A flying biological recognition method based on natural reserve multi-source data

Abstract

This invention relates to the field of aerial biometrics in nature reserves, and discloses an aerial biometrics method based on multi-source data from nature reserves, comprising: S1, acquiring and preprocessing multi-source data to obtain preprocessed multi-source data; S2, processing the preprocessed multi-source data and generating multi-source data feature vectors; S3, constructing an aerial biometrics target model; and S4, inputting the multi-source data feature vectors into the aerial biometrics target model for automatic identification and outputting the aerial biometrics identification results. This invention achieves comprehensive and all-time-period accurate monitoring, overcoming the limitations of single-source data. By integrating multi-source data, it achieves integrated "air-space-ground-sound" monitoring, capturing macro-environmental changes in aerial habitats and accurately identifying individual aerial organisms, populations, and behavioral states. Simultaneously, it enables continuous observation of fixed-point areas, solving the problems of existing methods such as single monitoring dimensions, incomplete time periods, and insufficient detail.

Description

Technical Field

[0001] This invention relates to the field of flight biometrics in nature reserves, and more particularly to a flight biometrics method based on multi-source data from nature reserves, applicable to the identification of various types of flying organisms such as birds and insects within nature reserves. Background Technology

[0002] Nature reserves (including national nature reserves, wetland parks, forest parks, etc.) are core areas for biodiversity conservation, an important component of the nature reserve system, and a vital vehicle for promoting ecological civilization and building a beautiful China. They play an important role in conserving water resources, maintaining soil, preventing wind erosion and sandstorms, regulating climate, and protecting rare and endemic species, typical ecosystems, and precious natural relics.

[0003] Establishing nature reserves is an effective measure to protect the ecological environment and natural resources, an important vehicle for protecting biodiversity and building ecological civilization, one of the most direct and effective ways to protect biodiversity, and a positive means to accelerate the transformation of economic development patterns and achieve sustainable development. Improving the scientific rigor and accuracy of decision-making in forestry nature reserve management is crucial. Establishing a comprehensive, interconnected, efficient, convenient, stable, and secure information system for forestry nature reserves, and leveraging advanced information technology to promote the construction of modern forestry nature reserves, is of great significance to promoting ecological civilization construction in my country.

[0004] Nature reserves are important habitats and stopover points for flying creatures (migratory birds, rare and protected birds, etc.), and the observation of flying creatures is the core foundation for biodiversity conservation, ecological environment assessment, and conservation of rare species.

[0005] In the observation of aerobatic creatures in nature reserves, single observation data is insufficient for comprehensive and high-precision monitoring. Satellite remote sensing enables large-scale, all-weather macroscopic monitoring, capturing the overall distribution of aerobatic creature habitats and changes in the ecological environment, providing macroscopic support for the delineation of aerobatic species migration routes and habitat ranges. UAV imagery enables close-range, high-precision microscopic monitoring, capturing the individual morphology of aerobatic creatures and the aggregation of small-scale populations, supplementing the lack of detail in satellite remote sensing and complementing radar and camera data from aerobatic creature observation devices. Fixed cameras enable fixed-point, continuous, and routine observation, accurately capturing the activity patterns and behavioral details of aerobatic creatures in specific areas (such as habitats, foraging areas, and migratory resting points), compensating for the limitations of "instantaneous observation" in satellite remote sensing and UAV imagery. Acoustic features, as unique identifiers of aerobatic creatures (especially birds and insects), enable rapid identification of aerobatic species and judgment of behavioral states (such as breeding calls and warning calls), compensating for the limitations of image observation at night and in scenarios with complex vegetation obstruction.

[0006] Currently, data processing methods for aerial wildlife observation in nature reserves have many limitations and are difficult to adapt to the specific needs of aerial wildlife observation: First, existing processing methods mostly process satellite remote sensing images or UAV images alone, failing to achieve the synergistic fusion of these three types of images with voiceprint features and fixed camera images. This fails to leverage the complementary advantages of "macro + micro + fixed-point + voiceprint" and makes it difficult to achieve multi-dimensional, all-time accurate monitoring of aerial wildlife. Second, image processing does not combine the core needs of aerial wildlife observation, failing to accurately extract key information related to aerial wildlife (such as habitat vegetation growth, population aggregation areas, and flight activity hotspots). Furthermore, the processing of UAV images and fixed camera images does not take into account low interference. The observation requirements are as follows: First, the denoising and enhancement algorithms are not adapted to the characteristics of their respective interferences (drone airflow vibration, fixed camera lighting changes, vegetation obstruction, and blurred motion of flying organisms); second, the voiceprint feature processing is not adapted to the complex environment of nature reserves, and is easily interfered with by environmental noise such as wind, water flow, and vegetation friction, resulting in low voiceprint recognition accuracy and failure to link with various image data, making it impossible to achieve accurate species identification through "sound-image matching"; third, the processing results of various data are disconnected, and collaborative analysis with flying organism observation data (such as wingbeat frequency and flight trajectory) is not achieved, resulting in insufficient practicality; in addition, the existing processing methods are inefficient and cannot meet the rapid monitoring needs of key periods such as the migration and breeding seasons of flying organisms.

[0007] Therefore, we need to develop a recognition scenario that is suitable for flying creatures in nature reserves, taking into account the collaborative processing of satellite remote sensing images, UAV images, fixed camera images and voiceprint features. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of the prior art by providing a flight biometric identification method based on multi-source data from nature reserves. This method aims to overcome the limitations of single-data identification in existing methods, as well as the deficiencies of traditional monitoring methods such as single monitoring dimensions, incomplete time periods, and insufficient details, thereby improving the accuracy and precision of flight biometric identification.

[0009] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0010] A flight biometric identification method based on multi-source data from nature reserves includes the following steps:

[0011] S1. Multi-source data acquisition and preprocessing to obtain preprocessed multi-source data;

[0012] Acquire multi-source data, including satellite remote sensing images, UAV images, fixed camera images, and voiceprint features of accompanying flying organisms observed in nature reserves;

[0013] Preprocessing of multi-source data yields preprocessed multi-source data, which includes preprocessed satellite remote sensing images, preprocessed UAV images, preprocessed fixed camera images, and preprocessed voiceprint features.

[0014] S2. Process the preprocessed multi-source data and generate multi-source data feature vectors;

[0015] Preprocessed satellite remote sensing images, preprocessed UAV images, preprocessed fixed camera images, and preprocessed voiceprint features are processed separately to extract key information related to flight biological observation from various types of data, forming feature vectors for satellite remote sensing images, UAV images, fixed camera images, and voiceprint feature recognition and behavior judgment.

[0016] S3. Construct a flight biometric target model;

[0017] An initial flight biometric model is constructed, a training dataset is built, and the training dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The initial flight biometric model is trained on the training dataset until the loss function converges, and the target flight biometric model is obtained.

[0018] S4. Flight biometrics based on flight biometric target model;

[0019] The multi-source data feature vectors are input into the flight biometric target model for automatic identification, and the flight biometric results are output.

[0020] Preferably, step S1 includes the following steps:

[0021] S11. Acquisition and preprocessing of satellite remote sensing images:

[0022] Acquisition of satellite remote sensing imagery:

[0023] Multispectral and high-resolution panchromatic images of the entire nature reserve and its surrounding areas are acquired through a satellite remote sensing platform. The image resolution is no less than 10 meters, and the corresponding metadata is acquired simultaneously.

