Intelligent identification and density monitoring method and device for intertidal zone beach benthic animals

Abstract

The present application provides a kind of intertidal zone beach benthic animal intelligent identification and density monitoring method and device, in the target area of investigation light beach, multiple independent sample quadrats are laid according to random way, video is collected using unmanned aerial vehicle and is pretreated;In uniform video environment, the identification and spatial scale calibration of sample quadrat are carried out synchronously, and the target identification of beach benthic animal is carried out;The actual beach area corresponding to the video picture is determined according to the sample quadrat identification result, and the representative quantity and biological density of the species are calculated based on the identification result of target species using median statistical method.The non-contact, automatic and high-precision monitoring of intertidal zone beach benthic animal is realized through the cooperative application of unmanned aerial vehicle aerial photography and improved YOLO identification model, the interference of artificial investigation is reduced, the stability, accuracy and monitoring efficiency of density estimation are improved, and it is suitable for large-scale, high-frequency investigation and monitoring of intertidal zone beach benthic animal resources.

Description

Technical Field

[0001] This invention relates to the field of ecological environment monitoring technology, and in particular to a method and device for intelligent identification and density monitoring of benthic animals in intertidal mudflats. Background Technology

[0002] Currently, intertidal mudflat ecosystems are crucial interfaces for land-sea interaction, with large benthic animals such as fiddler crabs and mudskippers playing key roles in material cycling and energy flow. Traditional monitoring methods primarily rely on manual observation and sampling, which is not only time-consuming, labor-intensive, and prone to interference, but also often leads to significant statistical errors due to the agile movements of the target organisms. Furthermore, existing fixed camera or trap-based monitoring methods are significantly affected by tides, sunlight, and equipment placement, making continuous and stable dynamic monitoring difficult. Simultaneously, calculations of biological quantity or density based on image pixels lack reliable spatial scale calibration, failing to accurately reflect the actual mudflat area and resulting in monitoring results lacking spatial accuracy.

[0003] In recent years, with the development of UAV (Unmanned Aerial Vehicle) aerial photography technology, ecological remote sensing monitoring has gradually shifted towards high resolution and low interference. However, the massive amount of video data and the large differences in target scale make it difficult for traditional image recognition algorithms (such as threshold-based or edge-detection-based methods) to effectively distinguish small benthic animals. The YOLO (You Only Look Once, a real-time target detection method based on deep learning) series of algorithms has been widely used in the field of target detection due to its high detection accuracy and real-time performance, but its application in tidal flat ecological monitoring still faces challenges such as complex lighting, cluttered backgrounds, and similar organism morphologies. Simultaneously, it is necessary to solve the problem of synchronous spatial scale calibration to achieve accurate conversion between image pixel scale and actual tidal flat area, ensuring the accuracy and reliability of quantity and density statistics. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and device for intelligent identification and density monitoring of benthic animals in intertidal mudflats, and to propose an improved UAV-based integrated monitoring system that combines the YOLO recognition model with the dynamic environment of the intertidal zone.

[0005] In a first aspect, the present invention provides an intelligent identification and density monitoring method for benthic animals in intertidal mudflats. The method includes: randomly or quasi-randomly deploying multiple independent quadrats in a target area of ​​the surveyed mudflat, and collecting video of each quadrat; wherein each quadrat is dropped horizontally, and the distance between each quadrat is greater than a preset distance threshold; preprocessing the video to construct a unified video processing environment, performing target identification processing in the video processing environment, and performing spatial scale calibration based on the quadrat reference information in the target identification results; determining the representative number of target species in the fixed area using median statistics based on the number of target species in each video frame within the fixed area, and calculating the biological density of target species in the fixed area in combination with the actual physical coverage area of ​​the video image.

[0006] In an optional embodiment of this application, the above-mentioned target recognition processing includes: inputting video frames into a pre-trained recognition model and outputting detection results; wherein, the recognition model is obtained by structural optimization based on the YOLO11 model for the intertidal ecological monitoring task; the detection results include: the bounding box, category, and confidence of the target; the target includes: intertidal benthic animals and fixed reference objects for spatial scale calibration; the intertidal benthic animals include: fiddler crabs and mudskippers, and the fixed reference objects include: square quadrats; spatial scale calibration includes: based on the number of pixels occupied by the square quadrats detected by the recognition model in the video frame, combined with the known real area of ​​the square quadrats, establishing a correspondence between the image pixel scale and the actual mudflat area, and calculating the actual physical coverage area of ​​the video frame as a fixed region; In an optional embodiment of this application, the step of collecting video of each quadrat includes: maneuvering the drone to fly directly above each quadrat and adjusting the gimbal shooting angle to be perpendicular to the quadrat; adjusting the video parameters and collecting video of each quadrat; and obtaining the environmental parameters when collecting video of each quadrat.

[0007] In optional embodiments of this application, the above-described preprocessing steps for the video include: extracting image frames from the video; enhancing the biological and quadrat contours of the image frames through histogram equalization and background modeling; and removing noisy regions of the image frames through threshold segmentation and edge detection.

[0008] In optional embodiments of this application, the structure of the above-mentioned recognition model includes: a backbone network, a feature enhancement module, and a detection head; the mechanism of the recognition model includes: cross-stage feature reuse, spatial pyramid pooling, and parallel attention mechanism; the recognition model outputs feature maps of the original input image in the following manner: the recognition model halves the spatial resolution of the original input image through a convolutional layer to extract the primary features of the original input image; the recognition model performs multiple downsampling operations on the primary features, and the multiple downsampling operations output multiple feature maps of different scales respectively; wherein, each downsampling operation halves the resolution of the input feature map and increases the number of channels of the input feature map.

[0009] In optional embodiments of this application, the feature map includes: shallow features and deep features; the shallow features contain details and edge information for identifying targets at smaller scales; the deep features contain global semantics and contextual information for identifying targets at larger scales or objects in complex backgrounds.

[0010] In an optional embodiment of this application, the total loss function of the above-mentioned recognition model is determined based on the intersection-union ratio loss, classification loss, and distributed bounding box loss.

