A method and system for evaluating water ecological environment quality based on fish monitoring

Abstract

The present invention discloses a method and system for evaluating the quality of water ecological environment based on fish monitoring. The method provided by the present invention includes: obtaining information of fish in a pre-selected water body area based on echo detection; collecting fish images according to the information of fish; inputting the fish images into a neural network model to identify the species of fish in the fish images; obtaining the total number of fish species in the pre-selected water body area, and obtaining the fish integrity index according to the total number of fish species and the conservative coefficient of the fish species; and judging the water ecological environment quality level of the pre-selected water body area according to the fish integrity index. The present invention can timely, accurately and comprehensively obtain the water ecological environment quality level in the water body area for use in water body ecological environment status evaluation and early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of water ecological environment quality evaluation, and in particular to a method and system for evaluating water ecological environment quality based on fish monitoring. Background Art

[0002] Usually, people mainly use physical, chemical means and biological monitoring to evaluate the status of water ecological environment. Among them, biological monitoring has been widely used in water ecological environment monitoring, which mainly includes biological index method, biodiversity index method, microbial community monitoring method, bacteriological inspection monitoring method, biological toxicity test method, biomarker method, etc. Common indicator organisms include protists, phytoplankton, benthic organisms, amphibians and fish.

[0003] Fish are at a high level in the aquatic food chain. They can reflect the impact of pollutants on other organisms in the water as well as the environmental stress effects of pollutants on themselves. They are currently the most widely used indicator organisms.

[0004] Traditional fish monitoring requires regular and periodic sampling of fish catches to record the species and quantity of fish, which is time-consuming and labor-intensive. Moreover, the same survey team conducts surveys, but the data obtained are not comparable due to different daily survey periods, durations, and seasons. At the same time, in the process of using the fish biodiversity index to assess the quality of the water ecological environment, the calculation results of the diversity index are only related to the number of fish species, and have little to do with the ecological attributes of the fish themselves. Therefore, the assessment results are not very sensitive to habitat differences.

[0005] Therefore, there is an urgent need for an automatic monitoring and non-invasive fish survey method to evaluate the water ecological environment status of the water body in which the fish is located based on the situation of the fish. Summary of the invention

[0006] In order to overcome the above problems, the inventors propose a method and system for evaluating the quality of water ecological environment based on fish monitoring.

[0007] In a first aspect, the present invention provides a method for evaluating the quality of aquatic ecological environment based on fish monitoring, comprising:

[0008] S101, obtaining information about fish in a preselected water area based on echo detection, wherein the information about fish includes the location and / or distribution of the fish;

[0009] S102, collecting fish images according to the information of the fish;

[0010] S103, inputting the fish image into the neural network model to identify the species of the fish in the fish image;

[0011] S104, obtaining the total number of fish species in the pre-selected water area, and obtaining a fish integrity index according to the total number of fish species and a conservative coefficient of the fish species; and

[0012] S105. Determine the water ecological environment quality level of the pre-selected water area based on the fish integrity index.

[0013] In a second aspect, the present invention provides a system for evaluating the quality of aquatic ecological environment based on fish monitoring, which comprises:

[0014] An echo sounder, which obtains information about fish in a preselected water area based on echo sounding, wherein the information about fish includes the location and / or distribution of the fish;

[0015] A video acquisition device, which acquires fish images based on the information of the fish;

[0016] A recognition module, which inputs the fish image into the neural network model to identify the species of the fish in the fish image;

[0017] An analysis module, which obtains the total number of fish species in a pre-selected water area, and obtains a fish integrity index according to the total number of fish species and a conservative coefficient of the fish species; and

[0018] The grade judgment module judges the water ecological environment quality grade of the pre-selected water area according to the fish integrity index.

[0019] The method and system for evaluating water ecological environment quality based on fish monitoring provided by the present invention have the following beneficial effects:

[0020] (1) The method provided by the present invention can timely, accurately and comprehensively obtain the water ecological environment quality level in the water area, so as to be used for water ecological environment status evaluation and early warning;

[0021] (2) The method provided by the present invention is simple to operate, can be automated and intelligent, and can also be conveniently remotely controlled, thereby using very little manpower and material resources to monitor a wide range of water areas, facilitating the promotion and application of technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram showing a flow chart of a method for evaluating water ecological environment quality based on fish monitoring according to the present invention;

[0023] Figure 2 A schematic diagram showing the structure of a system for evaluating water ecological environment quality based on fish monitoring of the present invention;

[0024] Figure 3 A schematic diagram of the water ecological environment status of Baiyangdian in 2019 is shown in Example 1. DETAILED DESCRIPTION

[0025] The present invention will be further described in detail below through the accompanying drawings and embodiments. Through these descriptions, the characteristics and advantages of the present invention will become more clear and distinct.

[0026] The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise noted.

[0027] Traditional fish survey methods are mainly based on netting, but this process has high labor costs and is harmful to fish. In recent years, acoustic methods have also been increasingly popularized, which have the advantages of being fast, convenient, and having a low risk of damaging biological resources; however, the use of acoustic methods will still interfere with fish populations to a certain extent. With the increasing improvement of current underwater photography technology, if fish identification can be directly carried out by combining photography with image recognition assistance, the interference with fish can be minimized.

