Weather bird monitoring method based on deep learning and application

Abstract

The invention discloses a migrant bird monitoring method based on deep learning and application. The monitoring method comprises the following steps: initializing hardware equipment and acquiring data; preprocessing and denoising the data; identifying a migrant bird image and estimating a distance; data storage, segmented analysis and feature extraction; migrant bird migration behavior analysis; migrant bird migration trend prediction; and carrying out data visualization display. According to the invention, the real-time monitoring and accurate analysis of the activities of the migrant birds are realized through the initialization of hardware equipment and the collection of environmental parameters in combination with deep learning and data processing technologies. The data accuracy is improved by adopting Gaussian weighted moving average filtering and dimension reduction technologies, and meanwhile, the target detection precision is improved through an improved MPDIOU-ODConv-YOLOv5 model. The migration behavior is classified and analyzed by using a k-means clustering algorithm, and the migration trend is predicted based on a Markov chain, so that a scientific basis and data support are provided for migrant bird protection and ecological management.

Description

Technical Field

[0001] The present invention relates to the field of biological monitoring technology, and in particular to a migratory bird monitoring method and application based on deep learning. Background Art

[0002] Migratory birds play a key role in ecosystems and are of great significance for maintaining ecological balance and functional integrity. As important indicator organisms of wetland, lake and river ecosystems, the community characteristics of migratory birds are key indicators for monitoring the health of ecosystems.

[0003] However, with the continuous expansion of human activities, the migration routes and habitats of migratory birds are facing unprecedented threats. Environmental pollution, climate change, illegal hunting and other factors have led to a significant decrease in the number of migratory bird populations, and some rare migratory birds are even on the verge of extinction. In order to protect migratory bird resources and maintain ecological balance and biodiversity, governments and scientific research institutions have continuously strengthened the monitoring and protection of migratory birds in recent years. However, traditional migratory bird monitoring methods mainly rely on manual observation and recording, which not only has limited monitoring coverage and low efficiency, but also has technical problems such as difficulty in dynamic tracking and insufficient data accuracy and consistency.

[0004] In response to these problems, we have developed an efficient and accurate migratory bird monitoring technology. By developing advanced recognition models and innovative monitoring methods, this technology aims to improve the efficiency and accuracy of migratory bird monitoring, provide strong technical support for migratory bird protection, and promote the modernization and intelligence of protection work. Summary of the invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a migratory bird monitoring method and application based on deep learning, which solves the technical problems that the traditional migratory bird monitoring method mainly relies on manual observation and recording, which not only has limited monitoring coverage and low efficiency, but also has difficulties in dynamic tracking and insufficient data accuracy and consistency.

[0006] To achieve the above purpose, the following technical solutions can be used:

[0007] The present invention provides a migratory bird monitoring method based on deep learning, comprising the following steps:

[0008] S1. Initialization of hardware equipment and data collection: Turn on the hardware equipment, connect the network camera, related sensors and GPS positioning device, obtain real-time video data of migratory bird activities through the network camera, and collect environmental parameters, including temperature, humidity, light intensity and geographic location information;

[0009] S2, data preprocessing and denoising: perform format conversion and frame extraction processing on the collected video data, and use Gaussian weighted moving average filtering to denoise the original data;

[0010] S3. Migratory bird image recognition and distance estimation: Use the self-developed MPDIoU-ODConv-YOLOv5 model to identify and extract the species and number of migratory birds in the denoised video, calculate the relative size of the migratory birds based on the average size of the detection frame, and estimate the distance between the migratory birds and the camera based on the pinhole imaging principle;

[0011] S4, data storage, segment analysis and feature extraction: the identified migratory bird species, quantity and environmental monitoring data are stored in the database, and features are extracted by time segment. The average temperature, average light and maximum humidity are selected as features. The feature data are reduced in dimension by principal component analysis, and a subspace projection matrix is ​​constructed to reduce the data dimension to form a reduced dimension data set.

[0012] S5. Analysis of migratory bird migration behavior: Based on the reduced-dimensional data, the k-means algorithm was used to divide the migratory status of migratory birds into three categories, and the characteristics of cluster centers were analyzed to identify the migration patterns and their influencing factors;

[0013] S6. Prediction of migratory bird migration trends: Based on the historical data of migratory birds, a state transition matrix based on Markov chain is constructed to quantify the transition rules between migration states. The multi-step transition matrix is ​​used to predict the changing trends of migratory bird species and numbers in the future, providing support for the analysis of migratory bird migration rules.

