A method and system for monitoring fish ecology in a protected area

Abstract

The application relates to the technical field of ecological monitoring, and discloses a fish ecological monitoring method and system for a protection zone, wherein the system comprises a data acquisition module, a data preprocessing module, a visual recognition module, a shape estimation module, a multi-source fusion module, a habitat assessment module, a risk early warning module and a data management module. The system combines underwater video, acoustic detection and eDNA three types of sensing technologies, combines habitat suitability assessment and ecological risk early warning, and realizes real-time and non-interference monitoring of fish populations and water environments in the protection zone. The application solves the technical problems of low monitoring frequency, insufficient data precision and large disturbance to fish habitats in traditional manual sampling monitoring, and achieves the technical effects of automation, real-time and non-interference of fish ecological monitoring in the protection zone.

Description

Technical Field

[0001] This invention relates to the field of ecological monitoring technology, specifically to a method and system for ecological monitoring of fish in protected areas. Background Technology

[0002] The conservation of fish populations in the waters of nature reserves relies on continuous monitoring of population dynamics and aquatic habitats. Traditional monitoring methods mainly rely on manual sampling, including net fishing, electrofishing, and underwater observation, to obtain data on population size, species composition, and individual morphology through regular sampling.

[0003] The aforementioned manual sampling methods have the following shortcomings: the monitoring range is limited by the personnel's operational capabilities, making it difficult to cover all waters of the protected area; the sampling frequency is low, failing to continuously reflect population dynamics; the sampling process directly disturbs individual fish and aquatic habitats; the detection capability for rare and endangered species is insufficient, and the data real-time performance is poor. These shortcomings make it difficult for protected area managers to comprehensively, accurately, and promptly grasp changes in fish population status and habitat quality, failing to meet the needs of precise and routine monitoring. Summary of the Invention

[0004] This invention provides a method and system for monitoring fish ecology in protected areas, which solves the technical problems of single source of fish ecological data in protected area waters, insufficient ability to integrate multi-source data, and lagging ecological risk identification in related technologies.

[0005] This invention discloses a method for monitoring fish ecology in a protected area, comprising at least the following steps: collecting underwater video data, acoustic echo data, eDNA detection data, and aquatic environmental parameter data from underwater camera units, acoustic detection units, eDNA monitoring units, and multi-parameter water quality monitoring instruments deployed in the protected area's waters, generating a multi-source raw monitoring dataset; cleaning and standardizing the multi-source raw monitoring dataset to generate effective video frame sequences, effective acoustic echo sequences, a standardized species detection result list, and an environmental parameter time series; inputting the effective video frame sequences into a fish target detection model to generate a target bounding box set; inputting the fish image regions corresponding to each bounding box in the target bounding box set into a species classification model to generate visual recognition results; and utilizing... The three-dimensional spatial coordinates of each fish target are calculated using the binocular parallax information of the underwater camera unit, and then the body length is calculated. Body weight is calculated based on the allometric growth relationship between body length and weight, generating individual morphological data. Visual recognition results, effective acoustic echo sequences, and a standardized list of species detection results are fused to generate fused population statistics for each monitoring point. The fused population statistics and environmental parameter time series are input into a habitat suitability assessment model to generate habitat suitability assessment results including habitat suitability levels and key environmental parameter identifiers. Based on the fused population statistics, environmental parameter time series, and habitat suitability assessment results, ecological risk events are identified, risk events are classified, early warning information is generated, and the early warning information is pushed to management personnel.

[0006] Furthermore, the multi-source fusion adopts a weighted voting method. For the same species at the same monitoring point, the statistical count of visual recognition results, the estimated count of effective acoustic echo sequence analysis, and the existence information of eDNA detection are assigned corresponding weights. The weights are determined based on the statistical accuracy of each technical means in the historical data of that point. When the results of the three sources are consistent, the confidence of the corresponding species quantity estimate is increased. When there is a conflict, the source with higher confidence is given priority, and the conflict is marked. Among them, species with confidence exceeding a preset threshold in the eDNA detection results are retained separately, and these species are included in the fused population statistics and marked as eDNA-specific detected species.