[0028] Therefore, the present invention provides a method and system for evaluating the quality of water ecological environment based on fish monitoring. In this method, the total number of fish species in the pre-selected water area is obtained by combining the acoustic method and the optical method, and the fish integrity index is calculated by combining the conservative coefficient of each fish species, thereby obtaining the water ecological environment quality grade. The present invention can realize the organic combination of fish monitoring and water ecological environment quality evaluation, thereby providing data reference for water ecological evaluation and biodiversity protection and restoration.

[0029] In a first aspect, the present invention provides a method for evaluating the quality of aquatic ecological environment based on fish monitoring. Figure 1 As shown, the method mainly includes:

[0030] S101. Obtain information about fish in a preselected water area based on echo detection.

[0031] The information about fish includes the location and / or distribution of the fish.

[0032] In the present invention, the size of the pre-selected water body area is selected according to the actual situation of the water body, for example, an entire lake, a section of a river flowing through a city, etc. can be selected.

[0033] Preferably, step S101 may further include:

[0034] S101-1. Select multiple sampling points in the preselected water area, for example, at least 3 sampling points;

[0035] S101-2. Perform at least one echo detection at at least one sampling point.

[0036] Preferably, each sampling point performs at least three echo detections within a preselected time period to ensure that information about fish in the preselected water area is obtained as much as possible.

[0037] The present invention does not impose any specific restrictions on the location and number of sampling points, which can be determined by those skilled in the art based on the actual situation of the pre-selected water area. Preferably, the number of sampling points is proportional to the size of the pre-selected water area, and the distribution of the sampling points should be relatively uniform, so that most or even all fish in the pre-selected water area can be fully located.

[0038] In the present invention, an echo sounder may be used to implement echo detection.

[0039] The transducer, the main component of the echo sounder, can convert electronic signals into sound pulses and emit them into the water. When the sound waves encounter obstacles during propagation in the water, part of the sound signals are backscattered to the transducer. If the signal level exceeds the preset threshold, it will be received by the transducer and converted into electronic signals to generate an echo map. According to the interval time of receiving the fish echo signal after the sound wave is emitted and the propagation speed of the sound wave in the water, the position of the fish can be measured, and the strength and structure of the echo signal can be analyzed to estimate the information of the fish. When the echo sounder is used to investigate fish, it has the characteristics of fast speed, real-time tracking within a certain period of time, and no damage to the fish. The data obtained can provide a basis for further research.

[0040] Specifically, the echo sounder is installed on an unmanned vessel that is capable of navigating in a preselected water area and mooring at a sampling point.

[0041] Among them, there are usually aquatic plants such as floating plants and submerged plants, or obstacles such as artificial nets in the pre-selected water area. If the unmanned ship sails arbitrarily in the pre-selected water area, it may destroy the growth of aquatic plants or be caught by artificial nets. Therefore, the unmanned ship needs to sail according to the preset navigation route so that the unmanned ship can avoid obstacles in time.

[0042] Therefore, in the present invention, it is also necessary to correct the navigation route of the unmanned ship in real time to ensure that the unmanned ship can navigate to the sampling point according to the preset navigation route, further saving sampling time and improving sampling efficiency.

[0043] Specifically, the deflection angle of the propeller of the unmanned ship is obtained in real time based on the forward-looking distance, and the deflection angle is transmitted to the propeller to control the deflection of the propeller, so that the unmanned ship follows a preset navigation route.

[0044] More specifically, the process of obtaining the propeller deflection angle in real time includes:

[0045] (1-1) For each value of the foresight distance, the corresponding comprehensive deviation at the next moment is obtained, and the foresight distance that minimizes the comprehensive deviation at the next moment is selected.

[0046] In this process, the coordinates are horizontal plane coordinates, i.e., two-dimensional coordinates, in which the X-axis direction is toward the east and the Y-axis direction is toward the north.

[0047] Among them, the foresight distance l d is the straight-line distance between the center position of the propeller of the unmanned ship and the foresight point. The foresight point is a point in the preset navigation route, which is determined by selecting the foresight distance that minimizes the comprehensive deviation at the next moment.

[0048] In the present invention, the range and value accuracy of the foresight distance are first determined according to the situation of the unmanned ship, and then each value of the foresight distance is obtained according to the range and value accuracy of the foresight distance. d The value range is 3m≤l d ≤60m; l d The value accuracy is 2m, traversing l d Each value of l d =3m, 5m, 7m...60m.

[0049] In the present invention, a corresponding comprehensive deviation is obtained based on the value of each foresight distance.

[0050] The comprehensive deviation at the next moment is obtained by formula 1:

[0051] e(t+1)=F1+F2+F3 Formula 1

[0052] e(t+1) represents the comprehensive deviation at the t+1th moment;

[0053] in,

[0054] F1=w1(ge(t)-N(t+1))+w2(gn(t)-E(t+1))

[0055] F2=w1vt0cosθ(t)+w2vt0sinθ(t)

[0056]

[0057] Wherein, θ(t) represents the heading angle at the tth moment, and the heading angle is the angle between the navigation direction of the unmanned ship and the X-axis in the horizontal coordinate; the heading angle is obtained by an inertial navigation device set on the unmanned ship, such as a gyroscope.