[0024] Preprocessing of satellite remote sensing images includes:

[0025] Based on the image quality evaluation indicators, valid images are selected, and invalid images with excessive cloud cover, excessive shadow coverage, or severe noise are removed.

[0026] The selected valid images are subjected to radiometric correction, geometric correction, and atmospheric correction in sequence. The radiative transfer model is used to eliminate the influence of sensor error, solar altitude angle, and topographic relief, and the image gray values ​​are converted into the true surface reflectance to ensure the radiometric consistency of images in different time periods and regions. Based on the high-precision geographic coordinates of the nature reserve, a polynomial fitting algorithm is used for geometric correction, and finally the preprocessed satellite remote sensing images are obtained.

[0027] S12. Acquisition and preprocessing of UAV imagery:

[0028] Acquisition of drone imagery:

[0029] The drone imagery of the biological activity area was acquired by the low-noise drone observation unit, including visible light images with a resolution of no less than 4K and thermal images with a resolution of no less than 1280×1024.

[0030] Preprocessing of UAV imagery:

[0031] Invalid images caused by airflow turbulence, shooting angle deviation, and blurred motion of flying creatures are removed, and valid images that are clear, without obvious distortion, and can clearly identify individual flying creatures or groups are retained.

[0032] Distortion correction, shading, and geometric registration are performed on the effective images; the camera intrinsic parameter correction algorithm is used to eliminate lens distortion, and the motion blur recovery algorithm is used to correct the image blur caused by airflow turbulence and restore the details of individual flying organisms; the UAV images are geometrically registered with the geographic coordinates of satellite remote sensing images as a reference to ensure accurate spatial alignment, and finally the pre-processed UAV images are obtained.

[0033] S13. Acquisition and preprocessing of images from a fixed camera:

[0034] Acquiring images from a fixed camera:

[0035] A fixed camera is used to photograph flying creatures and generate images of the fixed camera, while simultaneously recording the geographic coordinates, shooting time, light intensity, temperature and humidity of the images;

[0036] Preprocessing of images from a fixed camera:

[0037] Remove invalid images caused by poor lighting, blurry lenses, or obstructed shooting angles, and retain valid images that are clear, without obvious distortion, and can identify individual flying creatures, populations, or behavioral details.

[0038] The effective images undergo lighting correction, noise reduction, distortion correction, and geometric registration. An adaptive lighting equalization algorithm is used to correct brightness differences in images under different time periods and lighting conditions. A Gaussian filtering algorithm is used to remove image noise and preserve details of individual flying organisms. A camera intrinsic parameter correction algorithm is used to eliminate lens distortion and restore the true shape of the image. Using the geographic coordinates of satellite remote sensing images as a reference, geometric registration is performed on fixed camera images to accurately align the spatial positions of satellite remote sensing images and UAV images, ultimately obtaining preprocessed fixed camera images.

[0039] S14. Acquisition and preprocessing of voiceprint features:

[0040] Acquisition of voiceprint features:

[0041] Voiceprint features are obtained using a voiceprint machine;

[0042] Preprocessing of voiceprint features:

[0043] Remove invalid voiceprint data containing strong environmental noise and retain valid voiceprint data with a signal-to-noise ratio ≥20dB;

[0044] The effective voiceprint data is denoised, normalized, and truncated. A wavelet threshold denoising algorithm is used to remove environmental noise and retain the core features of the voiceprint of the flying organism. The voiceprint data is normalized to adjust the amplitude to the same range, eliminating the influence of the acquisition equipment and acquisition distance, and finally obtaining the preprocessed voiceprint features.

[0045] Preferably, step S2 includes the following steps:

[0046] S21. Processing of pre-processed satellite remote sensing images:

[0047] S211, Denoising and Enhancement;

[0048] A hierarchical denoising algorithm is adopted to denoise the images of nature reserves by region and type. An improved median filtering algorithm is used to remove atmospheric noise and sensor noise, while retaining vegetation and water body edge information to avoid image blurring. Histogram equalization combined with band fusion enhancement algorithm is used to enhance the contrast between vegetation, water body and background, highlighting the vegetation growth and water body boundary information of the habitat of flying organisms.

[0049] S212, Image Fusion;

[0050] The SIFT feature matching algorithm is used to accurately register the S211-processed multispectral image with the high-resolution panchromatic image, so that the spatial positions are completely aligned; then, the wavelet transform fusion algorithm is used to deeply fuse the spatial detail information of the high-resolution panchromatic image with the spectral information of the multispectral image.

[0051] S213, Information Extraction;

[0052] Key information is extracted from the data processed by S212, and satellite remote sensing image feature vectors are output.

[0053] S22. Processing of pre-processed UAV images:

[0054] S221, Denoising and Enhancement;

[0055] To address the interference characteristics of UAV imagery, an adaptive filtering algorithm is used for noise reduction, and a sharpening enhancement algorithm is employed to improve the clarity of flight organism morphology, feather features, and population aggregation status. Temperature threshold segmentation technology is used to process thermal imaging images.

[0056] S222, Information Extraction;

[0057] Using the YOLO8 target detection algorithm, we accurately extract individual flying organisms and population aggregation areas from UAV visible light and thermal images; we statistically analyze the population size and individual size parameters of flying organisms, and combine them with UAV flight parameters to calculate the flight altitude and speed of flying organisms to assist in the analysis of flight trajectories; for migratory bird populations, we extract population aggregation density and distribution range, mark flight activity hotspots, and output UAV image feature vectors.

[0058] S23. Preprocessing of fixed camera images:

[0059] S231, Denoising and Enhancement;

[0060] To address the interference characteristics of fixed camera images, an adaptive Gaussian filtering algorithm is used for noise reduction, preserving individual details and behavioral characteristics of flying organisms. Sharpening and contrast adjustment algorithms are employed to enhance the clarity of the morphological and behavioral details of flying organisms and correct brightness deviations in nighttime infrared images.

[0061] S232, Information Extraction;

[0062] An improved target detection and behavior recognition algorithm is used to extract information on individual and population flying organisms from visible light and infrared images from fixed cameras, and to count the activity frequency and dwell time of flying organisms in specific areas. The behavior status of flying organisms is accurately identified, and the activity patterns of flying organisms are analyzed by combining the shooting time series. Detailed features of individual flying organisms are extracted, and combined with activity trajectories and behavioral details, a fixed-point observation dataset is formed, and fixed-camera image feature vectors are output.

[0063] S24. Preprocessing of voiceprint features:

[0064] S241. Voiceprint feature extraction;

[0065] By using Mel-frequency cepstral coefficients (MFCC) combined with short-time Fourier transform, the temporal and frequency domain features of the flight biological voiceprint are extracted from the effective segments of the preprocessed voiceprint features in S14, forming a standardized feature vector for recognition and behavior judgment.

[0066] S242, Voiceprint recognition;

[0067] A database of voiceprint features of flying organisms in nature reserves was constructed, including voiceprint features, species information, and protection levels of common flying organisms. The Support Vector Machine (SVM) algorithm was used to match the extracted and labeled recognition and behavior judgment feature vectors with the features in the database to identify flying organism species with an accuracy of no less than 90%. Voiceprints that did not match were marked as unknown species and their feature parameters were recorded.