[0007] Furthermore, the allometric growth relationship between body length and body weight is expressed by the formula... It means that among them This is an estimated weight value. This is a measurement of body length. and The species-related allometric growth coefficient, and The values ​​are derived from the pre-stored species parameters in the fish species database of the protected area; when using parallax information to calculate the three-dimensional spatial coordinates, a water refractive index correction coefficient is introduced to correct the parallax value, and then the three-dimensional coordinates are calculated.

[0008] Furthermore, the identification of ecological risk events includes: calculating the deviation of the current value of each key parameter in the environmental parameter time series from the historical baseline mean; when the deviation exceeds a preset anomaly threshold and the duration exceeds a preset duration threshold, it is identified as a water quality anomaly risk event; calculating the population change rate of each species in the fused population statistics data within adjacent time windows; when the absolute value of the population change rate exceeds a preset change rate threshold and there is no significant deterioration in the habitat suitability assessment results, and the duration exceeds a preset duration threshold, it is identified as a fish abnormal mortality or fish abnormal aggregation risk event; and identifying illegal fishing and unauthorized entry behaviors in the human activity image areas around the water body detected in the effective video frame sequence using an image classification model, generating illegal behavior identification results.

[0009] Furthermore, based on the identification results of water quality abnormality risk events, abnormal fish mortality or abnormal fish gathering risk events, and illegal activities, the risk level is assessed into three levels: general, relatively large, and major, according to the risk type and degree of deviation, and an early warning message is generated that includes the risk type, the location of occurrence, the risk level, and the disposal recommendations.

[0010] Furthermore, the habitat suitability assessment model takes the statistical characteristics of environmental parameter time series and the species diversity index and population density from the integrated population statistics as input features, and the habitat suitability level label as output. It completes supervised training on the historical monitoring data of the protected area and the habitat suitability samples marked by experts. The habitat suitability assessment model simultaneously outputs the contribution ranking of each input environmental parameter to the habitat suitability assessment result, and marks the parameters with the highest contribution as key environmental parameters and incorporates them into the habitat suitability assessment result.

[0011] Furthermore, when performing target detection on the effective video frame sequence, the target trajectories between adjacent frames are also associated, and the bounding box sequence of the same fish target in consecutive frames is associated as the target trajectory. Based on the target trajectory, the missed detection caused by single-frame occlusion is filled in.

[0012] Furthermore, when cleaning underwater video data, the frame extraction frequency is dynamically adjusted according to the motion amplitude between video frames: when the motion amplitude between frames is lower than a preset threshold, the extraction frequency is reduced; when the motion amplitude between frames exceeds the preset threshold, the extraction frequency is increased.

[0013] This invention discloses a fish ecological monitoring system for a protected area, comprising: a data acquisition module for acquiring underwater video data, acoustic echo data, eDNA detection data, and aquatic environmental parameter data from an underwater camera unit, an acoustic detection unit, an eDNA monitoring unit, and a multi-parameter water quality monitor, generating a multi-source raw monitoring dataset; a data preprocessing module for cleaning and standardizing the multi-source raw monitoring dataset, generating a valid video frame sequence, a valid acoustic echo sequence, a standardized species detection result list, and an environmental parameter time series; a visual recognition module for inputting the valid video frame sequence into a fish target detection model to generate a target bounding box set, inputting the fish image regions corresponding to each bounding box in the target bounding box set into a species classification model to generate visual recognition results; and a morphology estimation module for using the binocular parallax information of the underwater camera unit to calculate... The system calculates the three-dimensional spatial coordinates of each fish target, then calculates its body length and weight based on allometric growth relationships, generating individual morphological data; a multi-source fusion module is used to fuse visual recognition results, effective acoustic echo sequences, and a standardized species detection result list to generate fused population statistics for each monitoring point; a habitat assessment module is used to input the fused population statistics and environmental parameter time series into a habitat suitability assessment model to generate habitat suitability assessment results including habitat suitability levels and key environmental parameter identifiers; a risk warning module is used to identify ecological risk events based on fused population statistics, environmental parameter time series, and habitat suitability assessment results, classify risk events, and generate warning information; a data management module is used to push warning information to management personnel, store various monitoring data and analysis results in a database, and generate monitoring reports periodically.