[0058] (ge(t),gn(t)) represents the coordinates of the center position of the propeller of the unmanned boat at the tth moment. In the actual working process, the center position of the propeller of the unmanned boat can be identified as the geometric center position of the propeller, and the relative position relationship between the center position and the unmanned boat can be known when the unmanned boat leaves the factory; ge(t) represents the coordinates of the center position of the propeller of the unmanned boat on the X-axis at the tth moment, and gn(t) represents the coordinates of the center position of the propeller of the unmanned boat on the Y-axis at the tth moment.

[0059] (E(t+1), N(t+1)) represents the coordinates of the point on the preset navigation route that is closest to the propeller center position coordinates (ge(t+1), gn(t+1)) at the t+1 moment. Correspondingly, E(t+1) represents the coordinates of the point on the X-axis, and N(t+1) represents the coordinates of the point on the Y-axis.

[0060] v represents the navigation speed of the unmanned ship, and t0 represents the sampling period, for example, t0 = 0.3s, that is, the interval between time t and time t+1 is 0.3s.

[0061] (E aim ,N aim ) represents the coordinates of the foresight point at time t, that is, E aim represents the coordinate of the foresight point on the X-axis at the tth moment, N aim Represents the coordinate of the foresight point on the Y axis at the tth moment;

[0062] After obtaining the value of each foresight distance, the coordinates of the foresight point (E aim ,N aim ):

[0063] (ge(t)-E aim ) 2 +(gn(t)-N aim ) 2 = l d 2 Formula 2

[0064] This formula reflects the process of drawing a circle. If there is a solution, that is, a circle with the center position of the propeller of the unmanned boat as the center and the foresight distance as the radius intersects with the preset navigation route, there is a corresponding foresight point. If there is no solution, it means that there is no intersection and there is no foresight point.

[0065] Among them, when there is only one intersection point, the intersection point is the front view point. When there are two intersection points, the intersection point closer to the front is the front view point. In the present invention, the navigation direction of the unmanned ship is the front.

[0066] w1 represents the weight of the lateral deviation, and w2 represents the weight of the longitudinal deviation. Preferably, the weight w1 of the lateral deviation and the weight w2 of the longitudinal deviation are obtained by the following steps:

[0067] With the center position of the propeller of the unmanned boat (ge(t),gn(t)) as the center of the circle, and the minimum foresight distance l dmin Detect points on the preset navigation route within a circular area with a radius of l; the minimum foresight distance is l dmin is the minimum value of the foresight distance, i.e. 3m;

[0068] If there is no point on the preset navigation route in the circular area, then w1=w2=0.5;

[0069] If there are points on the preset navigation route in the circular area, then find the point (E1, N1) with the smallest lateral deviation e1 from the center of the circle from the points on the preset navigation route in the circular area, find the point (E2, N2) with the smallest longitudinal deviation e2 from the center of the circle from the points on the preset navigation route in the circular area, and calculate the minimum lateral deviation e 1min and the minimum longitudinal deviation e 2min ;

[0070] When e 1min ≥e 2min When w1=0.9, w2=0.1;

[0071] When e 1min <e 2min , w1=0.8, w2=0.2.

[0072] The purpose of this method is to make the comprehensive deviation approach zero. Although it is impossible to make the comprehensive deviation always approach zero under various practical constraints, this method fully considers the coupling factors of lateral deviation and longitudinal deviation, and can minimize the comprehensive deviation within the achievable value range.

[0073] In the present invention, the traversal range of the foresight distance is set in combination with the speed of the unmanned ship and the sampling period, which can ensure that at least one intersection point is obtained in each period, so that the foresight distance that minimizes the comprehensive deviation at the next moment can be selected; after the foresight distance is determined, the coordinates of the foresight point can be determined accordingly, and then the comprehensive deviation can be obtained by combining Formula 1 and Formula 2.

[0074] (1-2) Obtain the propeller deflection angle δ(t) according to the foresight distance;

[0075] Preferably, the propeller deflection angle is expressed by formula 3:

[0076]

[0077] Among them, L represents the length of the unmanned ship, and α and β represent two intermediate parameters.

[0078] The above steps (1-1) and (1-2) are repeated once in each sampling period, and the propeller is controlled according to the obtained propeller deflection angle.

[0079] S102: Collect fish images based on the information of the fish.

[0080] Among them, after obtaining the information (position and / or distribution) of the fish at the sampling point using the echo sounder, the video acquisition device is immediately turned on. The video acquisition device is also installed on the unmanned boat.

[0081] Among them, since there are too many repeated elements in adjacent frame images in the continuous video of the video acquisition device, if all the continuous frame images are input into the neural network model, a lot of redundant information will be included. Therefore, in the present invention, the video acquired by the video acquisition device is extracted according to a preset time interval to obtain the fish image. For example, the preset time interval is 0.1 to 0.5 seconds, so as to ensure the continuity of the fish image.