[0014] S7. Data visualization display: Display the data generated by steps S1-S6 through a Web platform.

[0015] As a further solution of the present invention, in step S2, the specific method steps for denoising the original data by using Gaussian weighted moving average filtering are as follows:

[0016] S21, determine the filter window size N and standard deviation σ, calculate the Gaussian weight w(i) and perform normalization processing to ensure that the sum of the weights satisfies ∑w(i)=1, where, i is the relative position of the sample, σ is the standard deviation of the Gaussian distribution;

[0017] S22, the filter window slides to cover the original data, and the smoothing value is calculated using the Gaussian weight. The smoothing value calculation formula is: Where y[k] is the kth value of the output sequence, w[i] is the i-th Gaussian weighting coefficient, and x[k-(N-1) / 2+i] is the kth value in the original sequence;

[0018] S23, using data filling strategies to avoid filtering errors at both ends of the sequence, including repeating the first and last elements of the sequence or performing period extension;

[0019] S24. Calculate the mean absolute error (MAE) and the root mean square error (RSME) to evaluate the denoising effect and ensure that the smoothed data accurately reflects the actual environment. The relevant calculation formula is:

[0020] Among them, xi is the processed data and yi is the original data.

[0021] As a further solution of the present invention, in step S3, the self-developed MPDIoU-ODConv-YOLOv5 model is obtained by improvement based on the classic YOLOv5 target detection framework, and the specific improvement method steps are as follows:

[0022] S31. Replace the original target box regression loss function GIoU in YOLOv5 with MPDIoU, introduce the maximum pooling operation, and optimize the model's ability to capture the distance and position dislocation between target boxes. The calculation formula of MPDIoU is:

[0023] Among them, |A∩B| is the intersection area of ​​the predicted box A and the real box B, |A∪B| is the union area of ​​the predicted box A and the real box B, d1 and d2 are the Euclidean distances between the center points of the predicted box A and the real box B, λ1 and λ2 are dynamically adjusted weight factors, and w and h are the width and height of the joint box;

[0024] S32. Use the full-dimensional dynamic convolution ODConv module to replace the conventional convolution in YOLOv5, and optimize the dynamic weight generation in combination with the attention mechanism to improve the model's adaptability to the diversity characteristics of migratory birds. The calculation formula of ODConv is: Among them, a i is the dynamic weight associated with the input data x, W i is the i-th convolution kernel, x and y are the input and output feature maps respectively;

[0025] S33, replace the Backbone of YOLOv5 with the lightweight network ob_mbilenetv2_075 that integrates ODConv and Ghost modules. By optimizing the fusion of multi-scale features, the computational complexity is reduced while enhancing the model's adaptability to targets of different sizes, thereby improving detection accuracy and efficiency.

[0026] S34. Integrate MPDIoU and ODConv modules in the training process of YOLOv5, optimize the loss function through multiple rounds of iterations, and improve the target detection performance.

[0027] As a further solution of the present invention, in step S4, the characteristic data is reduced in dimension by principal component analysis, and a subspace projection matrix is ​​constructed to reduce the data dimension. The specific method steps for forming a reduced dimension data set are as follows:

[0028] S41. Calculate sample average vector

[0029] S42. Subtract the mean of the corresponding feature from each feature value in the data set to complete the data centering process.

[0030] S43. Calculate the covariance matrix Σ of the centralized data set.

[0031] S44, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​λ1,λ2,…,λ n and the corresponding eigenvectors v1,v2,…,v n , according to the size of the eigenvalue, select the eigenvectors corresponding to the first k eigenvalues ​​as the principal components to form a projection matrix W;

[0032] S45. Project the data set into the subspace formed by the first k eigenvectors: X reduced =X centered W, thereby achieving data dimensionality reduction.

[0033] As a further solution of the present invention, the specific steps of step S5 are as follows:

[0034] S51, obtaining the species and number of migratory birds in the original data from step S3, and obtaining the main environmental characteristics after dimensionality reduction from step S4;

[0035] S52, selecting the k-means clustering algorithm as the analysis tool, setting the number of clusters to k=3, and dividing the migratory status of migratory birds into three categories;

[0036] S53. Analyze the characteristics of cluster centers in the clustering results, clarify the characteristics of the three migration states, and identify the key factors affecting the migration behavior of migratory birds;

[0037] S54. The clustering results are displayed visually through charts, and different colors are used to represent the three types of migration status for easy comparison and analysis.