[0014] Furthermore, the multi-source fusion module uses a weighted voting method to fuse the same species at the same monitoring point, with the weights determined based on the statistical accuracy of each technical means in the historical data of that point; species with a confidence level exceeding a preset threshold in the eDNA detection results are retained separately, included in the fused population statistics, and labeled as eDNA-specific detected species; the risk warning module requires that abnormal deviations in parameters or abnormal changes in quantity persist for multiple consecutive collection times, and only triggers risk event identification after reaching a preset duration threshold.

[0015] This invention addresses the technical problems of limited coverage and insufficient rare species detection capabilities of single monitoring methods by fusing multi-source sensing data from underwater video, acoustic detection, and eDNA, thereby improving the completeness and accuracy of population statistics. A binocular parallax-based body length and weight estimation method replaces contact measurement, solving the problem of interference with individual fish during traditional sampling and achieving non-contact acquisition of individual morphological data. The habitat suitability assessment model combines environmental parameters and population data as inputs, addressing the lack of objective quantitative evidence for habitat quality assessment and enabling the identification of key environmental parameters to guide habitat restoration. The introduction of a duration threshold in the risk warning process solves the problem of false alarms caused by short-term fluctuations, improving the effectiveness of warning information. Overall, this invention solves the technical problems of low frequency, insufficient data accuracy, and significant interference with fish habitats caused by traditional manual sampling monitoring, achieving automated, real-time, and interference-free fish ecological monitoring in protected areas. Attached Figure Description

[0016] Figure 1 This is a flowchart of the fish ecological monitoring method in protected areas provided in this embodiment of the invention; Figure 2 This is a flowchart of the preprocessing of multi-source raw monitoring datasets in the fish ecological monitoring method for protected areas provided in this embodiment of the invention; Figure 3 This is a sub-flowchart of the method for monitoring fish ecology in protected areas and generating early warning information, provided in this embodiment of the invention. Detailed Implementation

[0017] Fish ecological monitoring in nature reserves faces challenges due to the limited monitoring range, low frequency, poor real-time performance, and significant disturbance to fish and aquatic habitats caused by traditional manual sampling methods. Existing monitoring techniques are insufficient to comprehensively grasp the dynamic changes in fish populations and habitat influencing factors, failing to meet the needs of precise and routine monitoring in nature reserves. This embodiment provides a method for fish ecological monitoring in nature reserves, achieving real-time, non-disruptive, and precise monitoring of fish populations and the aquatic environment through multi-technology integration of underwater sensing and acquisition, multimodal data preprocessing, artificial intelligence identification and analysis, habitat suitability assessment, and ecological risk early warning. This method operates on a data processing system with network communication capabilities, which is connected to underwater camera units, acoustic detection units, eDNA monitoring units, and multi-parameter water quality monitors deployed in the protected area's waters.

[0018] Underwater video data, acoustic echo data, eDNA detection data, and aquatic environmental parameter data are collected from underwater camera units, acoustic detection units, eDNA monitoring units, and multi-parameter water quality monitoring instruments deployed at various monitoring points in the protected area, generating a multi-source raw monitoring dataset.

[0019] It should be noted that underwater video data was acquired by a stereo-BRUV (stereo-visual unit). A stereo-BRUV consists of two underwater cameras at a fixed baseline distance, simultaneously capturing images of the same fish target. Utilizing the principle of binocular parallax, it provides the fish's three-dimensional spatial coordinates, supporting depth data for subsequent length and weight estimation. Acoustic echo data was acquired by a fish finder through the transmission and reception of sonar signals, carrying information on the location, density, and distribution range of underwater fish schools. eDNA detection data was generated by automated water sampling equipment after collecting water samples, followed by DNA extraction and macro-barcode amplification, carrying genetic marker information of fish species in the water. Aquatic environmental parameters, including water temperature, pH, dissolved oxygen, turbidity, ammonia nitrogen, total phosphorus, total nitrogen, water depth, current velocity, and water level, were acquired in real-time by a multi-parameter water quality monitor.