[0082] Preferably, the fish images are stored in chronological order to obtain video recordings as historical data for subsequent analysis.

[0083] S103: input the fish image into the neural network model to identify the species of the fish in the fish image.

[0084] In the prior art, before an image is input into a neural network model, it is necessary to perform pre-processing operations on the image, including image scaling, normalization, etc. When scaling the image size to the size specified by the neural network model, the current conventional method is to directly adjust the image to the size specified by the neural network model, but this operation will cause the image to change its original scale, resulting in distortion and deformation of the target in the image, resulting in unsatisfactory target detection results.

[0085] Therefore, in order to ensure that the fish image maintains the original scale and improve the detection effect, the present invention may further include the following steps before step S103:

[0086] (2-1) Create a blank image, where the size of the blank image is the size specified by the neural network model;

[0087] (2-2) Paste the fish image data onto the blank picture.

[0088] Preferably, the format of the blank picture can be common formats such as img, jpg, tiff, etc. In this embodiment, there is no restriction on the format of the blank picture, and technical personnel in this field can set it according to actual conditions, but the size of the blank picture needs to comply with the size specified by the neural network model.

[0089] At present, when using neural network models to recognize images, it is necessary to detect the target first. Among them, the point-supervised target detection method can improve the discriminative target location information and has higher accuracy in complex scenes.

[0090] Since the underwater background and lighting are relatively complex, the present invention preferably adopts a point-supervised target detection method.

[0091] Preferably, the neural network model is obtained through training, and during the training process, the historical fish images in the training samples are annotated by points. The historical fish images refer to the fish images in the training samples. The training samples can be a public data set of fish4knowledge or a data set obtained based on local fish annals.

[0092] At present, point-supervised target detection methods cannot utilize the information of labeled points, which makes different targets easily confused and results in low target recognition accuracy.

[0093] In order to improve the accuracy of target recognition, the neural network model of the present invention includes a point-to-frame network and a target detection network, wherein the point-to-frame network is used to convert point annotations in historical fish images into pseudo-frame annotations, and the target detection network is trained using historical fish images with pseudo-frame annotations.

[0094] The target detection network can be any image recognition network, such as Faster-RCNN, RetinaNet, Reppoint, preferably Faster-RCNN.

[0095] In the present invention, the process of converting the point annotations in the historical fish images into pseudo-frame annotations by the point-to-frame network can be:

[0096] The point-to-box network generates different proposal boxes centered on the annotation points;

[0097] The classifier is trained through the generated proposal boxes to obtain the pseudo-annotated boxes of fish.

[0098] In the present invention, the proposal frames are balanced by the number of fish species, thereby improving the training effect of the classifier and thus improving the recognition accuracy.

[0099] Furthermore, each annotation point generates multiple proposal boxes. The present invention does not impose a specific limit on the number of proposal boxes generated for each annotation point, and those skilled in the art can determine the optimal number of proposal boxes generated for each annotation point based on multiple experiments.

[0100] Since there may be multiple annotation points on the historical fish images, in the present invention, p j =(p jx ,p jy), j∈R, j represents the serial number of different annotation points on a frame of historical fish image.

[0101] Specifically, the proposal frame can be expressed by formula 4:

[0102] b jo =(b jx ,b ky ,b jiw ,b jih )=(p jx ,p jy ,v ji ·s ji ,s ji / v ji ) Formula 4

[0103] Among them, p jx ,p jy The annotation point p j The coordinates of the proposed box are taken as the midpoint b jx ,b jy , subscript i represents the i-th proposal box, b jiw is the width of the i-th proposal box, b jiw =v ji ·s ji ; b jih represents the height of the i-th proposal box, b jih =s ji / v ji ;s ji represents the size of the i-th proposal box, v ji is the adjustment ratio of the i-th proposal box.

[0104] By adjusting s ji and v ji By taking the value of , you can get multiple suggestion boxes.

[0105] The study found that when s ji If it is set too large, it will cause the proposal box to exceed the boundary of the historical fish image and will cause meaningless padding.

[0106] In a preferred embodiment, after obtaining a plurality of proposal frames, the proposal frames are further optimized so that the proposal frames are within the range of historical fish images.

[0107] Specifically, the width and height of the proposal box are constrained by formula 5:

[0108] b jiw =min(v ji ·s ji ,2(p jx -0),2(W r -p jx ))

[0109] b jih =min(s ji / v ji ,2(p jy -0),2(H r -p jy )) Formula 5

[0110] Among them, W r represents the width of the historical fish image in frame r, H r Represents the height of the r-th frame of the historical fish image.

[0111] Among them, the classifier is a two-stream structure MIL classifier, which is a classifier proposed in the paper Bilen, H., Vedaldi, A.: Weakly supervised deep detection networks. In: CVPR (2016) 1, 4, 7, including the classification branch f cls and instance branch f ins , its specific structure is not described in detail in the present invention.

