A method for measuring the number of fish cultured in deep-sea cages

Abstract

The present invention discloses a method for measuring the number of fish schools in deep-sea cage culture, comprising: using an omnidirectional sonar to perform an overall scan of a target cage to convert it into a single cage scanning video; training to obtain a fish school recognition model; obtaining the spatial three-dimensional coordinates of each identified fish in each frame of the sonar scanning image of the target single cage scanning video through the overall scan and the fish school recognition model; setting multiple water layers; performing density clustering processing on the spatial three-dimensional coordinates of each water layer to obtain the water layer clustering estimated fish quantity of each water layer; obtaining a sufficient number of neural network training data; training to obtain a cage total fish quantity fitting model; at the target time, through steps three to five and the cage total fish quantity fitting model, outputting the cage total fish quantity measurement value of the target cage. The present invention can use the omnidirectional sonar to achieve the measurement of the cage total fish quantity in the target cage in a relatively simple and fast data acquisition method, with a high measurement accuracy.

Description

Technical Field

[0001] The invention relates to fishery resource estimation and measurement, in particular to a method for measuring the number of fish schools cultured in deep-sea cages. Background Art

[0002] Due to the increasing depletion of marine fishery resources and the deterioration of shallow water aquaculture environment, deep-water cage aquaculture has developed rapidly in recent years. With the development of fishery science, modern fisheries have gradually formed a "from science to management" model, that is, from the past non-systematic empirical fishery decision-making to fishery management based on the best scientific evidence, which has also made the role of fishery resource assessment in fishery management increasingly apparent. The credibility of resource assessment results depends partly on model assumptions and partly on the quality of input data. It can be seen that fishery data is an important basis for scientific fishery management. On this basis, various fish monitoring methods such as optics and acoustics have been developed. However, in the monitoring process, there are still major problems in how to conduct statistical analysis on the size of the fish population in the cage. As an important part of cage aquaculture, fish quantity monitoring has far-reaching significance in the following aspects: first, the economic benefits of stocking different numbers of fry are different; second, accurately knowing the number of aquaculture is particularly important for obtaining accurate catches as data for resource assessment; third, knowing the amount of fish clearly is also of great help in making accurate fishery risk assessments.

[0003] In response to the above requirements, although some different fish population monitoring methods have been proposed, their estimation effects on fish population are not ideal and the cost is high. In addition, the monitoring requirements for breeding conditions are high, making it difficult to implement sonar monitoring and fish population assessment conveniently and effectively. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for measuring the number of fish schools cultured in deep-sea cages.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for measuring the number of fish in deep-sea cage culture, characterized by comprising:

[0007] Step 1: Place the omnidirectional sonar in the target cage, and drive the omnidirectional sonar to move at a constant speed between the water surface and the bottom of the cage in the vertical direction, so as to perform an overall scan of the target cage through the omnidirectional sonar during the movement, wherein the omnidirectional sonar from the bottom of the cage to the water surface, or from the water surface to the bottom of the cage can be regarded as an overall scan;

[0008] In addition, the data output by the omnidirectional sonar after an overall scan is played back through the supporting software of the omnidirectional sonar to be converted into a single cage scanning video, and the video frame of the single cage scanning video is recorded as the sonar scanning image, see Figure 3 ;

[0009] Step 2: Perform multiple overall scans of the target cage according to step 1 to obtain a sufficient number of single cage scanning videos as training data, so as to obtain a fish school recognition model through deep learning model training. The fish school recognition model can identify the position of each identified fish in each frame of the sonar scanning image of the single cage scanning video, see Figure 4 ;

[0010] Step 3: perform an overall scan of the target cage according to step 1 to obtain a target single cage scanning video, and use a fish school recognition model to identify the position of each identified fish in each frame of the sonar scanning image of the target single cage scanning video;

[0011] Furthermore, the spatial three-dimensional coordinates (x, y, d) of each identified fish in each sonar scanning image frame of the target single cage scanning video are extracted and converted from the recognition result of the fish school recognition model, that is, from the pixel coordinates and water depth obtained by recognition, wherein x and y represent the horizontal plane abscissa and horizontal plane ordinate of the identified fish at the time of recognition (i.e., the time when the omnidirectional sonar scanning obtains the corresponding sonar scanning image), respectively, and d represents the water depth of the identified fish at the time of recognition;

[0012] The pixel coordinates in the sonar scanning image can be converted into x and y in the three-dimensional spatial coordinates by using the conversion method in the prior art.