[0038] As a further solution of the present invention, the specific steps of step S6 are as follows:

[0039] S61, according to the migratory bird data processed in step S4, dividing the processed data into a finite number of state sets as state variables of the system;

[0040] S62, calculation state s i The initial probability of occurrence is calculated as follows: Among them, n(s i ) is in state si The number of data, N is the total number of data;

[0041] S63, establish a single-step transition probability matrix: The single-step transition probability matrix represents the probability of transitioning from one state to another state, and the calculation formula is: Among them, P(s j │s i ) is from state s i Transfer to state s j The probability of n(s i →s j ) is the number of state transitions, n(s i ) is state s i Total number of occurrences;

[0042] S64, constructing a step transfer probability matrix: Based on the single-step transfer probability matrix, the multi-step transfer probability matrix can be further calculated by the recursive formula: P (k) =P (k-1) ·P, where P (k) is the k-step transition probability matrix, and P is the single-step transition probability matrix;

[0043] S65. Predict future status: According to the state transition formula of the Markov chain model, predict the future migratory bird data: P t+k =P t ·P (k) , where P t+k is the predicted state distribution, P t is the current state distribution, P (k) is the k-step transition probability matrix.

[0044] As a further solution of the present invention, the method can be applied to migratory bird protection and ecological environment monitoring, and can be used to monitor migratory bird migration patterns, population sizes and habitat environmental parameters in real time, thereby providing data support for migration monitoring and habitat protection.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention realizes real-time and comprehensive monitoring of migratory bird activities through the initialization of hardware equipment and the collection of environmental parameters, combined with deep learning processing of video data, and provides multi-dimensional data support for accurate analysis of migration patterns.

[0047] 2. The present invention uses Gaussian weighted moving average filtering to denoise the data and combines dimensionality reduction technology to extract key features, effectively reducing the impact of noise on monitoring results and improving the accuracy and effectiveness of data analysis; dimensionality reduction processing optimizes the data structure through principal component analysis (PCA), reduces dimensional redundancy, helps to quickly extract environmental features and simplify subsequent analysis and calculations.

[0048] 3. The present invention uses the improved MPDIoU-ODConv-YOLOv5 model to effectively improve the accuracy and robustness of migratory bird target detection: the MPDIoU loss function more sensitively captures the distance and misalignment information between the predicted box and the real box, enhancing the positioning ability of the target detection model; the ODConv convolution module dynamically adapts to the diversity of migratory bird characteristics, significantly improving the recognition effect of migratory birds in complex backgrounds.

[0049] 4. The present invention classifies and analyzes the migratory behavior of migratory birds by using the k-means clustering algorithm, clarifies the migration patterns and key environmental influencing factors, and facilitates targeted protection; constructs a state transition matrix based on the Markov chain to predict migration trends and provide data support for the formulation of ecological protection measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the overall method steps of the present invention;

[0051] Figure 2 It is a system framework diagram involved in the monitoring method of the present invention;

[0052] Figure 3 A data graph of the number and species of original migratory birds identified by the present invention;

[0053] Figure 4 This is a data graph after Gaussian filtering is performed on the identified number and species of migratory birds in the present invention;

[0054] Figure 5 This is the network structure diagram of the self-developed MPDIoU-ODConv-YOLOv5 used in the present invention;

[0055] Figure 6 A characteristic parameter chart selected by the present invention for the collected data;

[0056] Figure 7 This is a PCA data dimension reduction analysis diagram of the data performed by the present invention;

[0057] Figure 8 The present invention performs K-means cluster analysis on the data;

[0058] Fig. 9 A Markov prediction analysis diagram of the data is performed for the present invention;

[0059] Fig.10 This is a screenshot of the monitoring interface of the web page platform of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Embodiment 1:

[0062] like Figure 1-10 As shown, a migratory bird monitoring method based on deep learning provided by an embodiment of the present invention includes the following steps:

[0063] S1. Initialization of hardware equipment and data collection: Turn on the hardware equipment, connect the network camera, GPS positioning device and related sensors (including temperature and humidity sensors, light sensors), obtain real-time video data of migratory bird activities through the network camera, and collect environmental parameters, including temperature, humidity, light intensity and geographic location information;

[0064] S2, data preprocessing and denoising: the collected video data is format converted and frame extracted, and the original data is denoised using Gaussian weighted moving average filtering;