[0011] In an optional embodiment of this application, the step of establishing a correspondence between the image pixel scale and the actual tidal flat area based on the number of pixels occupied by the square quadrat detected by the recognition model in the video frame and the known true area of ​​the square quadrat includes: obtaining the bounding box information of the square quadrat output by the recognition model; calculating the number of pixels occupied by the quadrat in the corresponding video frame based on the bounding box information; calculating the actual tidal flat area corresponding to a single pixel based on the true area of ​​the reference object and the corresponding number of pixels; calculating the actual tidal flat area corresponding to the entire image based on the pixel area conversion factor; and establishing a correspondence between the image pixel scale and the actual tidal flat area based on the actual tidal flat area corresponding to a single pixel and the actual tidal flat area corresponding to the entire image.

[0012] In optional embodiments of this application, after the steps of inputting video frames into a pre-trained recognition model and outputting detection results, the method further includes: filtering the detection results of each frame in the video frame and recording the number of valid biological targets (fiddler crabs and mudskippers); using a cumulative calculation method to determine the representative number (i.e., the median number) of benthic animals in a fixed area based on the number of valid biological targets; using a density median calculation method to determine the biological density of targets in a fixed area based on the number of valid biological targets and the actual physical area of ​​the fixed area obtained by spatial scale calibration; and using an average calculation method to calculate the biological density of species in the target area based on the biological density of targets in each fixed area.

[0013] Secondly, embodiments of the present invention also provide an intelligent identification and density monitoring device for benthic animals in intertidal mudflats. The device includes: a quadrat deployment and aerial data acquisition module, used to deploy multiple independent quadrats in a random or quasi-random manner in the target area of ​​the surveyed mudflat, and acquire video of each quadrat; wherein each quadrat is dropped horizontally, and the distance between each quadrat is greater than a preset distance threshold; a target identification processing and spatial scale calibration module, used to preprocess the video, construct a unified video processing environment, perform target identification processing in the video processing environment, and perform spatial scale calibration based on the quadrat reference information in the target identification results; the target identification processing and spatial scale calibration module includes: a density calculation submodule, used to determine the representative number of targets in the fixed area based on the number of benthic animal target species in each video frame in the fixed area using the median statistical method, and calculate the biological density of target species in the fixed area in combination with the actual mudflat area obtained by spatial scale calibration.

[0014] The embodiments of the present invention bring the following beneficial effects: This invention provides an intelligent identification and density monitoring method and device for intertidal benthic animals. It collects videos of intertidal mudflat quadrats via drone aerial photography and combines them with an identification model optimized for intertidal ecological monitoring tasks based on the YOLO11 model. This enables simultaneous automatic identification and classification of mudflat benthic animals (fiddler crabs and mudskippers) and spatial scale references (square quadrats). By introducing a spatial scale calibration method based on model identification results, it directly utilizes the pixel proportion of the identified quadrats in the image and their known true area to uniformly calibrate the scale of the entire video image, automatically converting the detection results from image pixel scale to the actual mudflat area. This method, without relying on camera pose parameters, can adapt to subtle fluctuations in shooting height and angle, providing a reliable physical spatial basis for accurate calculation of biological quantity and density. It employs median statistics to perform time-cumulative statistical analysis of the number of effective biological targets detected in the quadrat video sequence. Compared to traditional mean-based methods, this significantly reduces density estimation fluctuations caused by animals rapidly entering and leaving the field of view, mutual occlusion, and environmental changes, improving the stability and reliability of density statistics. This method reduces human interference, enabling non-contact automatic monitoring of intertidal benthic animals, improving the repeatability and applicability of monitoring results, and is suitable for long-term ecological monitoring and large-scale surveys. Compared with existing technologies, this invention achieves high-precision, stable, and automated monitoring of intertidal benthic animals, significantly improving monitoring efficiency and data reliability.

[0015] Other features and advantages of this disclosure will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0016] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an intelligent identification and density monitoring method for intertidal benthic animals provided in an embodiment of the present invention; Figure 2 A top view of a target area with multiple quadrats arranged in an embodiment of the present invention; Figure 3 A side view of a target area with multiple quadrats arranged in an embodiment of the present invention; Figure 4 A schematic diagram of a video frame provided in an embodiment of the present invention; Figure 5 A schematic diagram of a feature pyramid architecture provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a convolutional layer structure provided in an embodiment of the present invention; Figure 7 A schematic diagram of the architecture of an SPPF module provided in an embodiment of the present invention; Figure 8 A schematic diagram of the architecture of a C2PSA module provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of an output recognition result video frame provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of an intelligent identification and density monitoring method for intertidal benthic animals provided in an embodiment of the present invention; Figure 11 A schematic diagram of the structure of an intelligent identification and density monitoring device for intertidal benthic animals provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Currently, intertidal mudflat ecosystems are crucial interfaces for land-sea interaction, with large benthic animals such as fiddler crabs and mudskippers playing key roles in material cycling and energy flow. Traditional monitoring methods primarily rely on manual observation and sampling, which is not only time-consuming, labor-intensive, and prone to interference, but also often leads to significant statistical errors due to the agile movements of the target organisms. Furthermore, existing fixed camera or trap-based monitoring methods are significantly affected by tides, sunlight, and equipment placement, making continuous and stable dynamic monitoring difficult. Simultaneously, calculations of biological quantity or density based on image pixels lack reliable spatial scale calibration, failing to accurately reflect the actual mudflat area and resulting in monitoring results lacking spatial accuracy.

[0021] In recent years, with the development of drone aerial photography technology, ecological remote sensing monitoring has gradually shifted towards high resolution and low interference. However, the massive amount of video data and the large differences in target scale make it difficult for traditional image recognition algorithms (such as threshold-based or edge-detection-based methods) to effectively distinguish small benthic animals. The YOLO series of algorithms, due to their high detection accuracy and real-time performance, has been widely used in target detection, but its application in tidal flat ecological monitoring still faces challenges such as complex lighting, cluttered backgrounds, and similar organism morphologies. Simultaneously, the problem of synchronous spatial scale calibration needs to be solved to achieve accurate conversion between image pixel scale and actual tidal flat area, ensuring the accuracy and reliability of quantity and density statistics.