[0068] S243, Behavioral Judgment;

[0069] Based on changes in voiceprint features, the behavioral state of flying creatures can be determined, such as breeding calls, warning calls, and group communication, providing support for the analysis of flying creature behavior. At the same time, it is linked and verified with the behavioral features extracted from fixed camera images, and outputs voiceprint feature recognition and behavior judgment vectors.

[0070] Preferably, the flight biometric model includes:

[0071] Data Input Layer: The foundational layer of the model, used to receive feature vectors from satellite remote sensing images, UAV images, fixed camera images, voiceprint recognition and behavior judgment vectors. It also imports geographic coordinates, timestamps, and environmental parameter information corresponding to various types of data, and performs standardization and normalization processing on the input data to ensure that the feature vectors have uniform dimensions and consistent data format.

[0072] Spatial alignment layer: The core support layer of the model; Based on the WGS84 geographic coordinates of satellite remote sensing imagery, combined with geographic coordinate information from UAV imagery, fixed camera images, and voiceprint data, a coordinate mapping algorithm is used for spatial alignment. Spatial deviation is eliminated through iterative optimization, so that the alignment error is ≤0.5 meters.

[0073] Feature Fusion Layer: The core fusion layer of the model, it adopts an improved weighted fusion algorithm combined with an attention mechanism to achieve deep fusion of four feature vectors. Through the attention mechanism, it automatically identifies the importance of various features, dynamically allocates fusion weights, and performs feature splicing and dimensional fusion on various feature vectors during the fusion process to construct multi-dimensional observation feature vectors, thereby achieving complementary fusion of macroscopic information, microscopic information, fixed-point information, and voiceprint information.

[0074] Verification and Optimization Layer: The output optimization layer of the model is used to perform data verification, anomaly removal, and accuracy optimization on the fused feature vector. It combines radar data from the aerial biological observation device in the nature reserve to perform multi-dimensional verification on the fused feature vector and correct the extraction bias of various data sources. Anomaly detection algorithms are used to remove abnormal data. Finally, the verified fused feature vector is standardized and the results are output.

[0075] Preferably, the processing steps of the data input layer include:

[0076] Unified feature dimensions: Satellite remote sensing image feature vectors, UAV image feature vectors, fixed camera image feature vectors, voiceprint feature recognition and behavior judgment vectors are uniformly mapped to a 64-dimensional feature space through a feature mapping algorithm;

[0077] Data normalization: The Min-Max normalization algorithm is used to uniformly adjust the numerical range of various feature vectors to the [0,1] interval, eliminating the influence of different data source acquisition devices and acquisition conditions;

[0078] The normalization formula is:

[0079] ;

[0080] Where X is the original feature value, This is the minimum value of the feature. The maximum value of this feature. These are the normalized eigenvalues;

[0081] Data cleaning: Perform preliminary outlier detection on the input data and remove missing values ​​and extreme outliers.

[0082] Preferably, the processing steps of the spatial alignment layer include:

[0083] Unified coordinate reference: The geographic coordinates of UAV imagery, fixed camera images, and voiceprint data are uniformly converted into the WGS84 geographic coordinate system used by satellite remote sensing imagery. The coordinate system conversion is completed by using a coordinate transformation algorithm, and the conversion error is ≤0.1 meters.

[0084] Feature point matching: Extract geographic feature points from satellite remote sensing images, UAV images, and fixed camera images, and use the SIFT feature matching algorithm to accurately match feature points from satellite remote sensing images, UAV images, and fixed camera images, and to make satellite remote sensing images, UAV images, and fixed camera images spatially correspond to satellite remote sensing images.

[0085] Spatial deviation correction: Combine the geographic coordinates of the voiceprint data to calculate the spatial deviation between each data source and the satellite remote sensing image reference. The iterative least squares method is used to correct the deviation, with the number of iterations set to 10-15 times, until the spatial alignment error is ≤0.5 meters.

[0086] Time synchronization: Combine the timestamp of each data point to synchronize the time of the spatially aligned multi-source data, so that the multi-source data in the same observation period correspond and match.

[0087] Preferably, the processing steps of the feature fusion layer include:

[0088] Attention mechanism design: Channel weights are assigned to the standardized feature vectors to automatically identify the core channels relevant to flight biological observation in each feature. Attention weights are assigned to the core channels and the ineffective channels respectively. The attention weights are calculated using the Sigmoid activation function, and the formula is as follows:

[0089] ;

[0090] in, Let i be the attention weight for the i-th channel. Let N be the j-th feature value of the i-th channel, and N be the number of features. Use the Sigmoid activation function;

[0091] Dynamic weighted fusion: Combining attention weights, a dynamic weighted fusion algorithm is used to fuse all feature vectors. The fusion formula is as follows:

[0092] ;

[0093] Where F is the fused feature vector. For satellite remote sensing image feature vectors, For UAV image feature vectors, To fix the feature vector of the camera image, This is a vector for voiceprint feature recognition and behavior judgment. , , , For dynamic weights, + + + = 1;

[0094] Adaptive weight adjustment: manually adjusting weights , , , ;

[0095] Feature fusion optimization: The fused feature vectors are smoothed by convolution using a 3×3 convolution kernel to eliminate feature noise generated during the fusion process.

[0096] Preferably, the processing steps of the verification optimization layer include:

[0097] Multi-dimensional data verification: Combining radar data from the aerial biological observation device in the nature reserve, the flight parameters, population size, and behavioral status in the fused feature vector are verified; the deviation between the fused result and the radar data is calculated. If the deviation exceeds 5%, the weights of the corresponding feature vectors are adjusted and the data is re-fused until the deviation is ≤5%; at the same time, the voiceprint recognition results and the species recognition results from the fixed camera images are cross-validated. If the recognition results are inconsistent, the result with the higher confidence level is used to correct the fused result.

[0098] Outlier removal: The Isolation Forest algorithm is used to detect anomalies in the fused feature vectors, identify and remove outlier data; the anomaly detection threshold is set to m, and when the anomaly confidence is ≥ m, it is determined to be outlier data and removed.

[0099] Accuracy Optimization: The accuracy of the fused feature vector after verification is optimized using a gradient descent algorithm to minimize the error between the fused result and the actual observed data. The fusion parameters are iteratively optimized to improve the fusion accuracy. The objective function is:

[0100] ;

[0101] in, The error loss is M, and the amount of fused data is M. For the k-th fusion feature value, This is the k-th true observation;

[0102] Standardized output: The optimized fusion feature vector is converted into a standardized multi-dimensional fusion dataset, which is organized according to the nature reserve zones. It includes the species of flying organisms, protection level, population size, flight trajectory, behavioral status, fixed-point activity patterns, and habitat ecological environment parameters.

[0103] Preferably, step S4 includes the following steps:

[0104] S41. Input the feature vectors of multi-source data into the flight biometric target model for automatic identification, and output the flight biometric results;

[0105] S42. Analyze the results of flight biometric identification;

[0106] S421. Population dynamics analysis;

[0107] By combining time-series fusion data, we can analyze changes in the population size, migration routes, and aggregation areas of flying organisms, monitor the population migration dynamics during the migratory season of migratory birds, and monitor the population size changes of rare species during the breeding season; by combining fixed-point observation data from fixed cameras, we can analyze the activity patterns of flying organisms.