[0065] S3, migratory bird image recognition and distance estimation: The denoised video is input into the self-developed MPDIoU-ODConv-YOLOv5 model to identify the migratory birds in the image, extract the type and number of each migratory bird, calculate the relative size of the migratory birds based on the average size of the detection frame, and use the pinhole imaging principle to estimate the distance between the migratory birds and the camera;

[0066] S4, data storage, segment analysis and feature extraction: the identified migratory bird species, quantity and environmental monitoring data are stored in the database, and features are extracted by time segment. The average temperature, average light and maximum humidity are selected as features. The feature data are reduced in dimension by principal component analysis, and a subspace projection matrix is ​​constructed to reduce the data dimension to form a reduced dimension data set.

[0067] S5. Analysis of migratory bird migration behavior: Based on the k-means algorithm, cluster analysis was performed on the species, number and environmental parameters of migratory birds in the dimensionality reduction data set, the migratory status of migratory birds was divided into three categories, and the cluster centers were calculated to identify the migration patterns;

[0068] S6. Prediction of migratory bird migration trends: Based on the historical data of migratory birds, a state transition matrix based on Markov chain is constructed to quantify the transition rules between migration states. The multi-step transition matrix is ​​used to predict the changing trends of migratory bird species and numbers in the future, providing support for the analysis of migratory bird migration rules.

[0069] S7, data visualization display: display the data generated by steps S1-S6 through the web platform. It may include: dynamic maps (marking the current location and migration path of migratory birds), timeline graphs (changes in the number and species of migratory birds and environmental parameters), curve graphs of prediction model results (migration trends), and annotated displays of videos and images (migratory bird identification results), with the aim of providing users with intuitive analysis results and decision support.

[0070] Furthermore, in step S2, the specific method steps of using Gaussian weighted moving average filtering to denoise the original data are as follows:

[0071] S21, determine the filter window size N and standard deviation σ, calculate the Gaussian weight w(i) and perform normalization processing to ensure that the sum of the weights satisfies ∑w(i)=1, where, i is the relative position of the sample, σ is the standard deviation of the Gaussian distribution, which controls the smoothness of the Gaussian distribution;

[0072] S22, the filter window slides to cover the original data, and the smoothing value is calculated using the Gaussian weight. The smoothing value calculation formula is: Where y[k] is the kth value of the output sequence, w[i] is the i-th Gaussian weighting coefficient, and x[k-(N-1) / 2+i] is the kth value in the original sequence;

[0073] S23, using data filling strategies to avoid filtering errors at both ends of the sequence, including repeating the first and last elements of the sequence or performing period extension to ensure data integrity at both ends;

[0074] S24. Calculate the mean absolute error (MAE) and the root mean square error (RSME) to evaluate the denoising effect and ensure that the smoothed data accurately reflects the actual environment. The relevant calculation formula is:

[0075] Among them, xi is the processed data, yi is the original data, and n is the total number of data points, ensuring that the smoothed data accurately reflects the actual environment; when using these two indicators, MAE can help you understand the overall error level, while RMSE focuses more on the reduction of larger errors in the filtering process. The combination of the two can comprehensively evaluate the denoising effect and ensure the accuracy and stability of the data after processing.

[0076] It should be noted that some abnormal situations may occur during the data collection process, such as the camera being blocked, the hardware equipment being disturbed by animals at close range, or the camera being blinded by strong light, which will cause noise signals to be mixed in the original data. Since the original data is a non-periodic signal, the Gaussian weighted moving average filtering method is very suitable for processing such signals.

[0077] refer to Figure 3-Figure 4 , Figure 3 Indicates the original number and species of migratory birds, Figure 4 It shows the number and species of migratory birds after Gaussian filtering. It can be clearly seen that denoising helps ensure the accuracy of subsequent image recognition and analysis, making the results more stable and reliable and reducing errors introduced by noise.