[0022] Based on this, the present invention provides a method and device for intelligent identification and density monitoring of intertidal benthic animals, which can be applied to the fields of ecological environment monitoring and artificial intelligence identification. Specifically, it relates to an intelligent monitoring method that combines drone aerial photography and deep learning algorithm (YOLO) to achieve automatic identification and density calculation of intertidal benthic animals (including fiddler crabs and mudskippers).

[0023] To facilitate understanding of this embodiment, a detailed description of the intelligent identification and density monitoring method for intertidal benthic animals disclosed in this embodiment of the invention will be provided first.

[0024] Example 1: This invention provides a method for intelligent identification and density monitoring of benthic animals in intertidal mudflats. (See also...) Figure 1The flowchart shown illustrates a method for intelligent identification and density monitoring of intertidal benthic animals. This method includes the following steps: Step S102: In the target area of ​​the survey beach, multiple independent quadrats are set up in a random or quasi-random manner, and video of each quadrat is collected.

[0025] In this scenario, each quadrat is dropped horizontally, and the distance between each quadrat is greater than a preset distance threshold.

[0026] In this embodiment, quadrat layout and aerial data collection can be performed first. See also Figure 2 The diagram shows a top view of a target area with multiple quadrats arranged thereon. Figure 3 The diagram shows a side view of a target area where multiple quadrats are arranged. The target area for the survey (e.g., the low tide zone) is selected, and multiple quadrats are thrown in multiple directions in a random or quasi-random manner, ensuring that each quadrat falls horizontally and is spaced at least 10 meters apart. This reduces the spatial correlation of individual activity migration with the statistical data; the distance threshold can be set to 10 meters. This embodiment uses five quadrats as an example for illustration, and will not be repeated hereafter.

[0027] In some embodiments, the drone can be controlled to fly directly above each quadrat, and the gimbal angle can be adjusted to be perpendicular to the quadrat; video parameters can be adjusted to capture video of each quadrat; and environmental parameters can be obtained when capturing video of each quadrat.

[0028] In this embodiment, the drone can be slowly flown to a position 4-5 meters directly above the sample plot, and the gimbal angle can be adjusted to be as perpendicular to the sample plot as possible. The video parameters should be set to 4K resolution and 5-7 minutes in length to ensure the sample plot and organisms in the image are clearly visible. (See [reference needed]). Figure 4 The diagram shows a typical video feed. The same operations were performed simultaneously in all other quadrats. Additionally, environmental parameters such as temperature, wind speed, and light intensity were recorded concurrently during the video capture process.

[0029] Step S104: Preprocess the video, construct a unified video processing environment, perform target recognition processing in the video processing environment, and perform spatial scale calibration based on the sample reference information in the target recognition results.

[0030] After acquiring videos of multiple sample plots, this embodiment can preprocess the videos.

[0031] In some embodiments, image frames from a video can be extracted; the biological contours of the image frames can be enhanced by histogram equalization and background modeling; and noisy regions of the image frames can be removed by thresholding and edge detection.

[0032] This embodiment extracts image frames from the captured video at 25 frames per second, and enhances the biological contours through histogram equalization and background modeling (e.g., MOG2 (Mixture of Gaussians 2, a background subtraction algorithm based on Gaussian mixture models)). Noisy regions are removed through thresholding and edge detection, generating an image set suitable for recognition.

[0033] (1) Target recognition processing, including: inputting video frames into a pre-trained recognition model and outputting detection results; wherein, the recognition model is obtained by structural optimization based on the YOLO11 model for the intertidal ecological monitoring task; the detection results include: bounding boxes, categories and confidence scores of various targets; the targets include not only intertidal benthic animals (fiddler crabs and mudskippers), but also fixed reference objects (square quadrats) used for spatial scale calibration. In this embodiment, structural optimization can be performed on the YOLO11 model for intertidal ecological monitoring tasks to construct an identification model.

[0034] In some embodiments, the structure of the above-mentioned recognition model includes: a backbone network, a feature enhancement module, and a detection head; the mechanism of the recognition model includes: cross-stage feature reuse, spatial pyramid pooling, and parallel attention mechanism.

[0035] The overall structure of YOLO11 is mainly divided into three parts: the backbone network, the feature enhancement module, and the detection head. The model design incorporates cross-stage feature reuse, fast spatial pyramid pooling, and parallel attention mechanisms, which allows it to maintain good stability and accuracy when facing complex backgrounds and multi-scale targets.

[0036] This embodiment can be combined with detection scenarios involving fiddler crabs, mudskippers, and spatial scale references (square quadrats). YOLO11 can not only accurately draw bounding boxes for various targets and identify their categories, but also automatically analyze the actual physical coverage of the video footage based on the identified quadrat pixel information, thereby helping to achieve accurate statistics on the actual mudflat area, number of individuals, and density. This structure ensures detection speed while maintaining high detection performance under lightweight conditions, solving the problem of traditional visual monitoring lacking spatial scale benchmarks. It is very suitable for ecological monitoring and density analysis in fixed locations and dynamic environments.

[0037] In some embodiments, the recognition model outputs the feature map of the original input image in the following manner: the recognition model halves the spatial resolution of the original input image through a convolutional layer to extract the primary features of the original input image; the recognition model performs multiple downsampling operations on the primary features, and the multiple downsampling operations output multiple feature maps of different scales respectively; wherein, each downsampling operation halves the resolution of the input feature map and increases the number of channels of the input feature map.

[0038] The feature map includes shallow features and deep features. Shallow features contain details and edge information and are used to identify targets at smaller scales. Deep features contain global semantics and contextual information and are used to identify targets at larger scales or objects in complex backgrounds.