[0108] S422, Habitat quality assessment;

[0109] By combining habitat ecological environment parameters extracted from satellite remote sensing images, habitat details extracted from UAV images and fixed camera images, the quality level of the habitat of flying organisms is assessed, and the impact of habitat change on the population and behavior of flying organisms is analyzed.

[0110] S423, Collaborative and Interactive Applications;

[0111] By linking the fusion results with the data from the aerial biological observation device in the nature reserve, integrated monitoring of "air-space-ground-sound" can be achieved. The terminal device can be used to view multi-dimensional information of aerial organisms in real time and remotely adjust the observation parameters.

[0112] S424. Data storage and traceability;

[0113] The fusion results, each feature vector, and the original data are stored synchronously, supporting data export, backup, and traceability, with a focus on preserving relevant data on rare and protected flying organisms.

[0114] The beneficial effects of this invention are as follows:

[0115] 1. Multi-source data synergy and fusion: This invention achieves comprehensive and accurate monitoring at all times. It breaks through the limitations of single data and achieves integrated monitoring of "air-space-ground-sound" by integrating satellite remote sensing images, UAV images, fixed camera images and voiceprint features. It can capture macro-environmental changes in the habitat of flying organisms, accurately identify individual flying organisms, populations and behavioral states, and achieve continuous observation of fixed areas. It solves the problems of single monitoring dimension, incomplete time period and insufficient details of existing methods.

[0116] 2. This invention has strong adaptability and high accuracy: It adopts a modular design, introduces an attention mechanism and a dynamic weighting algorithm, and can adaptively adjust parameters according to the observation scenario. Combined with multi-dimensional data verification and accuracy optimization, it effectively improves the fusion accuracy, solves the defects of low accuracy and poor adaptability of existing fusion technologies, and ensures the accuracy and practicality of the fusion results.

[0117] 3. This invention is adaptable to zoning management and has strong versatility: Combining the zoning management requirements of nature reserves, it performs targeted processing and fusion of multi-source data from different regions. The fusion model can flexibly adjust parameters and is suitable for aerial biological observation in various types of nature reserves, such as national and provincial nature reserves, and different types of reserves such as forest land, wetland, and mountain. The parameters can be flexibly adjusted to adapt to different observation targets, such as birds and insects. Attached Figure Description

[0118] The invention will now be further described with reference to the accompanying drawings.

[0119] Figure 1 This is a system diagram of the flight biometric method in the embodiments of this application.

[0120] Figure 2 This is a schematic diagram of the flight biometric recognition results in this embodiment.

[0121] Figure 3 This is a schematic diagram of the flight biometric recognition results in this embodiment. Detailed Implementation

[0122] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0123] 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.

[0124] In the description of this invention, it should be noted that, 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 this invention based on the specific circumstances.

[0125] Reference Figure 1 This application discloses a flight biometric identification method based on multi-source data from nature reserves, specifically including the following steps:

[0126] S1. Multi-source data acquisition and preprocessing to obtain preprocessed multi-source data;

[0127] Acquire multi-source data, including satellite remote sensing images, UAV images, fixed camera images, and voiceprint features of accompanying flying organisms observed in nature reserves;

[0128] Preprocessing of multi-source data yields preprocessed multi-source data, which includes preprocessed satellite remote sensing images, preprocessed UAV images, preprocessed fixed camera images, and preprocessed voiceprint features.

[0129] Step S1 specifically includes the following steps:

[0130] S11. Acquisition and preprocessing of satellite remote sensing images:

[0131] Acquisition of satellite remote sensing imagery:

[0132] Multispectral and high-resolution panchromatic images of the entire nature reserve and its surrounding areas are acquired through a satellite remote sensing platform. The image resolution is no less than 10 meters, and the corresponding metadata is acquired simultaneously.

[0133] Preprocessing of satellite remote sensing images includes:

[0134] Based on the image quality evaluation indicators, valid images are selected, and invalid images with excessive cloud cover, excessive shadow coverage, or severe noise are removed.

[0135] The selected valid images are subjected to radiometric correction, geometric correction, and atmospheric correction in sequence. The radiative transfer model is used to eliminate the influence of sensor error, solar altitude angle, and topographic relief, and the image gray values ​​are converted into the true surface reflectance to ensure the radiometric consistency of images in different time periods and regions. Based on the high-precision geographic coordinates of the nature reserve, a polynomial fitting algorithm is used for geometric correction, and finally the preprocessed satellite remote sensing images are obtained.

[0136] S12. Acquisition and preprocessing of UAV imagery:

[0137] Acquisition of drone imagery:

[0138] The drone imagery of the biological activity area was acquired by the low-noise drone observation unit, including visible light images with a resolution of no less than 4K and thermal images with a resolution of no less than 1280×1024.

[0139] Preprocessing of UAV imagery:

[0140] Invalid images caused by airflow turbulence, shooting angle deviation, and blurred motion of flying creatures are removed, and valid images that are clear, without obvious distortion, and can clearly identify individual flying creatures or groups are retained.

[0141] Distortion correction, shading, and geometric registration are performed on the effective images; the camera intrinsic parameter correction algorithm is used to eliminate lens distortion, and the motion blur recovery algorithm is used to correct the image blur caused by airflow turbulence and restore the details of individual flying organisms; the UAV images are geometrically registered with the geographic coordinates of satellite remote sensing images as a reference to ensure accurate spatial alignment, and finally the pre-processed UAV images are obtained.

[0142] S13. Acquisition and preprocessing of images from a fixed camera:

[0143] Acquiring images from a fixed camera:

[0144] A fixed camera is used to photograph flying creatures and generate images of the fixed camera, while simultaneously recording the geographic coordinates, shooting time, light intensity, temperature and humidity of the images;

[0145] Preprocessing of images from a fixed camera:

[0146] Remove invalid images caused by poor lighting, blurry lenses, or obstructed shooting angles, and retain valid images that are clear, without obvious distortion, and can identify individual flying creatures, populations, or behavioral details.

[0147] The effective images undergo lighting correction, noise reduction, distortion correction, and geometric registration. An adaptive lighting equalization algorithm is used to correct brightness differences in images under different time periods and lighting conditions. A Gaussian filtering algorithm is used to remove image noise and preserve details of individual flying organisms. A camera intrinsic parameter correction algorithm is used to eliminate lens distortion and restore the true shape of the image. Using the geographic coordinates of satellite remote sensing images as a reference, geometric registration is performed on fixed camera images to accurately align the spatial positions of satellite remote sensing images and UAV images, ultimately obtaining preprocessed fixed camera images.

[0148] S14. Acquisition and preprocessing of voiceprint features:

[0149] Acquisition of voiceprint features:

[0150] Voiceprint features are obtained using a voiceprint machine;

[0151] Preprocessing of voiceprint features:

[0152] Remove invalid voiceprint data containing strong environmental noise and retain valid voiceprint data with a signal-to-noise ratio ≥20dB;

[0153] The effective voiceprint data is denoised, normalized, and truncated. A wavelet threshold denoising algorithm is used to remove environmental noise and retain the core features of the voiceprint of the flying organism. The voiceprint data is normalized to adjust the amplitude to the same range, eliminating the influence of the acquisition equipment and acquisition distance, and finally obtaining the preprocessed voiceprint features.