[0112] Furthermore, the process of training the classifier to obtain the pseudo-annotated box of the fish includes:

[0113] Get the feature F in the proposal box j , the feature F j Input classification branch f cls Get the classification score; set feature F j Enter instance branch f ins Get the instance score, perform Hadamard direct product of the classification score and the instance score to get the proposal box score S j ;

[0114] According to the classifier loss function, the relationship between the predicted species and the fish species corresponding to different proposal boxes is obtained, and the mean of the proposal boxes is calculated by weighted average using the score as the weight, and the mean is used as the pseudo-annotation box of the fish.

[0115] Preferably, take the top q1 highest scores S of fish j The proposed boxes are weighted averaged with the scores as weights to obtain the weighted average of the proposed boxes, and the mean is used as the pseudo-annotated box of the fish, thereby improving the robustness of the system.

[0116] Furthermore, the feature F in the proposal box jThe proposed box is obtained by processing the historical fish images through the RoIAlign layer and two convolutional layers. The details of the RoIAlign layer can be found in the paper He, K., Gkioxari, G., et al., PD: Mask R-CNN. In: ICCV (2017) 7, which will not be described here.

[0117] Furthermore, the classification score can be expressed by formula 6:

[0118]

[0119] The instance score can be expressed by formula 7:

[0120]

[0121] in, represents the classification score, Represents the instance score. Both the classification score and the instance score are U×K matrices, where U represents the labeled point p j The total number of corresponding proposal boxes, K represents the total number of fish species in the training samples, [·] uk Represents the element in the uth row and kth column of the matrix.

[0122] Proposal box score S j It can be expressed as: ⊙ represents the Hadamard direct product.

[0123] Further preferably, the classifier loss function It can be expressed by formula eight:

[0124]

[0125] Among them, c j Represents the annotation point p j Corresponding to the species label of fish, m represents the number of fish in the fish image, k represents the kth species of fish in the training sample, and K represents the total number of fish species in the training sample. Indicates the score S of the fish corresponding proposal box j The sum of .

[0126] In the present invention, by converting point annotations into pseudo-frame annotations, accurate fish species identification results are obtained while saving costs, thereby improving the accuracy of fish species identification and ultimately achieving reasonable water ecological environment quality grade evaluation results.

[0127] S104. Obtain the total number of fish species in the pre-selected water area, and obtain a fish integrity index based on the total number of fish species and a conservative coefficient of fish species.

[0128] Specifically, a video acquisition device is used to acquire video at different time periods at the sampling point for multiple times, for example, 10 to 50 times, and the video acquired by the video acquisition device is extracted at a preset time interval to obtain multiple frames of fish images. Step S103 is repeated to obtain most or all species of fish in the pre-selected water area as much as possible.

[0129] Preferably, the fish integrity index can be expressed by formula 9:

[0130]

[0131] Among them, QI fish Represents the fish integrity index; C m represents the conservative coefficient of the mth fish species in the pre-selected water area; n represents the total number of fish species in the pre-selected water area.

[0132] It can be seen from the above formula that the total number of fish species and the conservative coefficient of fish species jointly determine the fish integrity index of the pre-selected water area.

[0133] The conservative coefficient of each fish species can be obtained by expert assignment. For example, 5 to 7 experts can assign a conservative coefficient for each fish species based on Table 1, field survey conditions, and fish species characteristics, and then remove the highest and lowest values, and take the average of the remaining values ​​as the conservative coefficient C of the corresponding fish species. m .

[0134] Table 1 Conservative coefficient assignment standard

[0135]

[0136]

[0137] S105. Determine the water ecological environment quality level of the pre-selected water area based on the fish integrity index.

[0138] In the present invention, the water ecological environment quality is divided into five levels, namely poor, relatively poor, average, relatively good and good.

[0139] Among them, there is a corresponding relationship between the fish integrity index and the water ecological environment quality level of the pre-selected water body area. Therefore, the water ecological environment quality level of the pre-selected water body area can be evaluated based on the fish integrity index, thereby providing corresponding technical support for the protection or restoration of fish biodiversity in the pre-selected water body area.

[0140] Specifically, according to the calculated fish integrity index QI fish Classify the water ecological environment quality levels.

[0141] Select more than 5 sampling points as reference points in waters with similar natural geographical conditions and rich fish diversity as the pre-selected water area. fish The 95% quantile of the numerical distribution is used as the "good" grade evaluation standard, and the distribution range with values ​​less than the 95% quantile is divided into 4 equal parts, corresponding to good, average, poor and bad evaluation standards, thereby providing data reference for the protection and restoration of fish biodiversity in the pre-selected water areas.

[0142] If the pre-selected water area is a river, the reference point is selected from rivers with similar natural geographical conditions and rich fish diversity; if the pre-selected water area is a lake or reservoir, the reference point is selected from lakes or reservoirs with similar natural geographical conditions and rich fish diversity.