[0078] Further, such as Figure 5 As shown in step S3, the self-developed MPDIoU-ODConv-YOLOv5 model is improved on the basis of the classic YOLOv5 target detection framework. The specific improvement steps are as follows:

[0079] S31. Replace the original target box regression loss function GIoU in YOLOv5 with MPDIoU, introduce the maximum pooling operation, and optimize the model's ability to capture the distance and position dislocation between target boxes. The calculation formula of MPDIoU is:

[0080] Among them, |A∩B| is the intersection area of ​​the predicted box A and the real box B, |A∪B| is the union area of ​​the predicted box A and the real box B, d1 and d2 are the Euclidean distances between the center points of the predicted box A and the real box B, λ1 and λ2 are dynamically adjusted weight factors, and w and h are the width and height of the joint box (or minimum closure box);

[0081] S32. Use ODConv to replace the conventional convolution in YOLOv5 to improve the model's adaptability to the diverse characteristics of migratory birds. Conventional convolution uses a fixed static convolution kernel, which is independent of the input sample. Dynamic convolution linearly weights multiple convolution kernels and dynamically adjusts the weights according to the input, making it input-dependent. It is combined with the attention mechanism to optimize the generation of dynamic weights, thereby improving the model's adaptability to the diverse characteristics of migratory birds.

[0082] The calculation formula of ODConv is: Among them, a i is the dynamic weight associated with the input data x, W i is the i-th convolution kernel, x and y are the input and output feature maps respectively;

[0083] The MPDIoU-ODConv-YOLOv5 network structure is as follows Figure 5 As shown,

[0084] S33, replace the Backbone of YOLOv5 with the lightweight network ob_mbilenetv2_075 that integrates ODConv and Ghost modules (a technology used in convolutional neural networks (CNNs) in deep learning to improve efficiency by reducing the computational complexity of the model). By optimizing multi-scale feature fusion, the amount of calculation is reduced while enhancing the model's adaptability to targets of different sizes, thereby improving detection accuracy and efficiency;

[0085] S34. Integrate MPDIoU and ODConv modules in the training process of YOLOv5, and optimize the loss function through multiple rounds of iterations to improve the target detection performance. Through multiple rounds of iterative optimization of the loss function, the detection performance of the model can be continuously improved, especially in terms of precision and recall in target detection. Continuous optimization of the loss function helps improve the robustness of the model in complex environments, especially in the dynamically changing migratory bird environment, ensuring accuracy and stability during the monitoring process.

[0086] Further, in step S4, data storage, extraction of data features by time period, and feature extraction and dimensionality reduction are completed through the following specific steps:

[0087] The migratory bird identification results and environmental monitoring data are stored in a database (usually using a relational database such as MySQL) to ensure data traceability and provide a basis for subsequent analysis;

[0088] Extracting data features by time segment can reflect the laws of migratory bird activities and environmental changes in different time periods; due to the limited field of view of the camera, the results of migratory bird identification may differ between consecutive frames, so the data of one day is divided into three time periods, and the original data is split into multiple blocks. In order to evaluate the characteristics of the number and species of migratory birds in different time periods, the species, number and environmental characteristics of migratory birds in each time period, such as the highest temperature, the lowest temperature, and the temperature difference, are calculated. Feature parameters such as Figure 6 As shown, n is the number of data items in this period, and m is the number of data items for the whole day.

[0089] like Figure 7 As shown in the figure, due to the large number of features in the original data, there may be redundancy, which increases the computational complexity and easily leads to overfitting. To this end, we use the PCA (Principal Component Analysis) algorithm to reduce the dimensionality of the data and select average temperature, average light and maximum humidity as features. The PCA algorithm is used to map high-dimensional data to a low-dimensional space through linear transformation while retaining important information. The process includes calculating the eigenvectors and eigenvalues ​​of the data. The specific method steps are as follows:

[0090] S41. Calculate sample average vector Calculating the average vector of the sample is the basis for data centering. By calculating the average value of each feature, a benchmark can be provided for subsequent centralization processing.

[0091] S42. Subtract the mean of the corresponding feature from each feature value in the data set to complete the data centering process. Subtracting the mean from each eigenvalue ensures that the mean of the data is 0, thereby avoiding the impact of scale differences between features on the analysis results. This is the prerequisite for PCA and helps to improve the calculation accuracy during eigenvalue decomposition.

[0092] S43. Calculate the covariance matrix Σ of the centralized data set. The covariance matrix is ​​an important tool for describing the relationship between features, and measures the relationship between different features (such as linear correlation). By calculating the covariance matrix, we can further understand the dependencies between different features and help select the most meaningful features.

[0093] S44, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​λ1,λ2,…,λ n and the corresponding eigenvectors v1,v2,…,v n , according to the size of the eigenvalue, select the eigenvectors corresponding to the first k eigenvalues ​​as the principal components. These vectors can maximize the variability of the data and ultimately form the subspace W matrix after dimensionality reduction;

[0094] S45. Project the data set into the subspace formed by the first k eigenvectors: X reduced =X centered W, thereby achieving data dimensionality reduction. Through this operation, the original high-dimensional data is converted into a low-dimensional space while retaining most of the information in the data.

