 meters). These sensors are fixed to specially designed underwater supports to ensure stability and prevent displacement or damage from water flow. The sampling frequency range is set between 20kHz and 50kHz to accurately capture the sound signals generated by fish swimming, feeding, and breathing. By adjusting the gain parameters, the sensors adapt to different intensities of sound input, enabling the detection of weak, normal swimming sounds, as well as abnormal sounds from fish under stress or illness, ensuring the acquired sound data has rich information content.

[0026] The underwater data acquisition from the underwater data source is synchronized with the data acquisition from the surface data source.

[0027] The distribution evolution model includes data on the distribution evolution of fish schools over time, which is obtained through data preprocessing of visual image data from surface data sources and acoustic data from underwater data sources. Data preprocessing includes surface image denoising, underwater sound denoising, and data fusion. The specific method for denoising water images is as follows: Due to the complex outdoor environment, water images are easily affected by factors such as light reflection and water surface ripples, resulting in salt-and-pepper noise and Gaussian noise. For salt-and-pepper noise in water images, a median filtering algorithm (the verified filtering operator is 5×5) is used. This algorithm selects the median value within the neighborhood of a pixel to replace that pixel value, effectively removing isolated noise points. When dealing with blurring caused by subtle ripples on the water surface, Gaussian filtering is used through convolution kernel operations to smooth the image based on a Gaussian function to reduce the impact of noise (Gaussian filter operator size is 3×3 or 5×5). The underwater sound noise reduction method is as follows: The main background noise in aquaculture areas consists of water flow noise and equipment operation noise (such as aerators and water pumps), which interferes with the acquisition of fish sound signals. The method involves first acquiring the sound information of water flow and equipment, and then obtaining the frequency domain characteristics of the water flow and equipment through frequency domain transformation. After frequency domain transformation of the acquired data, the sound signal is decomposed into sub-bands of different frequencies. Low-frequency water flow noise and high-frequency equipment noise are removed to extract the pure fish sound signal. Fourier frequency domain transformation and Laplace transform are used for frequency domain feature transformation. The specific data fusion method is as follows: the surface images and underwater fixed-point sound data are fused at a unified time; when there are differences in the frequency and start time of image acquisition and sound acquisition, time synchronization algorithms using interpolation and extrapolation are used to make the two accurately aligned in the time dimension to ensure the continuity of the data.

[0028] The detection method includes the use of acoustic signature sensor matrix data inside the water body; The modeling process of the dynamic spatiotemporal model includes the extraction of the dynamic spatiotemporal characteristics of the fish school. The extracted spatiotemporal features include individual features, surface group features, and water body fish school features. The specific method for extracting individual features is as follows: In the preprocessed images of the data acquired from the water surface, target detection and recognition algorithms are used to identify the bounding box of each fish, and then the morphological features of the fish's body length, body width, and body shape coefficient are extracted; through continuous analysis of multiple frames of images using a tracking algorithm, the instantaneous swimming speed, acceleration, and turning angle of the fish are obtained; combined with the acoustic data of the uppermost layer near the water surface, the physiological characteristics of the fish's movement frequency and feeding activity are determined. The specific method for extracting the characteristics of the upper surface population is as follows: based on individual characteristic statistics, the distribution density of the activity level of the fish population is calculated, that is, the distribution of fish of different activity levels (preferably 5 levels: inactive, generally inactive, moderate, generally active, and active) within a unit water volume; cluster analysis and density peak algorithm are used to identify the aggregation center, subgroup division, and overall shape and size changes of the population; through vector analysis of the swimming direction of the fish population, the migration trend, diffusion or contraction rate of the population is determined to extract the dynamic evolution characteristics of the fish population in the upper water surface space. The specific method for extracting the characteristics of fish schools in water is as follows: Based on the characteristics of the fish population on the upper surface, the data collected by each sensor is compared and integrated by combining the data of the acoustic sensor matrix inside the water body. The triangulation method and hyperbolic positioning algorithm are used to establish the relationship between acoustic intensity, acoustic density, time interval and distance to the fish population, number of fish population and activity level. Based on the sound frequency change caused by the Doppler effect, the swimming speed of the fish school is calculated. The magnitude of the sound signal and the distance to the fish school are established through the inverse proportional relationship model. By combining the characteristics of fish schools on the surface and in the water body, the density, distribution, change cycle and activity probability distribution characteristics of the overall fish school can be obtained.