[0039] During the input phase, assuming the original input image has dimensions H (height) × W (width), it first passes through a convolutional layer with a stride of 2 to halve the spatial resolution and extract primary features. Subsequently, the network performs multiple downsampling operations, each reducing the feature map resolution by half while increasing the number of channels to enhance feature representation. Specifically, the model outputs five feature maps at different scales: P1 / 2, P2 / 4, P3 / 8, P4 / 16, and P5 / 32, where the subscripts represent the scaling ratio of the feature map relative to the original image. For example, P1 / 2 represents a size half that of the original image, with a resolution of approximately (H / 2) × (W / 2); P2 / 4 corresponds to (H / 4) × (W / 4), and so on, until P5 / 32 is (H / 32) × (W / 32). ; in,[ The symbol ] indicates a round-down operation. Indicates feature level, This represents the downsampling factor for the corresponding scale. This hierarchical structure ensures that the model can simultaneously capture semantic information at different scales: shallow features (such as P1 / 2, P2 / 4) contain more details and edge information, suitable for detecting small targets; while deep features (such as P4 / 16, P5 / 32) contain more global semantic and contextual information, helping to identify large targets or objects in complex backgrounds. Through this multi-scale feature extraction mechanism, the model achieves effective fusion of spatial and semantic information at different resolutions, providing a solid feature foundation for subsequent feature enhancement and detection heads. With this resolution pyramid, the model can simultaneously consider individuals of fiddler crabs or mudskippers at different scales. The feature pyramid architecture can be found in [reference needed]. Figure 5 The diagram shows a characteristic pyramid structure.

[0040] In the initial stage of the model, the backbone network extracts features from the input image through a series of basic convolutional units. The essence of the convolution operation is to utilize features of a size of... k × k The convolutional kernel slides across the input feature map, capturing spatial structure information through local weighted summation, thereby extracting low-level features such as edges and textures. ; in, It refers to the number of model parameters. The kernel size is [size]. , These represent the number of input and output channels, respectively. To improve the efficiency of feature reuse and gradient propagation, the C3k2 module is introduced. The core idea of ​​this module is to achieve diversified feature aggregation through cross-stage residual connections and dual-branch convolutional structures, enabling the network to maintain a lightweight design while still possessing powerful representational capabilities. Its computational form is as follows: ; In this formula, Indicates the input feature map, and These represent two different feature extraction paths. Features are linked through concatenation, with each path consisting of stacked 1×1 and 3×3 convolutional layers. The 1×1 convolutions are primarily used for channel compression and non-linear mapping, reducing the number of parameters and improving computational efficiency; while the 3×3 convolutions capture a wider range of spatial context information, extracting rich local structural features. The results extracted from the two paths are concatenated along the channel dimension, and then... The function performs channel integration. Implemented using 1×1 convolutions, this process achieves information fusion and channel number matching. Finally, the integrated features are element-wise added to the original input x, thus enabling cross-stage information transfer and efficient gradient backflow. For the convolutional layer structure, please refer to [link to documentation]. Figure 6 The diagram shows a convolutional layer structure.

[0041] In the high-level features of the backbone network, YOLO11 employs the Spatial Pyramid Pooling Fast (SPPF) module to expand the receptive field, the process of which can be represented as follows: ; in, Indicates the input feature map; p ( () indicates a max pooling operation; in this embodiment, 5×5 cores are used for multiple pooling operations; This indicates that the pooling outputs of each layer are spliced ​​along the channel dimension; g ( ) represents a 1×1 convolution operation, used to fuse spliced ​​multi-scale features and compress channel dimensions.

[0042] This embodiment introduces the SPPF module, which can be found in [reference needed]. Figure 7The diagram shows an architecture diagram of the SPPF module. Following SPPF, the model introduces a C2PSA module (Cross-Channel Parallel Self-Attention module, a feature enhancement module primarily used to improve the model's multi-scale feature extraction and fusion capabilities), the structure of which can be found in [reference needed]. Figure 8 The diagram shows an architecture of a C2PSA module.

[0043] The core of the C2PSA module is a parallel self-attention mechanism. For input features... The standard self-attention calculation method is as follows: ; in, Indicates input features The output features obtained after performing self-attention modeling; Q , K , V These respectively represent the input features The query matrix, key matrix, and value matrix obtained through linear transformation; , and This is the corresponding trainable parameter matrix; This is a scaling factor used to normalize the inner product result when calculating attention weights, in order to suppress problems such as excessively large values ​​or unstable gradients in high-dimensional feature spaces.

[0044] The C2PSA module adopts a parallel modeling strategy of channel attention and spatial attention in its structure: on the one hand, the channel branch automatically evaluates the importance of different feature channels in target recognition by weighted learning of the features of each channel, thereby highlighting the semantic dimension that contributes more to detection; on the other hand, the spatial branch captures the global dependencies between different regions in the image through adaptive spatial attention mapping, enabling the model to identify the structural differences between targets.

[0045] In the neck region, YOLO11 employs a dual-mode fusion strategy of top-down and bottom-up approaches. Let... Given the features of P5, P4, and P3 respectively, the fusion process can be represented as follows: ; in, This represents P4 features that incorporate high-level semantics; This indicates that the P3 features, which incorporate high-level semantics, are used for small target detection. This represents the P4 features after bottom-up fusion, which integrates low-level details and high-level semantics; , These represent 2x upsampling and downsampling, respectively. Indicates feature splicing, This is the C3k2 feature fusion module. This process ensures detail fidelity in small object detection and context enhancement in large object detection.

[0046] The detector head outputs class probabilities and bounding box regression parameters at three scales: P3 / 8, P4 / 16, and P5 / 32. When using an anchor-free decoding method, for the... l Layer-scale feature maps are located at grid positions The prediction results above, and the corresponding target bounding box are represented as follows: ; in, and These represent the horizontal and vertical grid indices of the predicted target on the feature map, respectively. , , and The bounding box regression offset parameters are output by the detection head; The sigmoid function maps the model's output values ​​to between 0 and 1, thus obtaining a probabilistic output result; the exponential function... This is used to map the width and height regression values ​​to positive values, ensuring the geometric validity of the predicted bounding box; Indicates the first The downsampling factor of the layer-scale feature map relative to the input image, in this embodiment... Take 8, 16, and 32 respectively.