[0154] S2. Process the preprocessed multi-source data and generate multi-source data feature vectors;

[0155] Preprocessed satellite remote sensing images, preprocessed UAV images, preprocessed fixed camera images, and preprocessed voiceprint features are processed separately to extract key information related to flight biological observation from various types of data, forming feature vectors for satellite remote sensing images, UAV images, fixed camera images, and voiceprint feature recognition and behavior judgment.

[0156] Step S2 specifically includes the following steps:

[0157] S21. Processing of pre-processed satellite remote sensing images:

[0158] S211, Denoising and Enhancement;

[0159] A hierarchical denoising algorithm is adopted to denoise the images of nature reserves by region and type. An improved median filtering algorithm is used to remove atmospheric noise and sensor noise, while retaining vegetation and water body edge information to avoid image blurring. Histogram equalization combined with band fusion enhancement algorithm is used to enhance the contrast between vegetation, water body and background, highlighting the vegetation growth and water body boundary information of the habitat of flying organisms.

[0160] S212, Image Fusion;

[0161] The SIFT feature matching algorithm is used to accurately register the S211-processed multispectral image with the high-resolution panchromatic image, so that the spatial positions are completely aligned; then, the wavelet transform fusion algorithm is used to deeply fuse the spatial detail information of the high-resolution panchromatic image with the spectral information of the multispectral image.

[0162] S213, Information Extraction;

[0163] Key information is extracted from the data processed by S212, and satellite remote sensing image feature vectors are output.

[0164] S22. Processing of pre-processed UAV images:

[0165] S221, Denoising and Enhancement;

[0166] To address the interference characteristics of UAV imagery, an adaptive filtering algorithm is used for noise reduction, and a sharpening enhancement algorithm is employed to improve the clarity of flight organism morphology, feather features, and population aggregation status. Temperature threshold segmentation technology is used to process thermal imaging images.

[0167] S222, Information Extraction;

[0168] Using the YOLO8 target detection algorithm, we accurately extract individual flying organisms and population aggregation areas from UAV visible light and thermal images; we statistically analyze the population size and individual size parameters of flying organisms, and combine them with UAV flight parameters to calculate the flight altitude and speed of flying organisms to assist in the analysis of flight trajectories; for migratory bird populations, we extract population aggregation density and distribution range, mark flight activity hotspots, and output UAV image feature vectors.

[0169] S23. Preprocessing of fixed camera images:

[0170] S231, Denoising and Enhancement;

[0171] To address the interference characteristics of fixed camera images, an adaptive Gaussian filtering algorithm is used for noise reduction, preserving individual details and behavioral characteristics of flying organisms. Sharpening and contrast adjustment algorithms are employed to enhance the clarity of the morphological and behavioral details of flying organisms and correct brightness deviations in nighttime infrared images.

[0172] S232, Information Extraction;

[0173] An improved target detection and behavior recognition algorithm is used to extract information on individual and population flying organisms from visible light and infrared images from fixed cameras, and to count the activity frequency and dwell time of flying organisms in specific areas. The behavior status of flying organisms is accurately identified, and the activity patterns of flying organisms are analyzed by combining the shooting time series. Detailed features of individual flying organisms are extracted, and combined with activity trajectories and behavioral details, a fixed-point observation dataset is formed, and fixed-camera image feature vectors are output.

[0174] S24. Preprocessing of voiceprint features:

[0175] S241. Voiceprint feature extraction;

[0176] By using Mel-frequency cepstral coefficients (MFCC) combined with short-time Fourier transform, the temporal and frequency domain features of the flight biological voiceprint are extracted from the effective segments of the preprocessed voiceprint features in S14, forming a standardized feature vector for recognition and behavior judgment.

[0177] S242, Voiceprint recognition;

[0178] A database of acoustic signatures of flying organisms in nature reserves was constructed, including acoustic signatures, species information, and protection levels of common flying organisms. The Support Vector Machine (SVM) algorithm was used to match the extracted standardized recognition and behavior judgment feature vectors with the features in the database to identify flying organism species with an accuracy of no less than 90%. For acoustic signatures that did not match successfully, they were marked as unknown species and their feature parameters were recorded.

[0179] S243, Behavioral Judgment;

[0180] Based on changes in voiceprint features, the behavioral state of flying creatures can be determined, such as breeding calls, warning calls, and group communication, providing support for the analysis of flying creature behavior. At the same time, it is linked and verified with the behavioral features extracted from fixed camera images, and outputs voiceprint feature recognition and behavior judgment vectors.

[0181] S3. Construct a flight biometric target model;

[0182] An initial flight biometric model is constructed, a training dataset is built, and the training dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The initial flight biometric model is trained on the training dataset until the loss function converges, and the target flight biometric model is obtained.

[0183] S4. Flight biometrics based on flight biometric target model;

[0184] The multi-source data feature vectors are input into the flight biometric target model for automatic identification, and the flight biometric results are output.

[0185] S4 specifically includes the following steps:

[0186] S41. Input the feature vectors of multi-source data into the flight biometric target model for automatic identification, and output the flight biometric results;

[0187] S42. Analyze the results of flight biometric identification;

[0188] S421. Population dynamics analysis;

[0189] By combining time-series fusion data, we can analyze changes in the population size, migration routes, and aggregation areas of flying organisms, monitor the population migration dynamics during the migratory season of migratory birds, and monitor the population size changes of rare species during the breeding season; by combining fixed-point observation data from fixed cameras, we can analyze the activity patterns of flying organisms.

[0190] S422, Habitat quality assessment;

[0191] By combining habitat ecological environment parameters extracted from satellite remote sensing images, habitat details extracted from UAV images and fixed camera images, the quality level of the habitat of flying organisms is assessed, and the impact of habitat change on the population and behavior of flying organisms is analyzed.

[0192] S423, Collaborative and Interactive Applications;

[0193] By linking the fusion results with the data from the aerial biological observation device in the nature reserve, integrated monitoring of "air-space-ground-sound" can be achieved. The terminal device can be used to view multi-dimensional information of aerial organisms in real time and remotely adjust the observation parameters.

[0194] S424. Data storage and traceability;

[0195] The fusion results, each feature vector, and the original data are stored synchronously, supporting data export, backup, and traceability, with a focus on preserving relevant data on rare and protected flying organisms.

[0196] The flight biometric model includes:

[0197] Data Input Layer: The foundational layer of the model, used to receive feature vectors from satellite remote sensing images, UAV images, fixed camera images, voiceprint recognition and behavior judgment vectors. It also imports geographic coordinates, timestamps, and environmental parameter information corresponding to various types of data, and performs standardization and normalization processing on the input data to ensure that the feature vectors have uniform dimensions and consistent data format.

[0198] The data input layer processing steps include:

[0199] Unified feature dimensions: Satellite remote sensing image feature vectors, UAV image feature vectors, fixed camera image feature vectors, voiceprint feature recognition and behavior judgment vectors are uniformly mapped to a 64-dimensional feature space through a feature mapping algorithm;

[0200] Data normalization: The Min-Max normalization algorithm is used to uniformly adjust the numerical range of various feature vectors to the [0,1] interval, eliminating the influence of different data source acquisition devices and acquisition conditions;

[0201] The normalization formula is:

[0202] ;

[0203] Where X is the original feature value, This is the minimum value of the feature. The maximum value of this feature. These are the normalized eigenvalues;

[0204] Data cleaning: Perform preliminary outlier detection on the input data and remove missing values ​​and extreme outliers.