[0143] In a second aspect, the present invention provides a system for evaluating the quality of aquatic ecological environment based on fish monitoring. Figure 2 As shown, the system mainly includes:

[0144] An echo sounder 201, which acquires information about fish in a preselected water area based on echo sounding, wherein the information about fish includes the location and / or distribution of the fish;

[0145] A video acquisition device 202, which acquires fish images based on the information of the fish;

[0146] The recognition module 203 inputs the fish image into the neural network model to recognize the species of the fish in the fish image;

[0147] An analysis module 204 is configured to obtain the total number of fish species in the pre-selected water area, and obtain a fish integrity index according to the total number of fish species and a conservative coefficient of the fish species; and

[0148] The grade judgment module 205 judges the water ecological environment quality grade of the pre-selected water area according to the fish integrity index.

[0149] The system for evaluating water ecological environment quality based on fish monitoring provided by the present invention can be used to execute the method for evaluating water ecological environment quality based on fish monitoring described in the first aspect above. Its implementation principle and technical effects are similar and will not be repeated here.

[0150] In the present invention, the echo sounder 201 and the video acquisition device 202 can be installed on an unmanned ship. The unmanned ship can also be provided with a GPS receiver, an electronic map module and a route planning module; after the sampling points are delineated by the electronic map, the route planning module presets the navigation route so as to facilitate sampling at the sampling points.

[0151] Preferably, a channel correction module is also provided on the unmanned ship, through which the preset navigation route of the unmanned ship can be corrected in real time. The channel correction module can correct in real time the problem that the unmanned ship deviates from the preset navigation route due to water flow and wind direction, so that the unmanned ship can follow the preset navigation route, further saving sampling time and improving sampling efficiency.

[0152] In the present invention, the channel correction module can obtain the deflection angle of the propeller of the unmanned ship in real time based on the foresight distance, and transmit the deflection angle to the propeller to control the propeller deflection, so that the unmanned ship can sail to the sampling point according to the preset navigation route.

[0153] The present invention is further described below by means of specific examples, but these examples are merely exemplary and do not constitute any limitation to the protection scope of the present invention.

[0154] Example

[0155] Example 1

[0156] Fish monitoring data were used to evaluate the water ecological environment of Baiyangdian in 2019, and sampling points were set up to cover lakes and ponds with a basic coverage area of ​​more than 100 mu. One sampling point was set up in lakes and ponds with a smaller area; 2 to 3 sampling points were set up in lakes and ponds with larger areas, such as Xiaobaiyangdian, Shaochedan, Chiyudian, Zaoshandian, etc., for a total of 35 sampling points. Unmanned boats carrying echo sounders and video acquisition equipment were controlled to cruise on the water body of Baiyangdian, and the navigation route was set and the navigation of the unmanned boat was controlled accordingly, and it was moored at 35 sampling points for sampling;

[0157] Using an echo sounder, 10 echo soundings were performed at each sampling point to obtain information about the fish at the sampling point, including their distribution and location.

[0158] After obtaining the fish information at each sampling point, the video acquisition device is immediately turned on according to the fish information to start collecting fish images.

[0159] The fish images are fed into the neural network model to identify the species of fish in the fish images.

[0160] The neural network model is obtained through training. During the training process, the historical fish images of the training samples are annotated by points. The training samples are a data set obtained from the Hebei Fish Records.

[0161] The neural network model includes a point-to-box network and a target detection network. The point-to-box network is used to convert point annotations in historical fish images into pseudo-box annotations. The target detection network is trained using historical fish images with pseudo-box annotations to obtain the final neural network model.

[0162] Among them, the target detection network is Faster-RCNN (see the paper Ren, S., He, K., et al., RBG: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE TPAMI (2017) 1, 4, 10, 11, 12, 13, 14 for details).

[0163] The point-to-box network generates different proposal boxes centered on the annotation points, and trains the classifier through the generated proposal boxes to obtain the pseudo-annotation boxes of fish.

[0164] The proposed frame is expressed by formula 4:

[0165] b ji =(b jx ,b jy ,b jiw ,b jih )=(p jc ,p jy ,v ji ·s ji ,s ji / v 2i ) Formula 4

[0166] Among them, p jx ,p jy The annotation point p j The coordinates of the proposed box are taken as the midpoint b jx ,b jy , subscript i represents the i-th proposal box, b jiw is the width of the i-th proposal box, b jiw =v ji ·s ji , b jih represents the height of the i-th proposal box, b jih =s ji / v ji ,s ji represents the size of the i-th proposal box, v ji is the adjustment ratio of the i-th proposal box.

[0167] The width and height of the proposed box are constrained by formula 5:

[0168] b jiw =min(v ji ·s ji ,2(p jx -0),2(W r -p jx ))

[0169] bjih =min(s ji / v ji ,2(p jy -0),2(H r -p jy )) Formula 5

[0170] Among them, W r Represents the width of the historical fish image in frame r, H r Represents the height of the r-th frame of the historical fish image.

[0171] Among them, the classifier is a two-stream structure MIL classifier, including the classification branch f cls and instance branch f ins .