[0020] In this embodiment, in order to ensure the availability of video data in complex underwater environments such as low light and high turbidity, the underwater camera unit is equipped with an underwater supplementary lighting device and a multispectral camera module. The multispectral camera module can switch to near-infrared band acquisition when the visible light imaging effect is limited, thereby improving the effective acquisition rate of video data in complex underwater environments.

[0021] The various types of data in the multi-source raw monitoring dataset are cleaned and standardized to generate standardized monitoring data.

[0022] Step 2 may specifically include: Step 201: Extract and convert the format of the underwater video data, remove blurry and duplicate frames, and generate a valid video frame sequence.

[0023] Step 202: Denoise the acoustic echo data to remove background noise from the water body and non-biological echo interference, and generate an effective acoustic echo sequence.

[0024] Step 203: Perform quality filtering and sequence denoising on the eDNA detection data, match the filtered sequences with the fish reference database to generate a standardized species detection result list. Each record in the standardized species detection result list includes the species name, detection confidence level and corresponding sampling point information.

[0025] Step 204: Detect and remove outliers from the aquatic environmental parameter data, and use a linear interpolation algorithm to fill in missing values ​​to generate a continuous time series of environmental parameters.

[0026] It should be noted that the frame extraction frequency in step 201 is dynamically adjusted according to the fish's activity speed: when the motion amplitude between video frames is lower than a preset threshold, the extraction frequency is reduced to reduce redundant data; when the motion amplitude exceeds the preset threshold, the extraction frequency is increased to preserve the complete motion trajectory, thereby achieving a balance between data volume and information integrity.

[0027] The effective video frame sequence is input into the fish target detection model to locate the fish targets in each frame and generate a set of target bounding boxes. The fish image regions corresponding to each bounding box in the target bounding box set are input into the species classification model to generate the species identification result and identification confidence of each target. The results are then summarized to generate the visual recognition result.

[0028] It should be noted that the fish target detection model uses either YOLOv8 or Cascade R-CNN. YOLOv8 is suitable for monitoring scenarios with high real-time requirements. Its input is a video frame image, and its output is a set of target bounding boxes and the confidence score of each box. The loss function is a weighted sum of cross-entropy loss and IoU regression loss, and it is trained using a stochastic gradient descent optimization strategy. Cascade R-CNN is suitable for scenarios with small targets and dense fish schools. Its input is also a video frame image, and its output is a set of target bounding boxes refined through multi-stage cascaded regression and their confidence scores. The loss function is the sum of the classification loss and bounding box regression loss at each stage, and it is trained using the Adam optimization algorithm. The species classification model takes the fish image region cropped within the target bounding box as input, outputs the probability distribution of each species category from the fully connected layer, uses cross-entropy loss as the loss function, and is trained using the Adam optimization algorithm.

[0029] In this embodiment of the application, in order to improve the recognition accuracy of overlapping targets in dense fish school scenes, when performing target detection on effective video frame sequences, the target trajectories between adjacent frames are also associated, and the bounding box sequences of the same fish target in consecutive frames are associated as target trajectories. Based on the target trajectory, the missed detection caused by single frame occlusion is filled in, thereby reducing the missed detection rate in dense fish school scenes.

[0030] For each fish target in the visual recognition results, the parallax information between the synchronized frames of the left and right cameras acquired by stereo-BRUV is used to calculate the three-dimensional spatial coordinates of the fish target, and then the actual body length of the fish target in three-dimensional space is calculated. Based on the allometric growth relationship between body length and body weight, the corresponding body weight is calculated, and individual morphological data containing target number, species tag, body length, and body weight are generated. Allometric growth refers to a power-law growth relationship between two traits of an organism that is not directly proportional, as opposed to isotropic growth: when body length doubles, body weight exactly doubles (y=kx linear). Fish do not exhibit isotropic growth; when a fish grows to twice its original length, its volume and weight will far exceed twice its original size. Therefore, an allometric model must be used for fitting.