[0095] Through the above steps, data dimension reduction can help reduce computational complexity and data redundancy, reduce the risk of overfitting, and improve model training efficiency and accuracy in migratory bird monitoring programs. At the same time, it can simplify data visualization and understanding, enhance data storage and management efficiency, and thus optimize the overall monitoring effect.

[0096] Further, such as Figure 8 As shown, in order to better describe the migratory status of migratory birds, when it is unclear what the label is, the k-means clustering algorithm is used to analyze the data. The specific steps of step S5 are as follows:

[0097] S51, extract the species and number of migratory birds in the original data from step S3, and obtain the main environmental characteristics after dimensionality reduction (such as average temperature, average light and maximum humidity) from step S4. These environmental characteristics will help the clustering algorithm to identify and group the migratory behavior of migratory birds.

[0098] S52. Select k-means clustering algorithm as the analysis tool, and set the number of clusters to k=3, that is, divide the migration status of migratory birds into three categories.

[0099] Given a data sample set G = {x1, x2, ..., x m}, the algorithm divides D = {D1, D2, ..., D k The minimized square error of} is: in It is cluster D i In this way, the k-means algorithm iterates continuously and finally obtains three clusters, and the center of each cluster represents a migratory state of migratory birds.

[0100] S53. Analyze the characteristics of cluster centers in the clustering results. By comparing the characteristics of different clusters (such as temperature, humidity, species and number of migratory birds, etc.), clarify the characteristics of the three migration states. Identifying these characteristics will help reveal the key environmental factors that affect the migration behavior of migratory birds.

[0101] S54. The clustering results are presented visually through graphs, using different colors to represent the three migration states (for example, red, blue, and green represent the three states respectively). The graphs can be used to visually observe the differences in temperature, humidity, light and other characteristics of the different states, which is convenient for researchers to further analyze and compare.

[0102] Specifically, Fig. 9 As shown, in step S6, a Markov chain model is used to perform statistical analysis on historical data of migratory birds to predict the number of migratory birds in the future.

[0103] Markov Process is an important method for studying the state space of discrete event dynamic systems. The future evolution of an event depends only on the current state of the event. The specific steps of prediction are as follows:

[0104] S61. According to the migratory bird data processed in step S4, the data is divided into a finite number of state sets as state variables of the system.

[0105] S62, calculation state s i The initial probability of occurrence is calculated as follows: Among them, n(s i ) is in state s i The number of data, N is the total number of data.

[0106] S63, establish a single-step transition probability matrix: The single-step transition probability matrix represents the probability of transitioning from one state to another state, and the calculation formula is: Among them, P(s j │si ) is from state s i Transfer to state s j The probability of n(s i →s j ) is the number of state transitions, n(s i ) is state s i The total number of occurrences.

[0107] S64, constructing a step transfer probability matrix: Based on the single-step transfer probability matrix, the multi-step transfer probability matrix can be further calculated by the recursive formula: P (k) =P (k-1) ·P, where P (k) is the k-step transition probability matrix, and P is the single-step transition probability matrix.

[0108] S65. Predict future status: According to the state transition formula of the Markov chain model, predict the future migratory bird data: P t+k =P t ·P (k) , where P t+k is the predicted state distribution, P t is the current state distribution, P (k) is the k-step transition probability matrix.

[0109] An embodiment of the present invention provides an application of a migratory bird monitoring method based on deep learning, which is applied to migratory bird protection and ecological environment monitoring, and is used for migratory bird migration monitoring and habitat protection, so as to provide real-time monitoring data on migration patterns, population sizes and environmental parameters, and support the formulation of relevant ecological protection measures.

[0110] It has the following significant application value and advantages:

[0111] The application value in migratory bird protection includes:

[0112] Dynamic monitoring of migratory bird populations: By acquiring real-time data on migratory bird activity through network cameras and combining the MPDIoU-ODConv-YOLOv5 model to identify species and numbers, the dynamic changes of migratory birds can be tracked quickly and accurately, providing data support for conservation work.

[0113] Optimization of migratory bird habitat protection: By combining the relationship between environmental data (such as temperature, humidity, and light intensity) and the distribution of migratory birds, cluster analysis can clarify the environmental characteristics of key habitats, thereby guiding the optimization and restoration of habitats.