[0047] In the detection head of the model, the class probabilities are also activated using the Sigmoid function. This approach allows for independent modeling of the prediction results for each class, unaffected by other classes, enabling the network to handle multi-label scenarios simultaneously. Through this design, the model can not only stably regress the target location in continuous space but also output a smooth and interpretable probability distribution for class discrimination, providing a reliable basis for subsequent target selection and confidence calculation. Here, the formula for obtaining the class probabilities through Sigmoid activation is given: ; Among them, among them, Indicates the detection model for the first The predicted probability of a target class is used to characterize the confidence that the target class exists in the current candidate detection region; (.) represents the Sigmoid activation function, which maps the model's output value to a probability range of 0 to 1; The output of the model detection head represents the first... The original predicted value corresponding to the target class; This represents the total number of target categories that the model can identify.

[0048] In some embodiments, the total loss function of the above recognition model is determined based on the intersection-union ratio loss, classification loss, and distributed bounding box loss.

[0049] During model training, YOLO11 uses a multi-task loss function for joint optimization, which typically includes IoU loss, classification loss, and distributed bounding box loss: ; Among them, among them, This represents the total loss function of the model; This represents the bounding box regression loss based on the intersection-union ratio, used to measure the degree of overlap between the predicted bounding box and the true bounding box; This represents the classification loss based on binary cross-entropy, used to constrain the accuracy of the target category prediction results; This represents the distributed bounding box regression loss, used for fine-grained modeling of bounding box coordinates; , and These represent the weighting coefficients corresponding to each loss term, and in this embodiment, they are all 0.5.

[0050] Through the joint optimization of the multi-task loss function, the model can simultaneously and stably detect intertidal benthic targets and quadrat references, balancing bounding box positioning accuracy and category discrimination reliability in complex scenarios, and providing reliable detection results for subsequent spatial scale calibration and biological density calculation.

[0051] (2) Spatial scale calibration, including: based on the number of pixels occupied by the square quadrat detected by the recognition model in the video frame, combined with the known real area of ​​the square quadrat, establish the correspondence between the image pixel scale and the actual tidal flat area, and calculate the actual physical coverage area of ​​the video frame as a fixed area.

[0052] The "fixed area" refers to the actual geographical coverage area projected by the camera's field of view (FOV) on the mudflat surface when the drone is capturing video at a single sample plot. The actual physical area of ​​this area (i.e., the total coverage area of ​​the video image) is calculated using a spatial scale calibration algorithm based on a fixed reference object (0.2m × 0.2m square sample plot) of known size in the image, and is used as a spatial benchmark for calculating biological density.

[0053] To address the challenge of directly using pixel scale for biomass density calculation under varying shooting heights, this embodiment incorporates quadrats into the detection target of the recognition model, achieving automatic scale calibration based on a fixed reference object. This method uniformly converts the detection results from image pixel space to the actual tidal flat area, automatically correcting scale errors caused by differences in shooting height without relying on camera pose parameters, thus improving the accuracy and comparability of biomass density calculation.

[0054] In some embodiments, the bounding box information of the square sample plots output by the recognition model can be obtained, and the number of pixels occupied by the sample plot in the corresponding video frame can be calculated based on the bounding box information; the actual tidal flat area corresponding to a single pixel can be calculated based on the actual area of ​​the reference object and the corresponding number of pixels; the actual tidal flat area corresponding to the whole image can be calculated based on the pixel area conversion factor; and the correspondence between the image pixel scale and the actual tidal flat area can be established based on the actual tidal flat area corresponding to a single pixel and the actual tidal flat area corresponding to the whole image.

[0055] In this embodiment, a pre-trained recognition model (improved YOLO11) can be used to automatically identify and locate square quadrats in the video frame, and obtain the number of pixels occupied by the quadrats in the current frame image (denoted as ). ); based on the true area of ​​the sample reference (denoted as ); ) and the corresponding number of pixels ( ), calculate the actual tidal flat area corresponding to a single pixel, i.e., the pixel area conversion factor ( k The actual tidal flat area corresponding to the entire image is calculated based on the pixel area conversion factor. ); Based on this, a real-time correspondence between image pixel scale and actual tidal flat area is established, providing a dynamically updated denominator for subsequent calculation of biological density (i.e., biological quantity / actual tidal flat area).

[0056] (3) Density calculation, including: based on the number of target species in each video frame within a fixed area, the representative number of target species within the fixed area is determined by median statistics, and the biological density of target species within the fixed area is calculated by combining the actual physical coverage area of ​​the video screen.

[0057] In some embodiments, the detection results of each frame in the video frame can be filtered to record the number of valid target species; the median of targets in a fixed area can be determined by a cumulative calculation method based on the number of valid target species; the biological density of targets in a fixed area can be determined by a density median calculation method based on the number of valid target species; and the species biological density of the target area can be calculated by a mean calculation method based on the biological density of target species in each fixed area.

[0058] In the process of image recognition and quantitative statistics of mudflat biological density, determining the reference scale is a crucial step in converting detection results from image space to real space. Due to the dynamic differences between the flight attitude of drones and the optical parameters of cameras in actual mudflat scenarios, without a unified scale benchmark, detection results obtained based on image pixels are difficult to directly use for ecological calculations of quantity and density.

[0059] To address the aforementioned problems, this embodiment provides a spatial scale calibration method based on automatic detection using a fixed reference object. This method selects a square quadrat with known physical dimensions as the reference object; the quadrat has a side length of 0.2m and its actual area is... =0.04m 2 .

[0060] During video capture, a target detection algorithm is used to identify the reference object in the sample plot and obtain its position in the [missing information]. t Each time step corresponds to the number of pixels occupied in the image, denoted as . The known true area based on a reference point. Based on the number of pixels corresponding to each pixel, the actual tidal flat area corresponding to a single pixel is calculated to obtain the pixel area conversion factor. The calculation formula is as follows: ; in k ( t ) indicates the first t Pixel area conversion factor for each time step.

[0061] Furthermore, using the pixel area conversion factor, the first... t The actual mudflat area corresponding to the entire video image at each time step is calculated using the following formula: ; in, This represents the total number of pixels in the entire image. For the first t The actual area of ​​the mudflats covered by the video footage at each time step. It should be noted that in this embodiment, the "fixed area" refers to the actual mudflat survey area corresponding to the entire video frame. Due to environmental factors such as wind, the hovering altitude of the drone may change slightly. Meeting time step t Dynamically updated.