[0029] The disease detection method is carried out through a dynamic spatiotemporal model, assuming that there are differences in the average amount of movement, cumulative amount of movement, instantaneous amount of movement, movement range and movement area between healthy fish groups and diseased fish groups inside and on the surface of the water.

[0030] The data and detection inside the water body are performed using acoustic sensors (vibration signals). However, due to the turbidity of the water inside the aquaculture body, the video acquisition range is limited to the upper surface layer of the water body (the activity of fish in the range of about 0~30CM depth).

[0031] The training method for the dynamic spatiotemporal model is as follows: Based on the collaborative data collection from surface optical and underwater acoustic sensors and the detection of fish school spatiotemporal features, the collected data is regularized into a data format that is easy to train a machine learning model. Preferably, this is a matrix of data points × time × feature dimensions. Fish populations mapped to disease severity levels The data format and mathematical formula are as follows: ; Where i represents the location of the i-th data point; j is the normalized time interval; A preferred example of data point arrangement is a clockwise arrangement from the water surface to the bottom, such as... Figure 2 As shown. A preferred example in this paper is to collect data every 15 minutes for a duration of 10 seconds, with a data frame rate of 10Hz.

[0032] k represents the data dimension, corresponding to the density, distribution, transformation period, and activity probability distribution characteristics. The value range of X is [0, 65535], corresponding to the signal strength; The value range of Y is [0,1,2,3,4,5], corresponding to the disease severity level; Mapping Machine learning models include Support Vector Machine (SVM), Linear Algorithm (LDA), Convolutional Neural Network (CNN), and Bayesian Model (BS). During the training of the model (mapping), the data collected in the early stage is divided into training set, validation set and test set (in a random ratio of 60%:20%:20%). The specific method for organizing the detection data into a matrix is ​​as follows: information on fish schools on and under the water surface is obtained through data collection; the data on the water surface is image information in the form of planar data; while the data under the water is the intensity data curve of the vibration sensor; the data form is a sequence of known data points. The matrixing of data points is achieved using spatial alignment; that is, the planar image data of the water surface is sliced ​​along the centerline of the underwater sensor location to ensure that all data have a consistent resolution, as expressed by the formula: Where I represents the image, M and N represent the image size, i represents the matrix-aligned sub-image, and ix and iy represent the underwater sensor arrangement. After the two forms of data are aligned in the geometric dimension of the data points through the result processing, the i-th dimension of the data point dimension of the above X_(i,j,k) matrix is ​​unified. The temporal resolution of image data acquisition is the video frame rate, while the data acquisition frame rate of the underwater vibration sensor is different. These two types of data have temporal differences and need to be aligned in the time dimension. The alignment process is as follows: by segmenting the vibration data points in time, the vibration curve data is aligned with the lower frame rate in terms of temporal resolution; the formula is expressed as follows. Where S represents vibration data and T represents the set of data points over time; This indicates an aligned segment, and j represents the j-th dimension of the time dimension of the X_(i,j,k) matrix; The feature dimensions of k correspond to the density, distribution, transformation period, and activity probability distribution features mentioned above. This part uses a method for extracting features of fish swarms in water.

[0033] The above processing forms a matrix of data points × time × feature dimensions, which facilitates the construction and use of subsequent disease detection models.

[0034] The disease detection method targets a single body of water containing only one group of fish, which is a mixed group of healthy and diseased fish. The value range of Y is [0,1,2,3,4,5], corresponding to disease levels 0 to 5, where 0 represents health; a value other than 0 indicates a severity of disease ranging from 1 to 5.

[0035] Example: like Figure 3 As shown in the figure, the surface layer data was analyzed using the established SVM model to obtain the activity levels of diseased and healthy individuals in a calm state (duration 15 minutes), presented statistically. It can be seen that although individuals in the population exhibit some similarity—for example, the activity level of a single individual is around 0.2, making it impossible to determine the individual's health status—the overall population distribution shows significant differences. As shown in the figure, the activity levels of diseased individuals in the calm state are concentrated below 0.2, while those of healthy individuals are significantly higher than 0.2. Similarly, statistical differences in population activity are also observed in the lower water layers.

Claims

1. A disease detection method integrating surface and underwater information on the dynamic spatiotemporal distribution characteristics of fish schools, characterized by: The detection method combines visual image data of a water surface data source and acoustic data of an underwater data source to judge the health degree of the fish group, measures the depth distribution of the fish group in water through a water pressure sensor, establishes a distribution evolution model of the fish group over time through dynamic space-time modeling, collects the gathering and scattering state of the fish group and the change law thereof at different time periods in view of the position relationship of the fish group in the three-dimensional space of the aquaculture water area, to obtain a dynamic space-time model of the fish group, and then identifies the abnormal change in density and the space-time dynamic law of the activity of the fish group through the dynamic space-time model, to detect the disease.