[0114] Prediction of migration patterns: Based on the Markov chain state transition model, it is possible to predict the future migration trends of migratory birds in advance and formulate relevant protection policies, such as delineating safe flight corridors and avoiding dangerous areas.

[0115] Preventing potential threats: Long-term monitoring data can be used to identify abnormal migration behaviors, such as population declines caused by climate change, habitat destruction or hunting, providing a basis for early intervention.

[0116] The application value in ecological environment monitoring includes:

[0117] Environmental health assessment: The number and migration behavior of migratory birds are important indicators of the ecological environment. By extracting reduced-dimensional environmental features (such as light intensity and humidity) and analyzing their impact on migratory bird behavior, the health of the regional ecosystem can be assessed.

[0118] Climate change response analysis: Based on long-term migratory bird data and changing trends in environmental parameters, it can reveal the impact of climate change on migratory bird behavior and habitat selection, and provide a reference for ecological strategies to cope with global climate change.

[0119] Ecological restoration effect evaluation: After ecological restoration is carried out in a certain area, this technology can monitor the recovery of migratory bird populations and behavioral changes in real time to verify the effectiveness of the restoration work.

[0120] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A migratory bird monitoring method based on deep learning, characterized in that: The following steps are involved: S1. Initialization of hardware equipment and data collection: Turn on the hardware equipment, connect the network camera, related sensors and GPS positioning device, obtain real-time video data of migratory bird activities through the network camera, and collect environmental parameters, including temperature, humidity, light intensity and geographic location information; S2, data preprocessing and denoising: perform format conversion and frame extraction processing on the collected video data, and use Gaussian weighted moving average filtering to denoise the original data; S3. Migratory bird image recognition and distance estimation: Use the self-developed MPDIoU-ODConv-YOLOv5 model to identify and extract the species and number of migratory birds in the denoised video, calculate the relative size of the migratory birds based on the average size of the detection frame, and estimate the distance between the migratory birds and the camera based on the pinhole imaging principle; S4, data storage, segment analysis and feature extraction: the identified migratory bird species, quantity and environmental monitoring data are stored in the database, and features are extracted by time segment. The average temperature, average light and maximum humidity are selected as features. The feature data are reduced in dimension by principal component analysis, and a subspace projection matrix is ​​constructed to reduce the data dimension to form a reduced dimension data set. S5. Analysis of migratory bird migration behavior: Based on the reduced-dimensional data, the k-means algorithm was used to divide the migratory status of migratory birds into three categories, and the characteristics of cluster centers were analyzed to identify the migration patterns and their influencing factors; S6. Prediction of migratory bird migration trends: Based on the historical data of migratory birds, a state transition matrix based on Markov chain is constructed to quantify the transition rules between migration states. The multi-step transition matrix is ​​used to predict the changing trends of migratory bird species and numbers in the future, providing support for the analysis of migratory bird migration rules. S7. Data visualization display: Display the data generated by steps S1-S6 through a Web platform.

2. A migratory bird monitoring method based on deep learning according to claim 1, characterized in that: In step S2, the specific method steps for denoising the original data by using Gaussian weighted moving average filtering are as follows: S21, determine the filter window size N and standard deviation σ, calculate the Gaussian weight w(i) and perform normalization processing to ensure that the sum of the weights satisfies ∑w(i)=1, where, i is the relative position of the sample, σ is the standard deviation of the Gaussian distribution; S22, the filter window slides to cover the original data, and the smoothing value is calculated using the Gaussian weight. The smoothing value calculation formula is: Where y[k] is the kth value of the output sequence, w[i] is the i-th Gaussian weighting coefficient, and x[k-(N-1) / 2+i] is the kth value in the original sequence; S23, using data filling strategies to avoid filtering errors at both ends of the sequence, including repeating the first and last elements of the sequence or performing period extension; S24. Calculate the mean absolute error (MAE) and the root mean square error (RSME) to evaluate the denoising effect and ensure that the smoothed data accurately reflects the actual environment. The relevant calculation formula is: Among them, xi is the processed data and yi is the original data.