[0062] In one embodiment, the video frame rate is set to 25 frames per second, meaning 25 frames of images are contained per second. Using one second as a time step, the target species in each video frame within that time step is analyzed. i The detection results are summarized to obtain the corresponding time steps. t biomass N i( t Based on this, for any target species i Based on the time step from the initial time step to the current time step t Biological population sequence, calculating time steps t The corresponding cumulative median number of organisms The calculation method is as follows: ; In this embodiment, the cumulative median statistical method is introduced for species... i At time step t The quantity estimation no longer relies on the detection results of a single time step, but comprehensively utilizes historical time-series detection data, thereby effectively suppressing the impact of outliers such as false positives and false negatives on the statistical results and improving the stability and reliability of the quantity estimation. As the number of time steps increases, the amount of data involved in the statistics gradually increases; for example, in... t At 40 seconds, a total of 40 time steps (corresponding to 1000 frames of images) are involved in the median calculation. t When the time is 100 seconds, data from 2500 frames of images are included in the statistics.

[0063] Similarly, this embodiment also calculates species based on the actual tidal flat area corresponding to the image. i The biological density was calculated and statistically processed using the cumulative median method. Specifically, in the... t At each time step, the corresponding instantaneous biological density is first calculated. The calculation formula is as follows: ; Based on this, based on the time step from the initial time step to the time step t Instantaneous biological density sequence, calculating time steps t Corresponding cumulative median biological density The calculation method is as follows: ; In a preferred embodiment, to avoid numerical fluctuations caused by excessively frequent updates of statistical results, the cumulative median result can be output at a preset time interval, such as updating the statistical result every 10 seconds, thereby further improving the stability and readability of the result while ensuring time resolution.

[0064] In summary, the automatic detection and spatial calibration of fixed reference objects (square quadrats) based on a recognition model employed in this embodiment not only theoretically solves the problem of converting pixel scale to actual area, but also provides a guarantee for stable statistics in complex tidal flat environments in practice. By analyzing the pixel parameters of the square boxes with fixed areas in real time, the detection results can be converted into comparable biological density values, thereby achieving consistent quantitative assessments across different scenarios and time periods. This method not only improves the scientific rigor and interpretability of the detection, but also provides a reliable technical path for subsequent ecological monitoring and long-term trend analysis.

[0065] This embodiment can also output a recognition result video (with borders and confidence scores), a temporal density curve, and a data summary file, which can be found in [reference needed]. Figure 9 The diagram shown is a schematic of a video output of the recognition result.

[0066] This embodiment also allows for statistical analysis and result output. The species and their densities in the mudflats are identified and calculated for each of the video samples from all quadrats. The biomass density of each species within the target area of ​​the surveyed mudflat is obtained by averaging (taking 5 quadrats as an example): ; in, i As a species, The corresponding sample plot video number, For species i The median density in the nth sample plot video.

[0067] This invention provides an intelligent identification and density monitoring method for intertidal mudflat benthic animals. Multiple independent quadrats are randomly or quasi-randomly deployed in the target area of ​​the surveyed mudflat, and video is collected for each quadrat. Each quadrat is dropped horizontally, and the distance between quadrats is greater than a preset distance threshold. The video is preprocessed to obtain an image set. The image set is input into a pre-trained recognition model, which outputs detection results. The recognition model is structurally optimized based on the YOLO11 model for intertidal ecological monitoring tasks. The detection results include: bounding boxes, categories, and confidence scores of various targets. Targets include not only intertidal mudflat benthic animals (fiddler crabs and mudskippers) but also fixed reference objects (square quadrats) used for spatial scale calibration. Spatial scale calibration is performed using the pixel information of the square quadrats detected by the recognition model, and the actual physical coverage area of ​​the video image is calculated. Based on the number of benthic animal targets in each video frame within the fixed area, the representative number of organisms within the fixed area is determined using median statistics, and the density of organisms within the fixed area is calculated in conjunction with the actual physical coverage area.

[0068] The intelligent identification and density monitoring method for intertidal benthic animals provided in this embodiment can calculate the biodensity based on the representative quantity of the target species. By performing time-cumulative statistics on the number of targets detected in the sample plot video sequence and using the median statistical method to determine the representative quantity of the targets, compared with the mean statistical method, it can significantly reduce the fluctuations in density estimation caused by the rapid entry and exit of fiddler crabs and mudskippers into and out of the field of view, mutual occlusion, and environmental changes, thereby improving the stability and reliability of the statistical results. This statistical method, combined with UAV aerial photography and YOLO recognition for intelligent identification and density monitoring of intertidal benthic animals, can reduce the interference of manual surveys on monitoring results, improve the automation and repeatability of density estimation, and is suitable for non-contact monitoring scenarios of intertidal benthic animals.

[0069] In summary, this invention addresses the technical challenges of "spatial migration + temporal fluctuation + scale instability" in monitoring intertidal benthic animals. By employing spatial decorrelation quadrat deployment, robust temporal median statistics, and an automatic scale calibration mechanism, it constructs a highly reliable biological density estimation method that can be implemented automatically.

[0070] See also Figure 10 The diagram shows a method for intelligent identification and density monitoring of benthic animals in intertidal mudflats. Figure 10 The image shows five quadrats (#1-#5, each 0.2m x 0.2m in size). The drone flies directly above quadrat #1 at a height of 3-4 meters. After capturing video of the quadrats, the drone can identify the main species and calculate the biological density. Figure 10 The image shows identified mudskippers, fiddler crabs, and other crab species. Figure 10 The image also shows the shooting area of ​​the drone gimbal, which can be determined based on the area of ​​the sample plot (0.04m²). 2 By analyzing the area of ​​the captured image and the percentage of pixels in it, the biological density of crabs and fish can be calculated.

[0071] The method provided in the embodiments of the present invention mainly includes the following: (1) Randomly throw multiple quadrats in multiple directions to ensure that each quadrat falls horizontally and is more than 10m apart from each other. Control the drone to fly slowly to a position 4-5m directly above the quadrat, and adjust the gimbal shooting angle to make it as perpendicular to the quadrat as possible; set the video parameters to 4K resolution and 5-7 minutes in duration to ensure that the quadrat and organisms in the picture are clear.