2. The disease detection method based on the fusion of surface and underwater information and the dynamic spatiotemporal distribution characteristics of fish schools according to claim 1, characterized in that: The water surface data source uses an image data acquisition device arranged above the water surface, and the image data acquisition device includes a camera with high-definition resolution and long-focus zooming function, which is installed on a fixed support above the fish group aquaculture area, and the installation height is determined according to the standard of clearly observing the dynamic of the upper fish group.

3. The disease detection method based on the fusion of surface and underwater information and the dynamic spatiotemporal distribution characteristics of fish schools according to claim 2, characterized in that: The data acquisition process is pre-set to have dynamic changes in the acquisition frequency and the acquisition duration.

4. The disease detection method of claim 1, wherein the method comprises: 1) collecting the water surface and underwater information of the fish; 2) determining the dynamic spatiotemporal distribution characteristics of the fish; and 3) detecting the disease of the fish based on the dynamic spatiotemporal distribution characteristics of the fish. The underwater data source uses an underwater sound data acquisition array, which includes a plurality of underwater acoustic sensors arranged at fixed intervals, and the underwater acoustic sensors are fixed at underwater supports. The sampling frequency range is set to be between 20 kHz and 50 kHz, so as to capture the sound signals generated by the swimming, feeding and breathing behaviors of the fish group. By adjusting the gain parameter, the sensor can adapt to different intensity of sound input, and can detect the normal swimming sound of the fish group and the abnormal sound of the fish group in the stress or disease state.

5. The disease detection method of claim 4, wherein the method comprises the steps of: 1) collecting the water surface and underwater information of the fish population; 2) determining the dynamic spatio-temporal distribution characteristics of the fish population; 3) determining the disease detection index of the fish population; and 4) determining the disease detection result of the fish population. The underwater data acquisition of the underwater data source is synchronized with the data acquisition of the water surface data source.

6. The disease detection method of claim 1, wherein the method comprises: The distribution evolution model includes distribution evolution data of the fish group over time, which is obtained through data preprocessing of the visual image data of the water surface data source and the acoustic data of the underwater data source. The data preprocessing includes water image denoising, underwater sound denoising and data fusion. The method of water image denoising is specifically: for the salt and pepper noise of the water image, a median filter algorithm is used to replace the pixel value by selecting the median value in the neighborhood of the pixel point, so as to effectively remove the isolated noise points. When processing the blur caused by the fine ripples on the water surface, a Gaussian filter is used to perform convolution kernel operation, and a Gaussian function is used to smooth the image to reduce the noise influence. The method of underwater sound denoising is specifically: the sound information of the water flow and the equipment is obtained in advance, and the frequency domain characteristics of the water flow and the equipment are obtained through frequency domain conversion. After frequency domain conversion of the obtained data, the sound signal is decomposed into subbands of different frequencies, the low-frequency water flow noise and the high-frequency equipment noise are removed, and the pure fish group sound signal is extracted. Fourier frequency domain conversion and Laplace conversion are used for frequency domain characteristic conversion. The method of data fusion is specifically: the water image and the underwater fixed-point sound data are fused according to the unified time. When there is a difference in the frequency and the starting time of the image acquisition and the sound acquisition, the time synchronization algorithm of the interpolation method and the extrapolation method is used to accurately align the two in the time dimension, so as to ensure the continuity of the data.