[0013] Step 4, starting from the water surface (i.e. d=0), multiple water layers are set according to the preset water depth interval △d, and two adjacent water layers have a preset water depth overlap interval δd, that is: dmax-dmin=△d, d1max-d2min=δd, dmax and dmin respectively represent the maximum water depth and the minimum water depth of any water layer, d1max and d2min respectively represent the maximum water depth of the upper water layer and the minimum water depth of the next water layer in two adjacent water layers; because it is taken into account that during measurement, fish may run to different water layers and generate similar signal data in different water layers, so ensuring that there is an overlap between two adjacent water layers can minimize the influence of the coordinates of the center point fish after clustering; and because the subsequent step is clustering, even if the counting is repeated, it will not have much impact on the clustering result.

[0014] Furthermore, according to the water depth d, the spatial three-dimensional coordinates (x, y, d) of each identified fish obtained in step 3 are respectively attributed to the corresponding water layer, wherein the spatial three-dimensional coordinates (x, y, d) within the overlapping interval δd of the water depth are attributed to both the upper and lower water layers;

[0015] Step 5: For each of the water layers, density clustering is performed on the spatial three-dimensional coordinates (x, y, d) belonging to the water layer to obtain a cluster of each water layer; wherein each cluster can be identified as a fish within the corresponding water layer range;

[0016] Furthermore, the number of clusters in each of the water layers is counted, and recorded as the estimated fish quantity of the water layer clusters;

[0017] For example: Figure 5 It is the spatial distribution map of 7681 three-dimensional coordinates (x, y, d) contained in a certain water layer. Figure 6 for Figure 5 The spatial distribution diagram of the 17 clusters obtained after the water layer is processed in step five is shown, wherein the 17 clusters, namely Cu0 to Cu16, are marked with different colors respectively.

[0018] Step 6: Under different test conditions, perform steps 3 to 5 respectively to obtain a sufficient amount of neural network training data;

[0019] The test condition includes at least one of the following conditions:

[0020] Condition 1: different numbers of fish are placed in the target cages, that is, the target cages contain different actual fish amounts placed in the cages;

[0021] Condition 2: different clustering radii are set in the density clustering process of step 5; for example, a clustering radius may be set every 0.1N within the range of 0.5N to 0.8N.

[0022] Step 7: Using the neural network training data obtained in step 6, the water layer clustering estimated fish quantity of multiple water layers calculated through steps 3 to 5 is used as input, and the actual amount of fish released into the cage is used as output, to train the neural network to obtain a fitting model for the total amount of fish in the cage; thereby, the nonlinear mapping problem between the water layer clustering estimated fish quantity of each water layer and the total amount of fish in the target cage as a whole can be solved.

[0023] Step 8. At the target time when the target cage needs to be measured for fish quantity, first calculate the estimated fish quantity by water layer clustering of multiple water layers according to steps 3 to 5, and then input the calculated estimated fish quantity by water layer clustering of multiple water layers into the cage total fish quantity fitting model to output the measured value of the total fish quantity in the target cage.

[0024] The following Table 1 is a test data set for testing the cage fish quantity measurement effect of the present invention, wherein the accuracy rate can reach 87.1% on average, and the precision is relatively high.

[0025] Preferably, in step seven, the neural network is a BP neural network (Back Propagation), and during the training process, preferably, the number of neurons in the hidden layer is 20, and Bayesian regularization is used for training.

[0026] Preferably: in the step one, the omnidirectional sonar is placed at the central axis position of the target cage and is driven by a vertical lifting drive device to move at a uniform speed between the water surface and the bottom of the cage in the vertical direction; wherein the vertical lifting drive device can be a mechanical lifting device driven by a motor, or a rope driven by manpower, and a scale can be set on the rope to facilitate depth control and recording.