[0205] Spatial alignment layer: The core support layer of the model; Based on the WGS84 geographic coordinates of satellite remote sensing imagery, combined with geographic coordinate information from UAV imagery, fixed camera images, and voiceprint data, a coordinate mapping algorithm is used for spatial alignment. Spatial deviation is eliminated through iterative optimization, so that the alignment error is ≤0.5 meters.

[0206] The processing steps for the spatial alignment layer include:

[0207] Unified coordinate reference: The geographic coordinates of UAV imagery, fixed camera images, and voiceprint data are uniformly converted into the WGS84 geographic coordinate system used by satellite remote sensing imagery. The coordinate system conversion is completed by using a coordinate transformation algorithm, and the conversion error is ≤0.1 meters.

[0208] Feature point matching: Extract geographic feature points from satellite remote sensing images, UAV images, and fixed camera images, and use the SIFT feature matching algorithm to accurately match feature points from satellite remote sensing images, UAV images, and fixed camera images, and to make satellite remote sensing images, UAV images, and fixed camera images spatially correspond to satellite remote sensing images.

[0209] Spatial deviation correction: Combine the geographic coordinates of the voiceprint data to calculate the spatial deviation between each data source and the satellite remote sensing image reference. The iterative least squares method is used to correct the deviation, with the number of iterations set to 10-15 times, until the spatial alignment error is ≤0.5 meters.

[0210] Time synchronization: Combine the timestamp of each data point to synchronize the time of the spatially aligned multi-source data, so that the multi-source data in the same observation period correspond and match.

[0211] Feature Fusion Layer: The core fusion layer of the model, it adopts an improved weighted fusion algorithm combined with an attention mechanism to achieve deep fusion of four feature vectors. Through the attention mechanism, it automatically identifies the importance of various features, dynamically allocates fusion weights, and performs feature splicing and dimensional fusion on various feature vectors during the fusion process to construct multi-dimensional observation feature vectors, thereby achieving complementary fusion of macroscopic information, microscopic information, fixed-point information, and voiceprint information.

[0212] The processing steps of the feature fusion layer include:

[0213] Attention mechanism design: Channel weights are assigned to the standardized feature vectors to automatically identify the core channels relevant to flight biological observation in each feature. Attention weights are assigned to the core channels and the ineffective channels respectively. The attention weights are calculated using the Sigmoid activation function, and the formula is as follows:

[0214] ;

[0215] in, Let i be the attention weight for the i-th channel. Let N be the j-th feature value of the i-th channel, and N be the number of features. Use the Sigmoid activation function;

[0216] Dynamic weighted fusion: Combining attention weights, a dynamic weighted fusion algorithm is used to fuse all feature vectors. The fusion formula is as follows:

[0217] ;

[0218] Where F is the fused feature vector. For satellite remote sensing image feature vectors, For UAV image feature vectors, To fix the feature vector of the camera image, This is a vector for voiceprint feature recognition and behavior judgment. , , , For dynamic weights, + + + = 1;

[0219] Adaptive weight adjustment: manually adjusting weights , , , ;

[0220] Feature fusion optimization: The fused feature vectors are smoothed by convolution using a 3×3 convolution kernel to eliminate feature noise generated during the fusion process.

[0221] Verification and Optimization Layer: The output optimization layer of the model is used to perform data verification, anomaly removal, and accuracy optimization on the fused feature vector. It combines radar data from the aerial biological observation device in the nature reserve to perform multi-dimensional verification on the fused feature vector and correct the extraction bias of various data sources. Anomaly detection algorithms are used to remove abnormal data. Finally, the verified fused feature vector is standardized and the results are output.

[0222] The processing steps of the verification optimization layer include:

[0223] Multi-dimensional data verification: Combining radar data from the aerial biological observation device in the nature reserve, the flight parameters, population size, and behavioral status in the fused feature vector are verified; the deviation between the fused result and the radar data is calculated. If the deviation exceeds 5%, the weights of the corresponding feature vectors are adjusted and the data is re-fused until the deviation is ≤5%; at the same time, the voiceprint recognition results and the species recognition results from the fixed camera images are cross-validated. If the recognition results are inconsistent, the result with the higher confidence level is used to correct the fused result.

[0224] Outlier removal: The Isolation Forest algorithm is used to detect anomalies in the fused feature vectors, identify and remove outlier data; the anomaly detection threshold is set to m, and when the anomaly confidence is ≥ m, it is determined to be outlier data and removed.

[0225] Accuracy Optimization: The accuracy of the fused feature vector after verification is optimized using a gradient descent algorithm to minimize the error between the fused result and the actual observed data. The fusion parameters are iteratively optimized to improve the fusion accuracy. The objective function is:

[0226] ;

[0227] in, The error loss is M, and the amount of fused data is M. For the k-th fusion feature value, This is the k-th true observation;

[0228] Standardized output: The optimized fusion feature vector is converted into a standardized multi-dimensional fusion dataset, which is organized according to the nature reserve zones. It includes the species of flying organisms, protection level, population size, flight trajectory, behavioral status, fixed-point activity patterns, and habitat ecological environment parameters.

[0229] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications 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.

Claims

1. A flight biometric identification method based on multi-source data from nature reserves, characterized in that, Includes the following steps: S1. Multi-source data acquisition and preprocessing to obtain preprocessed multi-source data; Acquire multi-source data, including satellite remote sensing images, UAV images, fixed camera images, and voiceprint features of accompanying flying organisms observed in nature reserves; Preprocessing of multi-source data yields preprocessed multi-source data, which includes preprocessed satellite remote sensing images, preprocessed UAV images, preprocessed fixed camera images, and preprocessed voiceprint features. S2. Process the preprocessed multi-source data and generate multi-source data feature vectors; Preprocessed satellite remote sensing images, preprocessed UAV images, preprocessed fixed camera images, and preprocessed voiceprint features are processed separately to extract key information related to flight biological observation from various types of data, forming feature vectors for satellite remote sensing images, UAV images, fixed camera images, and voiceprint feature recognition and behavior judgment. S3. Construct a flight biometric target model; An initial flight biometric model is constructed, a training dataset is built, and the training dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The initial flight biometric model is trained on the training dataset until the loss function converges, and the target flight biometric model is obtained. S4. Flight biometrics based on flight biometric target model; The multi-source data feature vectors are input into the flight biometric target model for automatic identification, and the flight biometric results are output.