[0031] It should be noted that the allometric growth relationship between body length and body weight is expressed by the formula... It means that among them This is an estimated weight value. This is a measurement of body length. and The species-related allometric growth coefficient, and The values ​​are derived from the species parameters pre-stored in the fish species database of the protected area. Specifically, for the target fish species, regression fitting is performed in advance using a large number of measured samples (body length and weight measured during catch) to obtain the species-specific parameters. and The coefficients are written into the algorithm model, and the weight is automatically calculated after the body length is measured by binoculars.

[0032] In this embodiment of the application, in order to reduce the impact of water refraction on the accuracy of three-dimensional coordinate calculation, when using parallax information to calculate three-dimensional spatial coordinates, a water refractive index correction coefficient is introduced to correct the parallax value. After correction, the three-dimensional coordinate calculation is performed, thereby improving the accuracy of body length estimation.

[0033] The species labels and quantity statistics from the visual recognition results, the estimated fish population and distribution range resolved from the effective acoustic echo sequence, and the species presence information from the standardized species detection result list are fused from multiple sources to generate fused population statistics for each monitoring point. The fused population statistics include the quantity estimate, confidence level, and spatial distribution information of each species.

[0034] It should be noted that the multi-source fusion adopts a weighted voting method: for the same species at the same location, the statistical count of visual recognition results, the estimated count of effective acoustic echo sequence resolution, and the existence information of eDNA detection are each assigned a corresponding weight, which is determined based on the statistical accuracy of each technical means in the historical data of that location; when the results of the three sources are consistent, the confidence level of the corresponding species quantity estimate is increased; when there is a conflict, the source with higher confidence level is given priority, and the conflict marker is recorded for subsequent verification.

[0035] In this embodiment of the application, in order to cover rare and endangered fish species that are difficult to detect effectively by both visual and acoustic monitoring, species with confidence levels exceeding a preset threshold in the eDNA detection results are retained separately during the fusion process. Even if the species is not detected in the visual recognition results and effective acoustic echo sequences, the species is included in the fused population statistics and marked as an eDNA-specific detected species, thereby ensuring the integrity of the monitoring coverage of rare species.

[0036] The fusion population statistics and corresponding environmental parameter time series of each monitoring point are input into the habitat suitability assessment model, and the habitat suitability assessment results of each monitoring point are output. The habitat suitability assessment results include the habitat suitability level and the identifiers of key environmental parameters affecting habitat quality.

[0037] It should be noted that the habitat suitability assessment model adopts a machine learning classification model. It takes the time series statistical characteristics of environmental parameters at each monitoring point (including the mean, variance and periodic variation of each parameter) and the species diversity index and population density from the integrated population statistics as input features, the habitat suitability level label as output, and the cross-entropy loss as the loss function. The Adam optimization algorithm is used to complete supervised training on the historical monitoring data of the protected area and the habitat suitability samples labeled by experts.

[0038] In this embodiment of the application, in order to identify the environmental factors that have the most significant impact on habitat suitability, the output of the habitat suitability assessment model is simultaneously output, ranking the contribution of each input environmental parameter to the habitat suitability assessment result. The parameters with the highest contribution ranking are marked as key environmental parameters and included in the habitat suitability assessment result, providing parameter-level support for habitat restoration decisions.

[0039] By comprehensively analyzing the time series changes in the number of each species in the integrated population statistics, the abnormal deviations of each parameter in the time series of environmental parameters, and the results of habitat suitability assessment, ecological risk events such as abnormal water quality, abnormal fish mortality, abnormal fish aggregation, and illegal fishing are identified. The identified risk events are classified according to risk level, and early warning information is generated.