[0027] Among them, the omnidirectional sonar can be a sonar that can simultaneously achieve 360° omnidirectional scanning, or a sonar that achieves 360° omnidirectional scanning by driving the scanning beam to rotate in the horizontal plane. When using the latter, the vertical movement speed of the omnidirectional sonar cannot be too fast to avoid failure to fully scan the entire space of the target cage.

[0028] Wherein, the beam opening angle range of the scanning beam of the omnidirectional sonar is [a°, b°];

[0029] In step 4, the water depth interval Δd is calculated according to the following formula:

[0030] △d=(△d1+△d2) / 2;

[0031] △d1=2*tan(a / 2)*R;

[0032] △d2=2*tan(b / 2)*R;

[0033] Where △d1 and △d2 are the water depth intervals when the beam opening angles are a° and b° respectively; R is the radius of the target cage.

[0034] Thus, the overall scanning of the target cage by the omnidirectional sonar can be realized more quickly and effectively.

[0035] Among them, in the step 4, the water depth overlap interval δd is taken as the smaller value between the average fish length and the average fish width of the identified fish.

[0036] Thus, it is possible to avoid fish wandering within the overlapping interval δd of water depth between two adjacent water layers, which would lead to repeated counting of fish in the two adjacent water layers.

[0037] In step 5, the DBSCAN clustering analysis method is used to perform density clustering processing on the spatial three-dimensional coordinates (x, y, d) belonging to the same water layer; wherein the clustering parameters are set in the following manner:

[0038] The clustering radius (epsilon) takes values ​​in the range of 0.5N to 0.8N, where N is the larger value of the average fish length and the average fish width of the identified fish; when the clustering radius is larger within this range, the number of clusters obtained by clustering will decrease, but considering the movement of fish in the water and the position changes between fish, the error of the number after clustering is small; the increase in the clustering radius will cause the offset of the center point of the fish, but it will have little effect on the spatial distribution characteristics of the overall fish in the cage.

[0039] The minimum number of samples (min_samples) is set to 1, which means that the cluster obtained by clustering can contain at least one sample, that is, the three-dimensional spatial coordinates (x, y, d).

[0040] Preferably, in step 2, the deep learning model adopts the YOLOv8 model, and the fish school recognition model is trained in the following manner:

[0041] Step 2-1: capture a sufficient number of sonar scanning images from a single cage scanning video, and annotate the position of each identified fish in each frame of the sonar scanning image using the MakeSense online data annotation software to output an annotation result file;

[0042] Step 2 of 2: Use a single cage scanning video as input data, use the annotation result file corresponding to the single cage scanning video as output data, and train the YOLOv8 model to obtain the fish school recognition model.

[0043] Preferably, the method for measuring the number of fish schools cultured in deep-sea cages further comprises: according to the water layer clustering estimated fish quantity of each water layer calculated in step 5, drawing a fish quantity change trend diagram of the water layer clustering estimated fish quantity in the target cage as the water depth changes, see Figure 7 , which can help farmers better understand the distribution of fish in the target cages.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] First, through steps one to eight, the present invention can use the omnidirectional sonar 1 to measure the total amount of fish in the target cage 2 in a relatively simple and fast data collection method, and after experimental verification, the accuracy of the measured value of the total amount of fish in the cage can reach an average of 87.1%, so it has a high measurement accuracy.

[0046] Second, since the present invention only needs to use the omnidirectional sonar 1 to obtain all the data required for the implementation method according to step one, the requirements for scanning the omnidirectional sonar 1 are relatively low. It only needs to scan the complete target cage 2 in the vertical direction at a uniform speed. It has the advantages of being simple, convenient, efficient and fast, and can solve the problems of high requirements for data monitoring and collection of the target cage 2 in the prior art, expensive monitoring equipment, complex operation and high cost.