7. The disease detection method of claim 6, wherein the method comprises the steps of: a) obtaining a plurality of images of the water body; b) extracting the features of the images; c) determining the dynamic spatio-temporal distribution of the fish population; and d) detecting the disease based on the dynamic spatio-temporal distribution of the fish population. The detection method includes the use of internal water body acoustic fingerprint sensor matrix data. The modeling process of the dynamic space-time model includes extraction of dynamic space-time characteristics of the fish population, and the extracted space-time characteristics include individual characteristics, upper surface population characteristics, and water body fish population characteristics, The extraction method of the individual characteristics specifically includes: in the image obtained after preprocessing the data on the water surface, the boundary box of each fish is recognized by using a target detection and recognition algorithm, and then the morphological characteristics of the fish, such as body length, body width and body shape coefficient, are extracted; the motion characteristics of the fish, such as instantaneous swimming speed, acceleration and turning angle, are obtained by continuously analyzing multiple images through a tracking algorithm; and the physiological characteristics of the fish, such as motion frequency and feeding activity, are determined in combination with the uppermost acoustic data near the water surface; The extraction method of the upper surface population characteristics specifically includes: based on the individual characteristics statistics, the distribution density of the activity level of the fish population is calculated, that is, the number distribution of the fish of different activity levels in a unit water volume; the clustering analysis and the density peak algorithm are used to identify the aggregation center in the fish population, the sub-population division, and the overall shape and size change of the population; and the migration trend, diffusion or contraction rate of the population are determined through vector analysis of the swimming direction of the fish population, so as to extract the dynamic evolution characteristics of the fish population in the upper space of the water surface; The extraction method of the water body fish population characteristics specifically includes: on the basis of the upper surface population characteristics, the data collected by each sensor are compared and integrated in combination with the acoustic sensor matrix data in the water body, the triangular positioning method and the hyperbolic positioning algorithm are used, the relationship between the acoustic intensity, acoustic density and time interval and the sub-fish population distance, the sub-fish population quantity and the sub-fish population activity model is established; and the fish swimming speed is converted according to the sound frequency change caused by the Doppler effect, and the distance model of the sound signal size and the fish population is established through the inverse proportional relationship model.

8. The disease detection method of claim 7, wherein the method comprises the steps of: obtaining the dynamic spatio-temporal distribution characteristics of the fish school from the fusion of the water surface information and the underwater information; and detecting the disease of the fish school according to the dynamic spatio-temporal distribution characteristics of the fish school. The disease detection method is performed through the dynamic space-time model, and the average movement amount, cumulative movement amount, instantaneous movement amount, movement range and movement area of the healthy fish population and the diseased fish population in the water body and on the water surface are different.

9. The training method of the dynamic space-time model is specifically: on the basis of the cooperation of the early water surface optical sensor and underwater acoustic sensor, and the detection information of the space-time characteristics of the fish school, the collected data is regularized into a matrix with the dimensions of data points x time x characteristic, and is mapped into the data form of the fish school with the disease level, and the mathematical formula is as follows: ;​​ wherein i represents the position of the ith data point; j is the regularized time interval; k is the data dimension, corresponding to the density, distribution, transformation period and activity probability distribution characteristics; The value range of X is [0, 65535], corresponding to the signal intensity; The value range of Y is [0, 1, 2, 3, 4, 5], corresponding to the disease level; Mapping For machine learning models, including support vector machines SVM, linear models LDA, convolutional neural networks CNN, Bayesian models BS; In the process of training the model, the data collected in the early stage is divided into a training set, a validation set and a test set; The disease detection method for fusing fish population dynamic space-time distribution characteristics of water surface and underwater information according to claim 8, wherein the specific method for regularizing the detection data into a matrix is: obtaining the information of the fish population on the water surface and underwater through data collection; the data on the water surface is image information in the form of planar data; and the data underwater is the intensity data curve of the vibration sensor; the data form is a sequence of known data points; The matrixing in the data point dimension is implemented using a space alignment method; that is, the planar image data of the water surface is cut according to the center line of the sensor position under the water, so that all data resolutions are consistent, and is expressed by a formula as follows: wherein I represents an image, M and N represent the size of the image, i represents a sub-image aligned in the matrix, ix and iy represent the arrangement of the underwater sensors, and the two forms of data after result processing are aligned in the geometric dimension of the data points, so that the i dimension of the data point dimension of the above X_(i,j,k) matrix is unified; The time resolution of the image data is the frame rate of the video, and the data acquisition frame rate of the underwater vibration sensor, and the two kinds of data are different in time, and need to be aligned in the time dimension. The method of the alignment process is: by cutting the vibration data points in time, the vibration curve data is aligned in the time resolution with the lower frame rate; the formula is: ​ Wherein, S represents the vibration data, T represents the data point set on time; represents the aligned split segment, and j represents the j dimension of the time dimension of the X_(i,j,k) matrix. The characteristic dimension of k corresponds to the density, distribution, transformation period, and activity probability distribution characteristics described above.

10. The disease detection method of claim 8, wherein the method comprises: The disease detection method detects a fish group in a single water body, and the water body has only one fish group, which is a diseased fish group formed by mixing healthy fish groups and diseased fish groups. The value range of Y is [0, 1, 2, 3, 4, 5], corresponding to disease levels 0 to 5, and 0 is healthy; values other than 0 are determined to be severity 1 to 5.

Images