[0172] The process of training the classifier to obtain the pseudo-annotation box of fish includes:

[0173] Get the feature F in the proposal box j , the feature F j Input classification branch f cls Get the classification score; set feature F j Enter instance branch f ins Get the instance score, perform Hadamard direct product of the classification score and the instance score to get the proposal box score S j ;

[0174] According to the classifier loss function, the relationship between the predicted species corresponding to different proposal boxes and the fish they belong to is obtained, and the top 5 fish with the highest scores S are taken. j The weighted average of the proposed boxes is calculated using the scores as weights, and the mean is used as the pseudo-annotated box of the fish.

[0175] The classification score is expressed by formula 6:

[0176]

[0177] The instance score is expressed by formula 7:

[0178]

[0179] in, represents the classification score, Represents the instance score. Both the classification score and the instance score are U×K matrices, where U represents the labeled point p j The total number of corresponding proposal boxes, K represents the total number of fish species in the training samples, [·] uk Represents the element in the uth row and kth column of the matrix.

[0180] Proposal box score S j It is expressed as: ⊙ represents the Hadamard direct product.

[0181] Classifier loss function Expressed in formula eight:

[0182]

[0183] Among them, c j Represents the annotation point p j Corresponding to the species label of fish, m represents the number of fish in the image, k represents the kth species of fish in the training sample, and K represents the total number of fish species in the training sample. Indicates the score S of the fish corresponding proposal box j The sum of .

[0184] The total number of fish species in Baiyangdian was calculated, and the fish integrity index was obtained based on the total number of fish species and the conservative coefficient of fish species.

[0185] Among them, the fish integrity index is expressed by formula 9:

[0186]

[0187] Among them, QI fish Represents the fish integrity index; C m represents the conservative coefficient of the mth fish species in each sampling point; n represents the total number of fish species in each sampling point.

[0188] Based on the fish integrity index, the water ecological environment quality level of each sampling point in Baiyangdian is judged.

[0189] Baiyangdian is the largest freshwater lake wetland in North China, with rich fish diversity, so the reference point was selected among the sampling points of Baiyangdian. fish The six sampling points with higher values, namely, Zaoshandian South, Zaoshandian North, Laowangdian, Shaochedian West, Damaidian, and Wangjiazhai, were used as reference points. fish The 95% quantile 10 of the numerical distribution is used as the evaluation standard for the "good" level, and 10 is divided into 4 equal parts, namely good, average, poor and bad. The results are shown in Table 2.

[0190] Table 2 Classification standards of Baiyangdian water ecological environment quality

[0191] Water ecological environment quality level Classification criteria good <![CDATA[QI fish ≥10]]> better <![CDATA[10<QI fish ≤7.5]]> generally <![CDATA[7.5<QI fish ≤5]]> Poor <![CDATA[5<QI fish ≤2.5]]> Difference <![CDATA[QI fish <2.5]]>

[0192] Total number of fish species, average conservative coefficient and QI of 35 sampling points in Baiyangdian Lake fish The values ​​are shown in Table 3 and Figure 3 .

[0193] Table 3 Results of fish survey and water ecological environment quality evaluation in Baiyangdian Lake

[0194]

[0195]

[0196] From Table 3 and Figure 3 It can be seen that the water ecological environment quality level in the northern part of Baiyangdian Lake is "good", the water ecological environment quality level in the middle part of the lake area is "relatively good", and the water ecological environment quality level in the southern part of the lake area is "average" and "poor". This is basically consistent with the evaluation results of the health status of the Baiyangdian water ecosystem in the literature (Li Wenjun, Kang Lixin, Zhao Yanchu, Kong Fanqing. Research on the evaluation of the health status of the Baiyangdian wetland water ecosystem based on the large benthic animal integrity index [J]. Environmental Science and Management, 2022, 47(03): 164-168).

[0197] Therefore, while protecting the water ecological environment quality in the north and central parts of Baiyangdian, it is necessary to strengthen the restoration of the water ecosystem in the southern waters to enhance its biodiversity.

[0198] The present invention has been described above in conjunction with preferred embodiments, but these embodiments are only exemplary and serve only as an illustration. On this basis, the present invention may be subjected to a variety of substitutions and improvements, all of which fall within the scope of protection of the present invention.