3. The migratory bird monitoring method based on deep learning according to claim 1, characterized in that: In step S3, the self-developed MPDIoU-ODConv-YOLOv5 model is obtained by improving the classic YOLOv5 target detection framework. The specific improvement method steps are as follows: S31. Replace the original target box regression loss function GIoU in YOLOv5 with MPDIoU, introduce the maximum pooling operation, and optimize the model's ability to capture the distance and position dislocation between target boxes. The calculation formula of MPDIoU is: Among them, |A∩B| is the intersection area of ​​the predicted box A and the real box B, |A∪B| is the union area of ​​the predicted box A and the real box B, d1 and d2 are the Euclidean distances between the center points of the predicted box A and the real box B, λ1 and λ2 are dynamically adjusted weight factors, and w and h are the width and height of the joint box; S32. Use the full-dimensional dynamic convolution ODConv module to replace the conventional convolution in YOLOv5, and optimize the dynamic weight generation in combination with the attention mechanism to improve the model's adaptability to the diversity characteristics of migratory birds. The calculation formula of ODConv is: Among them, a i is the dynamic weight associated with the input data x, W i is the i-th convolution kernel, x and y are the input and output feature maps respectively; S33, replace the Backbone of YOLOv5 with the lightweight network ob_mbilenetv2_075 that integrates ODConv and Ghost modules. By optimizing the fusion of multi-scale features, the computational complexity is reduced while enhancing the model's adaptability to targets of different sizes, thereby improving detection accuracy and efficiency. S34. Integrate MPDIoU and ODConv modules in the training process of YOLOv5, optimize the loss function through multiple rounds of iterations, and improve the target detection performance.

4. The migratory bird monitoring method based on deep learning according to claim 1, characterized in that: In step S4, the characteristic data is reduced in dimension by principal component analysis, and a subspace projection matrix is ​​constructed to reduce the data dimension. The specific method steps for forming a reduced dimension data set are as follows: S41. Calculate sample average vector S42. Subtract the mean of the corresponding feature from each feature value in the data set to complete the data centering process. S43. Calculate the covariance matrix Σ of the centralized data set. S44, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​λ1,λ2,…,λ n and the corresponding eigenvectors v1,v2,…,v n , according to the size of the eigenvalue, select the eigenvectors corresponding to the first k eigenvalues ​​as the principal components to form a projection matrix W; S45. Project the data set into the subspace formed by the first k eigenvectors: X reduced =X centered W, thereby achieving data dimensionality reduction.

5. The migratory bird monitoring method based on deep learning according to claim 1, characterized in that: The specific steps of step S5 are as follows: S51, obtaining the species and number of migratory birds in the original data from step S3, and obtaining the main environmental characteristics after dimensionality reduction from step S4; S52, selecting the k-means clustering algorithm as the analysis tool, setting the number of clusters to k=3, and dividing the migratory status of migratory birds into three categories; S53. Analyze the characteristics of cluster centers in the clustering results, clarify the characteristics of the three migration states, and identify the key factors affecting the migration behavior of migratory birds; S54. The clustering results are displayed visually through charts, and different colors are used to represent the three types of migration status for easy comparison and analysis.

6. The migratory bird monitoring method based on deep learning according to claim 1, characterized in that: The specific steps of step S6 are as follows: S61, according to the migratory bird data processed in step S4, dividing the processed data into a finite number of state sets as state variables of the system; S62, calculation state s i The initial probability of occurrence is calculated as: Among them, n(s i ) is in state s i The number of data, N is the total number of data; S63, establish a single-step transition probability matrix: The single-step transition probability matrix represents the probability of transitioning from one state to another state, and the calculation formula is: Among them, P(s j │s i ) is from state s i Transfer to state s j The probability of n(s i →s j ) is the number of state transitions, n(s i ) is state s i Total number of occurrences; S64, constructing a step transfer probability matrix: Based on the single-step transfer probability matrix, the multi-step transfer probability matrix can be further calculated by the recursive formula: P (k) =P (k-1) ·P, where P (k) is the k-step transition probability matrix, and P is the single-step transition probability matrix; S65. Predict future status: According to the state transition formula of the Markov chain model, predict the future migratory bird data: P t+k =P t ·P (k) , where P t+k is the predicted state distribution, P t is the current state distribution, P (k) is the k-step transition probability matrix.

7. A migratory bird monitoring method based on deep learning according to any one of claims 1 to 6, characterized in that: The method can be applied to migratory bird protection and ecological environment monitoring, and can be used to monitor migratory bird migration patterns, population numbers and habitat environmental parameters in real time, thereby providing data support for migration monitoring and habitat protection.

Images