2. The flight biometric identification method based on multi-source data from nature reserves according to claim 1, characterized in that, S1 includes the following steps: S11. Acquisition and preprocessing of satellite remote sensing images: Acquisition of satellite remote sensing imagery: Multispectral and high-resolution panchromatic images of the entire nature reserve and its surrounding areas are acquired through a satellite remote sensing platform. The image resolution is no less than 10 meters, and the corresponding metadata is acquired simultaneously. Preprocessing of satellite remote sensing images includes: Based on the image quality evaluation indicators, valid images are selected, and invalid images with excessive cloud cover, excessive shadow coverage, or severe noise are removed. The selected valid images are subjected to radiometric correction, geometric correction, and atmospheric correction in sequence. The radiative transfer model is used to eliminate the influence of sensor error, solar altitude angle, and topographic relief, and the image gray values ​​are converted into the true surface reflectance to ensure the radiometric consistency of images in different time periods and regions. Based on the high-precision geographic coordinates of the nature reserve, a polynomial fitting algorithm is used for geometric correction, and finally the preprocessed satellite remote sensing images are obtained. S12. Acquisition and preprocessing of UAV imagery: Acquisition of drone imagery: The drone imagery of the biological activity area was acquired by the low-noise drone observation unit, including visible light images with a resolution of no less than 4K and thermal images with a resolution of no less than 1280×1024. Preprocessing of UAV imagery: Invalid images caused by airflow turbulence, shooting angle deviation, and blurred motion of flying creatures are removed, and valid images that are clear, without obvious distortion, and can clearly identify individual flying creatures or groups are retained. Distortion correction, shading, and geometric registration are performed on the effective images; the camera intrinsic parameter correction algorithm is used to eliminate lens distortion, and the motion blur recovery algorithm is used to correct the image blur caused by airflow turbulence and restore the details of individual flying organisms; the UAV images are geometrically registered with the geographic coordinates of satellite remote sensing images as a reference to ensure accurate spatial alignment, and finally the pre-processed UAV images are obtained. S13. Acquisition and preprocessing of images from a fixed camera: Acquiring images from a fixed camera: A fixed camera is used to photograph flying creatures and generate images of the fixed camera, while simultaneously recording the geographic coordinates, shooting time, light intensity, temperature and humidity of the images; Preprocessing of images from a fixed camera: Remove invalid images caused by poor lighting, blurry lenses, or obstructed shooting angles, and retain valid images that are clear, without obvious distortion, and can identify individual flying creatures, populations, or behavioral details. The effective images undergo lighting correction, noise reduction, distortion correction, and geometric registration. An adaptive lighting equalization algorithm is used to correct brightness differences in images under different time periods and lighting conditions. A Gaussian filtering algorithm is used to remove image noise and preserve details of individual flying organisms. A camera intrinsic parameter correction algorithm is used to eliminate lens distortion and restore the true shape of the image. Using the geographic coordinates of satellite remote sensing images as a reference, geometric registration is performed on fixed camera images to accurately align the spatial positions of satellite remote sensing images and UAV images, ultimately obtaining preprocessed fixed camera images. S14. Acquisition and preprocessing of voiceprint features: Acquisition of voiceprint features: Acquire voiceprint features using a voiceprint machine; Preprocessing of voiceprint features: Remove invalid voiceprint data containing strong environmental noise and retain valid voiceprint data with a signal-to-noise ratio ≥20dB; The effective voiceprint data is denoised, normalized, and truncated. A wavelet threshold denoising algorithm is used to remove environmental noise and retain the core features of the voiceprint of the flying organism. The voiceprint data is normalized to adjust the amplitude to the same range, eliminating the influence of the acquisition equipment and acquisition distance, and finally obtaining the preprocessed voiceprint features.

3. The flight biometric identification method based on multi-source data from nature reserves according to claim 1, characterized in that, S2 includes the following steps: S21. Processing of pre-processed satellite remote sensing images: S211, Denoising and Enhancement; A hierarchical denoising algorithm is adopted to denoise the images of nature reserves by region and type. An improved median filtering algorithm is used to remove atmospheric noise and sensor noise, while retaining vegetation and water body edge information to avoid image blurring. Histogram equalization combined with band fusion enhancement algorithm is used to enhance the contrast between vegetation, water body and background, highlighting the vegetation growth and water body boundary information of the habitat of flying organisms. S212, Image Fusion; The SIFT feature matching algorithm is used to accurately register the S211-processed multispectral image with the high-resolution panchromatic image, so that the spatial positions are completely aligned; then, the wavelet transform fusion algorithm is used to deeply fuse the spatial detail information of the high-resolution panchromatic image with the spectral information of the multispectral image. S213, Information Extraction; Key information is extracted from the data processed by S212, and satellite remote sensing image feature vectors are output. S22. Processing of pre-processed UAV images: S221, Denoising and Enhancement; To address the interference characteristics of UAV imagery, an adaptive filtering algorithm is used for noise reduction, and a sharpening enhancement algorithm is employed to improve the clarity of flight organism morphology, feather features, and population aggregation status. Temperature threshold segmentation technology is used to process thermal imaging images. S222, Information Extraction; Using the YOLO8 target detection algorithm, we accurately extract individual flying organisms and population aggregation areas from UAV visible light and thermal images; we statistically analyze the population size and individual size parameters of flying organisms, and combine them with UAV flight parameters to calculate the flight altitude and speed of flying organisms to assist in the analysis of flight trajectories; for migratory bird populations, we extract population aggregation density and distribution range, mark flight activity hotspots, and output UAV image feature vectors. S23. Preprocessing of fixed camera images: S231, Denoising and Enhancement; To address the interference characteristics of fixed camera images, an adaptive Gaussian filtering algorithm is used for noise reduction, preserving individual details and behavioral characteristics of flying organisms. Sharpening and contrast adjustment algorithms are employed to enhance the clarity of the morphological and behavioral details of flying organisms and correct brightness deviations in nighttime infrared images. S232, Information Extraction; An improved target detection and behavior recognition algorithm is used to extract information on individual and population flying organisms from visible light and infrared images from fixed cameras, and to count the activity frequency and dwell time of flying organisms in specific areas. The behavior status of flying organisms is accurately identified, and the activity patterns of flying organisms are analyzed by combining the shooting time series. Detailed features of individual flying organisms are extracted, and combined with activity trajectories and behavioral details, a fixed-point observation dataset is formed, and fixed-camera image feature vectors are output. S24. Preprocessing of voiceprint features: S241. Voiceprint feature extraction; By using Mel-frequency cepstral coefficients (MFCC) combined with short-time Fourier transform, the temporal and frequency domain features of the flight biological voiceprint are extracted from the effective segments of the preprocessed voiceprint features in S14, forming a standardized feature vector for recognition and behavior judgment. S242, Voiceprint recognition; A database of voiceprint features of flying organisms in nature reserves was constructed, including voiceprint features, species information, and protection levels of common flying organisms. The Support Vector Machine (SVM) algorithm was used to match the extracted standardized recognition and behavior judgment feature vectors with the features in the database to identify flying organism species with an accuracy of no less than 90%. Voiceprints that did not match were marked as unknown species and their feature parameters were recorded. S243, Behavioral Judgment; Based on changes in voiceprint features, the behavioral state of flying creatures can be determined, such as breeding calls, warning calls, and group communication, providing support for the analysis of flying creature behavior. At the same time, it is linked and verified with the behavioral features extracted from fixed camera images, and outputs voiceprint feature recognition and behavior judgment vectors.