[0047] Third, since the present invention does not have strict requirements on the quality of the data obtained by the omnidirectional sonar 1 scanning, even if the fish in the target cage 2 are scanned repeatedly, the density clustering processing in step five can be used to reduce the impact of repeated scanning to obtain a higher measurement accuracy of the total fish volume in the cage, thereby solving the problem that the estimation effect of the prior art is not ideal. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0049] Figure 1 It is a flowchart of the method for measuring the number of fish schools cultured in deep sea cages of the present invention;

[0050] Figure 2 It is a schematic diagram of the omnidirectional sonar of the present invention being placed in a target cage;

[0051] Figure 3 is a schematic diagram of a sonar scanning image in step S1 of the present invention;

[0052] Figure 4 A schematic diagram of the fish school recognition model identifying the position of the identified fish in the sonar scanning image in step S2 of the present invention;

[0053] Figure 5 It is the spatial distribution map of 7681 three-dimensional spatial coordinates (x, y, d) contained in a certain water layer;

[0054] Figure 6 for Figure 5 The spatial distribution diagram of the clusters obtained after the water layer is processed in step S5 is shown;

[0055] Figure 7 This is a trend chart of fish quantity changes in a target cage. DETAILED DESCRIPTION

[0056] The present invention is described in detail below in conjunction with the embodiments and the accompanying drawings to help those skilled in the art better understand the inventive concept of the present invention. However, the protection scope of the claims of the present invention is not limited to the following embodiments. For those skilled in the art, all other embodiments obtained without creative work without departing from the inventive concept of the present invention belong to the protection scope of the present invention.

[0057] like Figures 1 to 4 As shown, the present invention discloses a method for measuring the number of fish schools cultured in deep-sea cages, comprising:

[0058] Step 1: Place the omnidirectional sonar 1 in the target cage 2, and drive the omnidirectional sonar 1 to move at a constant speed between the water surface 2a and the cage bottom 2b in the vertical direction, so as to perform an overall scan of the target cage 2 through the omnidirectional sonar 1 during the movement, wherein the omnidirectional sonar 1 from the cage bottom 2b to the water surface 2a, or from the water surface 2a to the cage bottom 2b, can be regarded as an overall scan;

[0059] In addition, the data outputted by the omnidirectional sonar 1 after an overall scan is played back through the supporting software of the omnidirectional sonar 1 to be converted into a single cage scanning video, and the video frame of the single cage scanning video is recorded as a sonar scanning image, see Figure 3 ;

[0060] Preferably: in the step one, the omnidirectional sonar 1 is placed at the central axis position of the target cage 2 and is driven by a vertical lifting drive device to move at a uniform speed between the water surface 2a and the bottom surface 2b of the cage in the vertical direction; wherein the vertical lifting drive device can be a mechanical lifting device driven by a motor, or a rope driven by manpower, and a scale can be set on the rope to facilitate depth control and recording.

[0061] Among them, the omnidirectional sonar 1 can adopt a sonar that can simultaneously realize 360° omnidirectional scanning, or it can adopt a sonar that realizes 360° omnidirectional scanning by driving the scanning beam 1a to rotate in the horizontal plane. When using the latter, the vertical movement speed of the omnidirectional sonar 1 cannot be too fast to avoid failure to fully scan the entire space of the target cage 2.

[0062] Step 2: Perform multiple overall scans on the target cage 2 according to step 1 to obtain a sufficient number of single cage scanning videos as training data, so as to obtain a fish school recognition model through deep learning model training. The fish school recognition model can identify the position of each identified fish in each frame of the sonar scanning image of the single cage scanning video, see Figure 4 ;

[0063] Preferably, in step 2, the deep learning model adopts the YOLOv8 model, and the fish school recognition model is trained in the following manner:

[0064] Step 2-1: capture a sufficient number of sonar scanning images from a single cage scanning video, and annotate the position of each identified fish in each frame of the sonar scanning image using the MakeSense online data annotation software to output an annotation result file;

[0065] Step 2 of 2: Use a single cage scanning video as input data, use the annotation result file corresponding to the single cage scanning video as output data, and train the YOLOv8 model to obtain the fish school recognition model.