[0040] Step 7 may specifically include: Step 701: For key parameters such as dissolved oxygen and ammonia nitrogen in the environmental parameter time series, calculate the deviation of the current value of each key parameter from the historical baseline mean. When the deviation exceeds the preset abnormal threshold, it is identified as a water quality abnormality risk event, and the abnormal parameter name, deviation amount and corresponding location information are recorded.

[0041] Step 702: For the population time series of each species in the fused population statistics, calculate the population change rate within adjacent time windows. When the absolute value of the population change rate exceeds the preset change rate threshold and there is no significant deterioration in the habitat suitability assessment results, it is identified as a risk event of abnormal fish mortality or abnormal fish aggregation.

[0042] Step 703: For the human activity image areas around the water detected in the effective video frame sequence, identify behaviors such as illegal fishing and unauthorized entry through an image classification model. The image classification model takes the human activity image area as input, outputs the probability distribution of behavior categories from the fully connected layer as output, uses cross-entropy loss as the loss function, and completes supervised training on labeled behavior samples using the Adam optimization algorithm; generate illegal behavior identification results, and record the behavior type, time of occurrence, and location information.

[0043] Step 704: Based on the various risk events identified in steps 701 to 703, assess the risk level according to the risk type and degree of deviation. The risk level is divided into three levels: general, major, and serious. Generate early warning information that includes the risk type, location of occurrence, risk level, and handling suggestions.

[0044] In this embodiment of the application, in order to reduce the false alarm rate caused by abnormal fluctuations at a single moment, during the abnormal identification process in steps 701 and 702, it is required that abnormal deviations in parameters or abnormal changes in quantity persist for multiple consecutive collection moments, and risk event identification is triggered only after a preset duration threshold is reached, thereby filtering out false alarms caused by short-term fluctuations.

[0045] Early warning information is pushed to relevant management and law enforcement personnel via SMS, system messages, etc.; population statistics, individual morphological data, environmental parameter time series, habitat suitability assessment results and early warning information are stored in the database according to location and time dimension; based on the stored data, monitoring reports containing species quantity statistics, population density distribution, body length and weight distribution, habitat suitability change trends and early warning event summaries are generated on a daily, weekly, monthly and yearly basis.

[0046] On the other hand, the present invention also proposes a fish ecological monitoring system for protected areas, comprising: The data acquisition module is used to collect underwater video data, acoustic echo data, eDNA detection data, and aquatic environmental parameter data from the underwater camera unit, acoustic detection unit, eDNA monitoring unit, and multi-parameter water quality monitor, and generate multi-source raw monitoring datasets. The data preprocessing module is used to clean and standardize various types of data in the multi-source raw monitoring dataset to generate effective video frame sequences, effective acoustic echo sequences, standardized species detection result lists, and continuous environmental parameter time series. The visual recognition module is used to input valid video frame sequences into the fish target detection model and species classification model to generate visual recognition results; The morphology estimation module is used to estimate the body length and weight of fish targets in the visual recognition results based on stereo-BRUV binocular parallax information, and generate individual morphology data. The multi-source fusion module is used to fuse visual recognition results, effective acoustic echo sequences, and a standardized list of species detection results to generate fused population statistics for each monitoring point. The habitat assessment module is used to input the integrated population statistics and environmental parameter time series into the habitat suitability assessment model to generate habitat suitability assessment results for each monitoring point. The risk warning module is used to identify ecological risk events and classify risks based on integrated population statistics, environmental parameter time series and habitat suitability assessment results, and generate warning information. The data management module is used to push early warning information to managers, store various monitoring data and analysis results in the database, and generate monitoring reports periodically.