4. The flight biometric identification method based on multi-source data from nature reserves according to claim 3, characterized in that, The flight biometric model includes: Data Input Layer: The foundational layer of the model, used to receive feature vectors from satellite remote sensing images, UAV images, fixed camera images, voiceprint recognition and behavior judgment vectors. It also imports geographic coordinates, timestamps, and environmental parameter information corresponding to various types of data, and performs standardization and normalization processing on the input data to ensure that the feature vectors have uniform dimensions and consistent data format. Spatial alignment layer: The core support layer of the model; Based on the WGS84 geographic coordinates of satellite remote sensing imagery, combined with geographic coordinate information from UAV imagery, fixed camera images, and voiceprint data, a coordinate mapping algorithm is used for spatial alignment. Spatial deviation is eliminated through iterative optimization, so that the alignment error is ≤0.5 meters. Feature Fusion Layer: The core fusion layer of the model, it adopts an improved weighted fusion algorithm combined with an attention mechanism to achieve deep fusion of four feature vectors. Through the attention mechanism, it automatically identifies the importance of various features, dynamically allocates fusion weights, and performs feature splicing and dimensional fusion on various feature vectors during the fusion process to construct multi-dimensional observation feature vectors, thereby achieving complementary fusion of macroscopic information, microscopic information, fixed-point information, and voiceprint information. Verification and Optimization Layer: The output optimization layer of the model is used to perform data verification, anomaly removal, and accuracy optimization on the fused feature vector. It combines radar data from the aerial biological observation device in the nature reserve to perform multi-dimensional verification on the fused feature vector and correct the extraction bias of various data sources. Anomaly detection algorithms are used to remove abnormal data. Finally, the verified fused feature vector is standardized and the results are output.

5. The flight biometric identification method based on multi-source data from nature reserves according to claim 4, characterized in that, The data input layer processing steps include: Unified feature dimensions: Satellite remote sensing image feature vectors, UAV image feature vectors, fixed camera image feature vectors, voiceprint feature recognition and behavior judgment vectors are uniformly mapped to a 64-dimensional feature space through a feature mapping algorithm; Data normalization: The Min-Max normalization algorithm is used to uniformly adjust the numerical range of various feature vectors to the [0,1] interval, eliminating the influence of different data source acquisition devices and acquisition conditions; The normalization formula is: ； Where X is the original feature value, This is the minimum value of the feature. The maximum value of this feature. These are the normalized eigenvalues; Data cleaning: Perform preliminary outlier detection on the input data and remove missing values ​​and extreme outliers.

6. The flight biometric identification method based on multi-source data from nature reserves according to claim 4, characterized in that, The processing steps for the spatial alignment layer include: Unified coordinate reference: The geographic coordinates of UAV imagery, fixed camera images, and voiceprint data are uniformly converted into the WGS84 geographic coordinate system used by satellite remote sensing imagery. The coordinate system conversion is completed by using a coordinate transformation algorithm, and the conversion error is ≤0.1 meters. Feature point matching: Extract geographic feature points from satellite remote sensing images, UAV images, and fixed camera images, and use the SIFT feature matching algorithm to accurately match feature points from satellite remote sensing images, UAV images, and fixed camera images, and to make satellite remote sensing images, UAV images, and fixed camera images spatially correspond to satellite remote sensing images. Spatial deviation correction: Combine the geographic coordinates of the voiceprint data to calculate the spatial deviation between each data source and the satellite remote sensing image reference. The iterative least squares method is used to correct the deviation, with the number of iterations set to 10-15 times, until the spatial alignment error is ≤0.5 meters. Time synchronization: Combine the timestamp of each data point to synchronize the time of the spatially aligned multi-source data, so that the multi-source data in the same observation period correspond and match.

7. The flight biometric identification method based on multi-source data from nature reserves according to claim 4, characterized in that, The processing steps of the feature fusion layer include: Attention mechanism design: Channel weights are assigned to the standardized feature vectors to automatically identify the core channels relevant to flight biological observation in each feature. Attention weights are assigned to the core channels and the ineffective channels respectively. The attention weights are calculated using the Sigmoid activation function, and the formula is as follows: ； in, Let i be the attention weight for the i-th channel. Let N be the j-th feature value of the i-th channel, and N be the number of features. Use the Sigmoid activation function; Dynamic weighted fusion: Combining attention weights, a dynamic weighted fusion algorithm is used to fuse all feature vectors. The fusion formula is as follows: ； Where F is the fused feature vector. For satellite remote sensing image feature vectors, For UAV image feature vectors, To fix the feature vector of the camera image, This is a vector for voiceprint feature recognition and behavior judgment. , , , For dynamic weights, + + + = 1; Adaptive weight adjustment: manually adjusting weights , , , ; Feature fusion optimization: The fused feature vectors are smoothed by convolution using a 3×3 convolution kernel to eliminate feature noise generated during the fusion process.

8. The flight biometric identification method based on multi-source data from nature reserves according to claim 4, characterized in that, The processing steps of the verification optimization layer include: Multi-dimensional data verification: Combining radar data from the aerial biological observation device in the nature reserve, the flight parameters, population size, and behavioral status in the fused feature vector are verified; the deviation between the fused result and the radar data is calculated. If the deviation exceeds 5%, the weights of the corresponding feature vectors are adjusted and the data is re-fused until the deviation is ≤5%; at the same time, the voiceprint recognition results and the species recognition results from the fixed camera images are cross-validated. If the recognition results are inconsistent, the result with the higher confidence level is used to correct the fused result. Outlier removal: The Isolation Forest algorithm is used to detect anomalies in the fused feature vectors, identify and remove outlier data; the anomaly detection threshold is set to m, and when the anomaly confidence is ≥ m, it is determined to be outlier data and removed. Accuracy Optimization: The accuracy of the fused feature vector after verification is optimized using a gradient descent algorithm to minimize the error between the fused result and the actual observed data. The fusion parameters are iteratively optimized to improve the fusion accuracy. The objective function is: ； in, The error loss is M, and the amount of fused data is M. For the k-th fusion feature value, This is the kth true observation; Standardized output: The optimized fusion feature vector is converted into a standardized multi-dimensional fusion dataset, which is organized according to the nature reserve zones. It includes the species of flying organisms, protection level, population size, flight trajectory, behavioral status, fixed-point activity patterns, and habitat ecological environment parameters.

9. The flight biometric identification method based on multi-source data from nature reserves according to claim 1, characterized in that, S4 includes the following steps: S41. Input the feature vectors of multi-source data into the flight biometric target model for automatic identification, and output the flight biometric results; S42. Analyze the results of flight biometric identification; S421. Population dynamics analysis; By combining time-series fusion data, we can analyze changes in the population size, migration routes, and aggregation areas of flying organisms, monitor the population migration dynamics during the migratory season of migratory birds, and monitor the population size changes of rare species during the breeding season; by combining fixed-point observation data from fixed cameras, we can analyze the activity patterns of flying organisms. S422, Habitat quality assessment; By combining habitat ecological environment parameters extracted from satellite remote sensing images, habitat details extracted from UAV images and fixed camera images, the quality level of the habitat of flying organisms is assessed, and the impact of habitat change on the population and behavior of flying organisms is analyzed. S423, Collaborative and Interactive Applications; By linking the fusion results with data from the aerial organism observation device in the nature reserve, integrated monitoring of "air-space-ground-sound" can be achieved. The terminal device can be used to view multi-dimensional information of aerial organisms in real time and remotely adjust the observation parameters. S424. Data storage and traceability; The fusion results, each feature vector, and the original data are stored synchronously, supporting data export, backup, and traceability, with a focus on preserving relevant data on rare and protected flying organisms.

Images