[0066] Step 3: perform an overall scan of the target cage 2 according to step 1 to obtain a target single cage scanning video, and use a fish school recognition model to identify the position of each identified fish in each frame of the sonar scanning image of the target single cage scanning video;

[0067] Furthermore, the spatial three-dimensional coordinates (x, y, d) of each identified fish in each sonar scanning image frame of the target single cage scanning video are extracted and converted from the recognition result of the fish school recognition model, that is, from the pixel coordinates and water depth obtained by recognition, wherein x and y respectively represent the horizontal plane abscissa and horizontal plane ordinate of the identified fish at the time of recognition (i.e., the time when the omnidirectional sonar 1 scans and obtains the corresponding sonar scanning image), and d represents the water depth of the identified fish at the time of recognition;

[0068] The pixel coordinates in the sonar scanning image can be converted into x and y in the three-dimensional spatial coordinates by using the conversion method in the prior art.

[0069] Step 4, starting from the water surface (i.e. d=0), multiple water layers are set according to the preset water depth interval △d, and two adjacent water layers have a preset water depth overlap interval δd, that is: dmax-dmin=△d, d1max-d2min=δd, dmax and dmin respectively represent the maximum water depth and the minimum water depth of any water layer, d1max and d2min respectively represent the maximum water depth of the upper water layer and the minimum water depth of the next water layer in two adjacent water layers; because it is taken into account that during measurement, fish may run to different water layers and generate similar signal data in different water layers, so ensuring that there is an overlap between two adjacent water layers can minimize the influence of the coordinates of the center point fish after clustering; and because the subsequent step is clustering, even if the counting is repeated, it will not have much impact on the clustering result.

[0070] Furthermore, according to the water depth d, the spatial three-dimensional coordinates (x, y, d) of each identified fish obtained in step 3 are respectively attributed to the corresponding water layer, wherein the spatial three-dimensional coordinates (x, y, d) within the overlapping interval δd of the water depth are attributed to both the upper and lower water layers;

[0071] Wherein, the beam opening angle range of the scanning beam 1a of the omnidirectional sonar 1 is [a°, b°];

[0072] In step 4, the water depth interval Δd is calculated according to the following formula:

[0073] △d=(△d1+△d2) / 2;

[0074] △d1=2*tan(a / 2)*R;

[0075] △d2=2*tan(b / 2)*R;

[0076] Wherein, △d1 and △d2 are the water depth intervals when the beam opening angles are a° and b° respectively; R is the radius of the target cage 2.

[0077] Thus, the overall scanning of the target cage 2 by the omnidirectional sonar 1 can be achieved more quickly and effectively.

[0078] Among them, in the step 4, the water depth overlap interval δd is taken as the smaller value between the average fish length and the average fish width of the identified fish.

[0079] Thus, it is possible to avoid fish wandering within the overlapping interval δd of water depth between two adjacent water layers, which would lead to repeated counting of fish in the two adjacent water layers.

[0080] Step 5: For each of the water layers, density clustering is performed on the spatial three-dimensional coordinates (x, y, d) belonging to the water layer to obtain a cluster of each water layer; wherein each cluster can be identified as a fish within the corresponding water layer range;

[0081] Furthermore, the number of clusters in each of the water layers is counted, and recorded as the estimated fish quantity of the water layer clusters;

[0082] For example: Figure 5 It is the spatial distribution map of 7681 three-dimensional coordinates (x, y, d) contained in a certain water layer. Figure 6 for Figure 5 The spatial distribution diagram of the 17 clusters obtained after the water layer is processed in step five is shown, wherein the 17 clusters, namely Cu0 to Cu16, are marked with different colors respectively.

[0083] In step 5, the DBSCAN clustering analysis method is used to perform density clustering processing on the spatial three-dimensional coordinates (x, y, d) belonging to the same water layer; wherein the clustering parameters are set in the following manner:

[0084] The clustering radius (epsilon) takes values ​​in the range of 0.5N to 0.8N, where N is the larger value of the average fish length and the average fish width of the identified fish; when the clustering radius is larger within this range, the number of clusters obtained by clustering will decrease, but considering the movement of fish in the water and the position changes between fish, the error of the number after clustering is small; the increase in the clustering radius will cause the offset of the center point of the fish, but it will have little effect on the spatial distribution characteristics of the overall fish in the cage.

[0085] The minimum number of samples (min_samples) is set to 1, which means that the cluster obtained by clustering can contain at least one sample, that is, the three-dimensional spatial coordinates (x, y, d).