[0047] This embodiment overcomes the limitations of single monitoring methods in terms of coverage, detection depth, and rare species detection capabilities by fusing data from three sensing technologies: underwater video, acoustic detection, and eDNA. This allows acoustic detection to supplement fish populations in visual blind spots, and eDNA detection to cover rare species that are difficult to detect visually and acoustically, thereby improving the completeness and accuracy of population statistics. A binocular parallax-based method for estimating body length and weight replaces contact measurement with three-dimensional spatial coordinate calculations, enabling non-contact acquisition of individual fish morphology. The habitat suitability assessment model uses environmental parameters and population data as input, ensuring that the assessment results reflect the actual impact of environmental factors on fish habitats. Identified key environmental parameters can directly guide habitat restoration efforts. A duration threshold is introduced in the risk warning step to prevent short-term fluctuations from triggering warnings, reducing false alarm rates and ensuring the effectiveness of warning information. Overall, the synergistic effect of these steps transforms fish ecological monitoring in protected areas from reliance on manual sampling to automated, real-time, and interference-free intelligent monitoring, solving the problems of low monitoring frequency, insufficient data accuracy, and significant interference with fish habitats inherent in traditional monitoring methods.

[0048] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for monitoring the ecological ecology of fish in a protected area, characterized in that, At least the following steps are included: Underwater video data, acoustic echo data, eDNA detection data, and aquatic environmental parameter data are collected from underwater camera units, acoustic detection units, eDNA monitoring units, and multi-parameter water quality monitoring instruments deployed in the protected area to generate a multi-source raw monitoring dataset; The multi-source raw monitoring dataset is cleaned and standardized to generate effective video frame sequences, effective acoustic echo sequences, standardized species detection result lists, and environmental parameter time series. The effective video frame sequence is input into the fish target detection model to generate a set of target bounding boxes. The fish image regions corresponding to each bounding box in the target bounding box set are input into the species classification model to generate visual recognition results. The three-dimensional spatial coordinates of each fish target are calculated using the binocular parallax information of the underwater camera unit, and then the body length is calculated. The body weight is calculated based on the allometric growth relationship between body length and body weight, and individual morphological data is generated. The visual recognition results, effective acoustic echo sequences, and standardized species detection result lists are fused from multiple sources to generate fused population statistics for each monitoring point. By inputting population statistics and environmental parameter time series data into the habitat suitability assessment model, a habitat suitability assessment result containing habitat suitability level and key environmental parameter identifiers is generated. Ecological risk events are identified by integrating population statistics, environmental parameter time series, and habitat suitability assessment results. These risk events are then classified, early warning information is generated, and the early warning information is pushed to management personnel.

2. The method for monitoring fish ecology in a protected area according to claim 1, characterized in that, The multi-source fusion adopts a weighted voting method. For the same species at the same monitoring point, the statistical count of visual recognition results, the estimated count of effective acoustic echo sequence analysis, and the existence information of eDNA detection are assigned corresponding weights. The weights are determined based on the statistical accuracy of each technical means in the historical data of that point. When the results of the three sources are consistent, the confidence level of the corresponding species quantity estimate is increased. When there is a conflict, the source with higher confidence level is given priority, and the conflict is marked. Among them, species with a confidence level of more than a preset threshold in the eDNA detection results are retained separately, and the species are included in the fused population statistics and marked as eDNA-specific detected species.

3. The method for monitoring fish ecology in a protected area according to claim 1, characterized in that, The allometric growth relationship between body length and body weight is expressed by the formula... It means that among them This is an estimated weight value. This is a measurement of body length. and The species-related allometric growth coefficient, and The values ​​are derived from the pre-stored species parameters in the fish species database of the protected area; when using parallax information to calculate the three-dimensional spatial coordinates, a water refractive index correction coefficient is introduced to correct the parallax value, and then the three-dimensional coordinates are calculated.

4. The method for monitoring fish ecology in a protected area according to claim 1, characterized in that, The identified ecological risk events include: The deviation of the current value of each key parameter in the time series of environmental parameters from the historical baseline mean is calculated. When the deviation exceeds the preset anomaly threshold and the duration exceeds the preset duration threshold, it is identified as a water quality anomaly risk event. Calculate the rate of change of the population of each species in the time series of the integrated population statistics within adjacent time windows. When the absolute value of the rate of change exceeds the preset rate of change threshold and there is no significant deterioration in the habitat suitability assessment results, and the duration exceeds the preset duration threshold, it is identified as a risk event of abnormal fish mortality or abnormal fish aggregation. For the human activity image areas around the water detected in the effective video frame sequence, illegal fishing and unauthorized entry behaviors are identified through image classification models, and illegal behavior identification results are generated.