[0072] (2) In this embodiment, the cumulative median of biological density is used as the characterization index of the target species density in the video, and it is calculated using a full-time cumulative method. Specifically, when calculating the median at any time, the statistical sample set covers all historical detection data from the initial monitoring time to the current time, rather than being limited to local short-term data segments. By introducing this cumulative statistical mechanism, the median result output by the model at each statistical point does not only reflect the current instantaneous state, but is a comprehensive evaluation based on the detection values ​​at multiple historical times, thereby effectively suppressing and smoothing random noise caused by occasional occlusion, changes in illumination, or fluctuations in detection confidence.

[0073] The method provided in the embodiments of the present invention has the following main technical advantages: (1) Non-contact intelligent monitoring: This embodiment combines low-altitude vertical aerial photography by UAV with deep learning algorithms to achieve remote, non-contact real-time monitoring of intertidal benthic animals, avoiding human interference and improving the natural authenticity and spatiotemporal continuity of sample data.

[0074] (2) Strong multi-scale target detection capability: This embodiment introduces the C3k2 module, fast spatial pyramid pooling (SPPF) and parallel attention mechanism (C2PSA) to enable the model to stably identify targets of different scales in complex tidal flat background.

[0075] (3) High robustness and anti-interference performance: This embodiment effectively eliminates instantaneous noise fluctuations caused by changes in illumination, reflections from ocean waves and biological occlusion through a sliding median filtering mechanism; the recognition model adopts an anchor point free decoding method, which enhances the continuity and stability of bounding box prediction.

[0076] (4) Automatic scale calibration and actual density conversion: In this embodiment, a fixed reference object (0.2m×0.2m quadrat) is used as one of the detection targets of the identification model. The model automatically identifies the quadrat and extracts its pixel parameters, directly completing the automatic conversion from image pixels to actual physical area, thereby outputting a biological density index with ecological significance, effectively avoiding manual calibration errors.

[0077] (5) Lightweight and high efficiency: The improved YOLO11 model used in this embodiment maintains high detection accuracy (mAP=0.89) while significantly reducing the number of parameters and computational load, and the inference speed reaches 60FPS, which is suitable for real-time processing at the edge of UAVs.

[0078] (6) High scalability and application prospects: The above-mentioned methods provided in this embodiment are not only applicable to the monitoring of large benthic species in tidal flats (mainly fiddler crabs and mudskippers), but can also be extended to other intertidal organisms (such as shellfish, snails, and crabs) and wetland ecological monitoring scenarios, providing technical support for intertidal flat restoration assessment, dynamic monitoring of coastal ecosystems and climate change response research.

[0079] Example 2: Corresponding to the above method embodiments, this invention provides an intelligent identification and density monitoring device for intertidal benthic animals, see [link to relevant documentation]. Figure 11 The diagram shows a structural schematic of an intelligent identification and density monitoring device for intertidal benthic animals. This device includes: The quadrat layout and aerial photography data acquisition module 1001 is used to lay out multiple independent quadrats in a random or quasi-random manner in the target area of ​​the survey beach and acquire video of each quadrat; wherein, each quadrat is dropped horizontally and the distance between each quadrat is greater than a preset distance threshold. The target recognition and spatial scale calibration module 1002 is used to preprocess the video, construct a unified video processing environment, perform target recognition processing in the video processing environment, and perform spatial scale calibration based on the quadrat reference information in the target recognition results. The target recognition and spatial scale calibration module 1002 includes: a target recognition processing submodule 1011, which is used to input video frames into a pre-trained recognition model and output detection results; wherein, the recognition model is obtained by structural optimization based on the YOLO11 model for the intertidal ecological monitoring task; the detection results include: the bounding box, category, and confidence of the target; the targets include: intertidal benthic animals and fixed references used for spatial scale calibration; the intertidal benthic animals include: fiddler crabs and mudskippers, and the fixed references include: square quadrats; The target recognition processing and spatial scale calibration module 1002 includes: a spatial scale calibration submodule 1012, which is used to establish the correspondence between the image pixel scale and the actual tidal flat area based on the number of pixels occupied by the square quadrat detected by the recognition model in the video image, combined with the known real area of ​​the square quadrat, and calculate the actual physical coverage area of ​​the video image as a fixed area. The target recognition processing and spatial scale calibration module 1002 includes a density calculation submodule 1013, which is used to determine the representative number of targets in the fixed area based on the number of benthic animal target species in each video frame within the fixed area, using the median statistical method, and calculate the biological density of target species in the fixed area in combination with the actual tidal flat area obtained by spatial scale calibration.

[0080] This invention provides an intelligent identification and density monitoring device for intertidal benthic animals. It collects video footage of intertidal mudflat quadrats via drone aerial photography and combines this with an identification model optimized for intertidal ecological monitoring tasks based on the YOLO11 model. This enables simultaneous automatic identification and classification of mudflat benthic animals (fiddler crabs and mudskippers) and spatial scale references (square quadrats). By introducing a spatial scale calibration method based on the model's identification results, it directly uses the pixel proportion of the identified quadrats in the image and their known actual area to uniformly calibrate the scale of the entire video image, automatically converting the detection results from image pixel scale to actual mudflat area. This method adapts to subtle fluctuations in shooting height and angle without relying on camera pose parameters, providing a reliable physical spatial basis for accurate calculation of biological abundance and density. It employs median statistics to accumulate the number of effective biological targets detected in the sample plot video sequence over time. Compared to traditional mean-based methods, this significantly reduces density estimation fluctuations caused by animals rapidly entering and leaving the field of view, mutual occlusion, and environmental changes, improving the stability and reliability of density statistics. This method reduces human interference, enabling non-contact automatic monitoring of intertidal benthic animals, improving the repeatability and applicability of monitoring results, and is suitable for long-term ecological monitoring and large-scale surveys. Compared to existing technologies, this invention achieves high-precision, stable, and automated monitoring of intertidal benthic animals, significantly improving monitoring efficiency and data reliability.

[0081] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the intelligent identification and density monitoring device for intertidal benthic animals described above can be referred to the corresponding process in the embodiments of the aforementioned intelligent identification and density monitoring method for intertidal benthic animals, and will not be repeated here.