[0086] Step 6: Under different test conditions, perform steps 3 to 5 respectively to obtain a sufficient amount of neural network training data;

[0087] The test condition includes at least one of the following conditions:

[0088] Condition 1: different numbers of fish are put into the target cage 2, that is, the target cage 2 contains different actual fish amounts put into the cage;

[0089] Condition 2: different clustering radii are set in the density clustering process of step 5; for example, a clustering radius may be set every 0.1N within the range of 0.5N to 0.8N.

[0090] Step 7: Use the neural network training data obtained in step 6, take the water layer clustering estimated fish quantity of multiple water layers calculated through steps 3 to 5 as input, and take the actual amount of fish released into the cage as output, to train the neural network to obtain a fitting model for the total amount of fish in the cage; thereby, the nonlinear mapping problem between the water layer clustering estimated fish quantity of each water layer and the total amount of fish in the target cage 2 as a whole can be solved.

[0091] Preferably, in step seven, the neural network is a BP neural network (Back Propagation), and during the training process, preferably, the number of neurons in the hidden layer is 20, and Bayesian regularization is used for training.

[0092] Step 8. At the target time when the fish quantity of the target cage 2 needs to be measured, first calculate the estimated fish quantity of the water layer clustering of multiple water layers according to steps 3 to 5, and then input the calculated estimated fish quantity of the water layer clustering of multiple water layers into the total fish quantity fitting model of the cage to output the total fish quantity measurement value of the target cage 2.

[0093] The following Table 1 is a test data set for testing the cage fish quantity measurement effect of the present invention, wherein the accuracy rate can reach 87.1% on average, and the precision is relatively high.

[0094] Table 1

[0095]

[0096] Preferably, the method for measuring the number of fish schools cultured in deep-sea cages further comprises: drawing a fish quantity variation trend diagram of the fish quantity estimated by water layer clustering in the target cage 2 as the water depth changes according to the water layer clustering estimated fish quantity of each water layer calculated in step 5, see Figure 7, so as to help farmers better understand the distribution of fish in the target cage 2.

[0097] The present invention is not limited to the above-mentioned specific implementation modes. According to the above-mentioned contents, in accordance with the common technical knowledge and customary means in the field, without departing from the above-mentioned basic technical ideas of the present invention, the present invention can also make other various forms of equivalent modifications, replacements or changes, all of which fall within the protection scope of the present invention.

Claims

1. A method for measuring the number of fish in deep-sea cage culture, characterized in that: include: Step 1: placing an omnidirectional sonar (1) in a target cage (2), and driving the omnidirectional sonar (1) to move vertically between the water surface (2a) and the cage bottom surface (2b) at a constant speed, so as to perform an overall scan of the target cage (2) through the omnidirectional sonar (1) during the movement; Furthermore, the data outputted by the omnidirectional sonar (1) after an overall scan is played back through the supporting software of the omnidirectional sonar (1) to be converted into a single cage scanning video, and the video frames of the single cage scanning video are recorded as sonar scanning images; Step 2: Performing multiple overall scans of the target cage (2) according to step 1 to obtain a sufficient number of single cage scanning videos as training data, so as to obtain a fish school recognition model through deep learning model training, wherein the fish school recognition model can recognize the position of each identified fish in each frame of the sonar scanning image of the single cage scanning video; Step 3, performing an overall scan of the target cage (2) according to step 1 to obtain a target single cage scanning video, and using a fish school recognition model to identify the position of each identified fish in each frame of the sonar scanning image of the target single cage scanning video; Furthermore, the spatial three-dimensional coordinates (x, y, d) of each identified fish in each frame of the sonar scanning image of the target single cage scanning video are extracted and converted from the recognition results of the fish school recognition model, wherein x and y represent the horizontal plane abscissa and horizontal plane ordinate of the identified fish at the time of recognition, respectively, and d represents the water depth of the identified fish at the time of recognition; Step 4: Starting from the water surface, multiple water layers are set according to a preset water depth interval △d, and two adjacent water layers have a preset water depth overlap interval δd, that is: dmax-dmin=△d, d1max-d2min=δd, dmax and dmin represent the maximum water depth and minimum water depth of any water layer, respectively, and d1max and d2min represent the maximum water depth of the upper water layer and the minimum water depth of the next water layer, respectively; Furthermore, according to the water depth d, the spatial three-dimensional coordinates (x, y, d) of each identified fish obtained in step 3 are respectively assigned to the corresponding water layer; Step 5: For each of the water layers, density clustering is performed on the spatial three-dimensional coordinates (x, y, d) belonging to the water layer to obtain a cluster of each water layer; Furthermore, the number of clusters in each of the water layers is counted, and recorded as the estimated fish quantity of the water layer clusters; Step 6: Under different test conditions, perform steps 3 to 5 respectively to obtain a sufficient amount of neural network training data; The test condition includes at least one of the following conditions: Condition 1: different numbers of fish are placed in the target cage (2), that is, the target cage (2) contains different actual fish amounts placed in the cage; Condition 2: setting different clustering radii in the density clustering process of step 5; Step 7: Using the neural network training data obtained in step 6, taking the estimated fish quantity of multiple water layers calculated in steps 3 to 5 as input and the actual fish quantity put into the cage as output, the neural network is trained to obtain a fitting model of the total fish quantity in the cage; Step 8. At the target time, first calculate the estimated fish quantity of the water layers by clustering in the multiple water layers according to steps 3 to 5, then input the calculated estimated fish quantity of the water layers by clustering in the multiple water layers into the total fish quantity fitting model of the cage to output the total fish quantity measurement value of the target cage (2).