5. A method for monitoring fish ecology in a protected area according to claim 4, characterized in that, Based on the identification results of comprehensive water quality abnormality risk events, abnormal fish mortality or abnormal fish gathering risk events, and illegal activities, the risk level is assessed into three levels: general, major, and serious, according to the risk type and degree of deviation. Early warning information is generated, which includes risk type, location of occurrence, risk level, and disposal recommendations.

6. The method for monitoring fish ecology in a protected area according to claim 1, characterized in that, The habitat suitability assessment model takes the statistical characteristics of environmental parameter time series and the species diversity index and population density from population statistics as input features, and habitat suitability level labels as output. It completes supervised training on historical monitoring data of the protected area and habitat suitability samples labeled by experts. The habitat suitability assessment model simultaneously outputs the contribution ranking of each input environmental parameter to the habitat suitability assessment results, and marks the parameters with the highest contribution as key environmental parameters and incorporates them into the habitat suitability assessment results.

7. The method for monitoring fish ecology in a protected area according to claim 1, characterized in that, When performing target detection on a valid video frame sequence, the target trajectories between adjacent frames are also associated. The bounding box sequence of the same fish target in consecutive frames is associated as the target trajectory, and the missed detection caused by single-frame occlusion is filled in based on the target trajectory.

8. A method for monitoring fish ecology in a protected area according to claim 1, characterized in that, When cleaning underwater video data, the frame extraction frequency is dynamically adjusted according to the motion amplitude between video frames: when the motion amplitude between frames is lower than a preset threshold, the extraction frequency is reduced; when the motion amplitude between frames exceeds the preset threshold, the extraction frequency is increased.

9. A fish ecological monitoring system for a protected area, characterized in that, include: The data acquisition module is used to collect underwater video data, acoustic echo data, eDNA detection data, and aquatic environmental parameter data from the underwater camera unit, acoustic detection unit, eDNA monitoring unit, and multi-parameter water quality monitor, and generate multi-source raw monitoring datasets. The data preprocessing module is used to clean and standardize the multi-source raw monitoring datasets, generating effective video frame sequences, effective acoustic echo sequences, standardized species detection result lists, and environmental parameter time series. The visual recognition module is used to input the effective video frame sequence into the fish target detection model to generate a set of target bounding boxes, and input the fish image regions corresponding to each bounding box in the target bounding box set into the species classification model to generate visual recognition results. The morphology estimation module is used to calculate the three-dimensional spatial coordinates of each fish target using the binocular parallax information of the underwater camera unit, and then calculate the body length and weight based on the allometric growth relationship to generate individual morphology data. The multi-source fusion module is used to fuse visual recognition results, effective acoustic echo sequences, and a standardized list of species detection results to generate fused population statistics for each monitoring point. The habitat assessment module is used to input integrated population statistics and environmental parameter time series into the habitat suitability assessment model to generate habitat suitability assessment results that include habitat suitability level and key environmental parameter identifiers. The risk warning module is used to identify ecological risk events based on integrated population statistics, environmental parameter time series and habitat suitability assessment results, classify risk events, and generate warning information. The data management module is used to push early warning information to managers, store various monitoring data and analysis results in the database, and generate monitoring reports periodically.

10. The protected area fish ecological monitoring system according to claim 9, characterized in that, The multi-source fusion module uses a weighted voting method to fuse the same species at the same monitoring point. The weights are determined based on the statistical accuracy of each technology in the historical data of that point. Species with a confidence level exceeding a preset threshold in the eDNA detection results are retained separately, included in the fused population statistics, and marked as eDNA-specific detected species. The risk warning module requires that abnormal deviations in parameters or abnormal changes in quantity persist for multiple consecutive collection times, and only triggers risk event identification after reaching a preset duration threshold.

Images