[0082] Example 3: This invention also provides an electronic device for running the above-described intelligent identification and density monitoring method for intertidal benthic animals; see also Figure 12 The diagram shows the structure of an electronic device, which includes a memory 100 and a processor 101. The memory 100 is used to store one or more computer instructions, which are executed by the processor 101 to realize the above-mentioned intelligent identification and density monitoring method for intertidal benthic animals.

[0083] Furthermore, Figure 12 The electronic device shown also includes a bus 102 and a communication interface 103, with the processor 101, the communication interface 103 and the memory 100 connected via the bus 102.

[0084] The memory 100 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 102 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 12 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0085] Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 100, and processor 101 reads information from memory 100 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0086] This invention also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the above-mentioned intelligent identification and density monitoring method for intertidal benthic animals. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0087] The computer program product of the intelligent identification and density monitoring method and device for intertidal benthic animals provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0088] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and / or device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0089] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0090] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0092] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent identification and density monitoring of benthic animals in intertidal mudflats, characterized in that, The method includes: Multiple independent fixed reference points are randomly or quasi-randomly deployed in the target area of ​​the survey beach, and video of each reference point is collected; wherein each reference point is dropped horizontally, and the distance between each reference point is greater than a preset distance threshold. The video is preprocessed to construct a unified video processing environment. Target recognition processing is then performed within this environment, and spatial scale calibration is performed based on the fixed reference object information in the target recognition results. Based on the number of target species in each video frame within the fixed area, the representative number of target species within the fixed area is determined using median statistics, and the biological density of target species within the fixed area is calculated in conjunction with the actual physical coverage area of ​​the video frame.

2. The method according to claim 1, characterized in that, The target recognition process includes: inputting video frames into a pre-trained recognition model and outputting detection results; wherein, the recognition model is obtained by structural optimization based on the YOLO11 model for intertidal ecological monitoring tasks; the detection results include: the target's bounding box, category, and confidence score; the targets include: intertidal benthic animals and fixed references for spatial scale calibration; the intertidal benthic animals include: fiddler crabs and mudskippers, and the fixed references include: square quadrats; The spatial scale calibration includes: based on the number of pixels occupied by the square quadrat in the video frame detected by the recognition model, combined with the known real area of ​​the square quadrat, establishing a correspondence between the image pixel scale and the actual tidal flat area, and calculating the actual physical coverage area of ​​the video frame as a fixed area.

3. The method according to claim 1, characterized in that, The steps for acquiring video of each sample plot include: Control the drone to fly directly above each of the sample plots, and adjust the gimbal's shooting angle to be perpendicular to the sample plot; Adjust the video parameters and acquire video for each sample plot; The environmental parameters were obtained when the video of each sample plot was collected.

4. The method according to claim 1, characterized in that, The steps for preprocessing the video include: Extract image frames from the video; The biological contours of the image frames are enhanced through histogram equalization and background modeling. Noise regions in the image frame are removed by thresholding and edge detection.

5. The method according to claim 1, characterized in that, The structure of the recognition model includes: a backbone network, a feature enhancement module, and a detection head; the mechanism of the recognition model includes: cross-stage feature reuse, spatial pyramid pooling, and parallel attention mechanism. The recognition model outputs the feature map of the original input image in the following manner: The recognition model halves the spatial resolution of the original input image through convolutional layers to extract primary features of the original input image; The recognition model performs multiple downsampling operations on the primary features, and each downsampling operation outputs multiple feature maps at different scales; wherein, each downsampling operation halves the resolution of the input feature map and increases the number of channels of the input feature map.

6. The method according to claim 5, characterized in that, The feature map includes: shallow features and deep features; The shallow features contain details and edge information, which are used to identify targets at smaller scales; The deep features contain global semantics and contextual information, which are used to identify targets at a large scale or objects in complex backgrounds.

7. The method according to claim 5, characterized in that, The total loss function of the recognition model is determined based on the intersection-union ratio loss, classification loss, and distributed bounding box loss.

8. The method according to claim 1, characterized in that, The step of establishing the correspondence between image pixel scale and actual tidal flat area based on the number of pixels occupied by the square quadrat in the video frame detected by the recognition model and the known actual area of ​​the square quadrat includes: Obtain the bounding box information of the square sample plot output by the recognition model, and calculate the number of pixels occupied by the sample plot in the corresponding video frame based on the bounding box information; Calculate the actual tidal flat area corresponding to a single pixel based on the actual area of ​​the reference object and the corresponding number of pixels. Calculate the actual tidal flat area corresponding to the entire image based on the pixel area conversion factor; A correspondence between image pixel scale and actual tidal flat area is established based on the actual tidal flat area corresponding to the individual pixel and the actual tidal flat area corresponding to the entire image.

9. The method according to claim 1, characterized in that, After the steps of inputting video frames into a pre-trained recognition model and outputting detection results, the method further includes: The detection results of each frame in the video are filtered, and the number of valid target species is recorded; The median of targets within the fixed area is determined by a cumulative calculation method based on the number of effective target species; The biological density of the target within the fixed area is determined based on the number of effective target species by using the median density calculation method. The species biomass of the target region is calculated by averaging the values ​​of the target species within each of the fixed regions.

10. A smart identification and density monitoring device for intertidal benthic animals, characterized in that, The device includes: The quadrat deployment and aerial data acquisition module is used to deploy multiple independent quadrats in a random or quasi-random manner in the target area of ​​the surveyed beach, and to acquire video of each quadrat; wherein each quadrat is dropped horizontally, and the distance between each quadrat is greater than a preset distance threshold. The target recognition and spatial scale calibration module is used to preprocess the video, construct a unified video processing environment, perform target recognition processing in the video processing environment, and perform spatial scale calibration based on the sample reference information in the target recognition results. The target identification and spatial scale calibration module includes a density calculation submodule, which is used to determine the representative number of targets in the fixed area based on the number of benthic animal target species in each video frame of the fixed area using the median statistical method, and calculate the biological density of target species in the fixed area in combination with the actual tidal flat area obtained by the spatial scale calibration.

Images