2. The method for measuring the number of deep-sea cage cultured fish according to claim 1, characterized in that: In the step seven, the neural network is a BP neural network.

3. The method for measuring the number of deep-sea cage cultured fish according to claim 1, characterized in that: In step 1, the omnidirectional sonar (1) is placed at the central axis of the target cage (2) and is driven by a vertical lifting drive device to move at a constant speed between the water surface (2a) and the bottom surface (2b) of the cage in a vertical direction.

4. The method for measuring the number of deep-sea cage cultured fish according to claim 3, characterized in that: The beam opening angle range of the scanning beam (1a) of the omnidirectional sonar (1) is [a°, b°]; In step 4, the water depth interval Δd is calculated according to the following formula: △d=(△d1+△d2) / 2; △d1=2*tan(a / 2)*R; △d2=2*tan(b / 2)*R; Where △d1 and △d2 are the water depth intervals when the beam opening angles are a° and b° respectively; R is the radius of the target cage (2).

5. The method for measuring the number of deep-sea cage cultured fish according to claim 3, characterized in that: In step 4, the water depth overlap interval δd is taken as the smaller value between the average fish length and the average fish width of the identified fish.

6. The method for measuring the number of deep-sea cage cultured fish according to claim 1, characterized in that: In step 5, the DBSCAN clustering analysis method is used to perform density clustering processing on the spatial three-dimensional coordinates (x, y, d) belonging to the same water layer; wherein the clustering parameters are set in the following manner: The cluster radius ranges from 0.5N to 0.8N, where N is the larger value of the average fish length and average fish width of the identified fish; The minimum number of samples is set to 1.

7. The method for measuring the number of deep-sea cage cultured fish according to any one of claims 1 to 6, characterized in that: In step 2, the deep learning model adopts the YOLOv8 model, and the fish school recognition model is trained in the following manner: Step 2-1: capture a sufficient number of sonar scanning images from a single cage scanning video, and annotate the position of each identified fish in each frame of the sonar scanning image using the MakeSense online data annotation software to output an annotation result file; Step 2 of 2: Use a single cage scanning video as input data, use the annotation result file corresponding to the single cage scanning video as output data, and train the YOLOv8 model to obtain the fish school recognition model.

8. The method for measuring the number of deep-sea cage cultured fish according to any one of claims 1 to 6, characterized in that: The method for measuring the number of fish schools cultured in deep-sea cages further comprises: drawing a fish quantity variation trend diagram of the fish quantity estimated by water layer clustering in the target cage (2) as the water depth changes, based on the estimated fish quantity by water layer clustering in each water layer calculated in step 5.

Images