Claims

1. A method for evaluating the quality of water ecological environment based on fish monitoring, characterized in that: include: S101, obtaining information about fish in a preselected water area based on echo detection, wherein the information about fish includes the location or distribution of the fish; S102, collecting fish images according to the information of the fish; S103, inputting the fish image into a neural network model to identify the species of the fish in the fish image; S104, obtaining the total number of fish species in the pre-selected water area, and obtaining a fish integrity index according to the total number of fish species and a conservative coefficient of fish species; S105, judging the water ecological environment quality level of the pre-selected water area according to the fish integrity index; Step S101 further includes: S101-1. Select multiple sampling points in the pre-selected water area; S101-2, performing at least one echo detection at at least one sampling point; The echo sounder is installed on the unmanned ship, and the unmanned ship sails according to the preset route. The route of the unmanned ship is corrected in real time to ensure that the unmanned ship can sail to the sampling point according to the preset route. The deflection angle of the propeller of the unmanned ship is obtained in real time based on the forward distance, and the deflection angle is transmitted to the propeller to control the propeller deflection, so that the unmanned ship can sail according to the preset route; The process of obtaining the propeller deflection angle in real time includes: (1-1) For each value of the foresight distance, obtain the corresponding comprehensive deviation at the next moment, and select the foresight distance that minimizes the comprehensive deviation at the next moment; The comprehensive deviation at the next moment is obtained by formula 1: e(t+1)=F1+F2+F3 Formula 1 e(t+1) represents the comprehensive deviation at the t+1th moment; in, F1=w1(ge(t)-N(t+1))+w2(gn(t)-E(t+1)) F2=w1vt0cosθ(t)+w2vt0sinθ(t) Wherein, θ(t) represents the heading angle at the tth moment; (ge(t), gn(t)) represents the coordinates of the center position of the propeller of the unmanned ship at the tth moment; ge(t) represents the coordinates of the center position of the propeller of the unmanned ship at the tth moment on the X-axis, and gn(t) represents the coordinates of the center position of the propeller of the unmanned ship at the tth moment on the Y-axis; (E(t+1), N(t+1)) represents the coordinates of the point on the preset navigation route that is closest to the coordinates of the propeller center position at the t+1th moment (ge(t+1), gn(t+1)), and accordingly E(t+1) represents the coordinates of the point on the X-axis, and N(t+1) represents the coordinates of the point on the Y-axis; v represents the navigation speed of the unmanned ship, t0 represents the sampling period, (E aim ,N aim ) represents the coordinates of the foresight point at time t, N aim represents the coordinate of the foresight point on the Y axis at the tth moment; w1 represents the weight of the lateral deviation, w2 represents the weight of the longitudinal deviation, l d Represents the foresight distance, which is the straight-line distance between the center of the propeller of the unmanned ship and the foresight point; (1-2) Obtain propeller deflection angle according to forward sight distance; Before step S103, the method further includes: Create a new blank image, where the size of the blank image is the size specified by the neural network model; Paste the data of the fish image onto a blank picture; The neural network model is obtained through training, and during the training process, historical fish images in the training samples are marked by points; The neural network model includes a point-to-frame network and a target detection network, wherein the point-to-frame network is used to convert the point annotations in the historical fish images into pseudo-frame annotations, and the target detection network is trained by the historical fish images with pseudo-frame annotations; The point-to-frame network is used to convert the point annotations in the historical fish image into pseudo-frame annotations, including: Generate different proposal boxes centered on the annotation points; The classifier is trained through the generated proposal boxes to obtain the pseudo-annotated boxes of fish; Improve the training effect of the classifier by balancing the proposal boxes by the number of fish species; Among them, each annotation point will generate multiple proposal boxes. After obtaining multiple proposal boxes, the proposal boxes are further optimized so that the proposal boxes are within the range of historical fish images. The optimization is to constrain the width and height of the proposal boxes: b jiw =min(v ji ·s ji ,2(p jx -0),2(W r -p jx )) b jih =min(s ji / v ji ,2(p jy -0),2(H r -p jy )) Among them, W r represents the width of the historical fish image in frame r, H r represents the height of the historical fish image in frame r, b jiw is the width of the i-th proposal box, b jih represents the height of the i-th proposal box, s ji represents the size of the i-th proposal box, v ji is the adjustment ratio of the i-th proposal box, p jx ,p jy The annotation point p j The coordinates of The process of training the classifier to obtain the pseudo-annotation box of fish includes: Get the feature F in the proposal box j , the feature F j Input classification branch f cls Get the classification score; set feature F j Enter instance branch f ins Get the instance score, perform Hadamard direct product of the classification score and the instance score to get the proposal box score S j ; According to the classifier loss function, the relationship between the predicted species corresponding to different proposal boxes and the fish to which they belong is obtained. The score is used as the weight for weighted averaging to obtain the mean of the proposal boxes, and the mean is used as the pseudo-annotation box of the fish.

2. The method for evaluating water ecological environment quality based on fish monitoring according to claim 1 is characterized in that: The proposal frame is represented by formula 4: b ji =(b jx ,b jy ,b jiw ,b jih )=(p jx ,p jy ,v ji ·s ji ,s ji / v ji ) Formula 4 Among them, p jx ,p jy The annotation point p j The coordinates of the proposed box are taken as the midpoint b jx ,b jy , subscript i represents the i-th proposal box, b jiw is the width of the i-th proposal box, b jiw =v ji ·s ji ; b jih represents the height of the i-th proposal box, b jih =s ji / v ki ;s ji represents the size of the i-th proposal box, v ji is the adjustment ratio of the i-th proposal box.

3. The method for evaluating water ecological environment quality based on fish monitoring according to claim 1 is characterized in that: The fish integrity index can be expressed by formula 9: Among them, QI fish Represents the fish integrity index; C m represents the conservative coefficient of the mth fish species in the pre-selected water area; n represents the total number of fish species in the pre-selected water area.

4. The method for evaluating water ecological environment quality based on fish monitoring according to claim 1 is characterized in that: The water ecological environment quality levels include poor, relatively poor, average, relatively good and good.

Images