A bird chirping sound recognition method based on a combination of voiceprints and spatial distribution

Abstract

The application discloses a bird chirp sound recognition method based on a combination of voiceprints and spatial distribution, and belongs to the technical field of intelligent sound classification and recognition. In view of the problems of ignoring geographical distribution prior knowledge and sample imbalance in the prior art, the application firstly constructs a voiceprint recognition model: a training data set is constructed by audio preprocessing, logarithmic mel spectrum and dynamic difference feature extraction, a model is trained based on DenseNet-121 by adopting a two-stage training strategy, and recognition confidence of each species is obtained; meanwhile, a spatial distribution model is constructed: based on public observation data, an average observer ability index is used to correct an original encounter rate, and spatial distribution probability of the species in a specific city is obtained; finally, a Sigmoid function is used to perform nonlinear fusion on the two, a joint recognition probability is calculated, and a classification result is output. The application introduces ecological spatial constraints into the recognition decision, effectively reduces false positive misjudgment, improves rare species monitoring capability, and makes the recognition result have ecological interpretability.

Description

Technical Field

[0001] This invention relates to the field of intelligent sound classification technology, and in particular to a method for bird call recognition based on a combination of voiceprint and spatial distribution. Background Technology

[0002] Birds serve as key indicator organisms in ecosystems, and their population structure, distribution patterns, and dynamic changes effectively reflect the ecological status and environmental health of their habitats. With rapid urbanization, bird species diversity is declining, potentially leading to impaired urban ecosystem functions, such as reduced seed dispersal capacity and decreased pest control efficiency. Therefore, spatial distribution analysis and dynamic monitoring of birds not only help identify core ecological source areas but also analyze the impact of environmental factors on species, providing data support for scientific conservation and ecological management.

[0003] In terms of monitoring technology, sound-based bird identification technology has unique advantages over traditional image-based methods. Its non-invasiveness, ability to operate continuously around the clock, wide coverage, and lower labor costs have led to its increasing importance in ecological surveys and monitoring. Bird calls exhibit significant species specificity and complex frequency structures, providing a reliable biological basis for the automated identification of acoustic features.

[0004] In recent years, with the advancement of deep learning and pattern recognition technologies, researchers have proposed various automated bird call recognition models, which can achieve high recognition accuracy in laboratory environments. However, in practical applications, these methods still face the following challenges: (1) Insufficient spatial information integration. Most models rely solely on sound features for training and classification, without fully considering the geographical distribution patterns of birds. The habitats of different species vary significantly, and ignoring spatial priors may lead to prediction errors in abnormal distribution areas, resulting in misjudgments that do not conform to the actual distribution; (2) Data imbalance. In bird call data, there are abundant samples of common species, while samples of rare or elusive species are extremely limited. This class imbalance affects the accuracy of the model when dealing with closely related species with similar characteristics but significantly different ecological habits, and also reduces the monitoring capabilities for rare and endangered species.

[0005] Therefore, there is an urgent need in this field for a technical solution that can integrate voiceprint features with spatial distribution information, overcome the problem of data imbalance, and improve the accuracy of bird call recognition in complex outdoor environments.

[0006] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a method for bird call recognition based on a combination of voiceprint and spatial distribution.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for bird call recognition based on a combination of voiceprint and spatial distribution includes: The voiceprint features of bird calls are obtained, and these features are input into a pre-trained voiceprint recognition model to obtain the recognition confidence score for each species. p i ; Based on the geographical location information of the collected bird calls, a pre-built spatial distribution model is input to obtain the spatial distribution probability of each species in the corresponding urban area. P s ; The recognition confidence level is determined using the Sigmoid function. P i With the spatial distribution probability P s Perform nonlinear fusion to calculate the joint recognition probability. P f The joint identification probability P f The calculation formula is:

[0009] in, k The steepness parameter is set to 10. a The value is 1; Based on the joint identification probability P f Output the bird call recognition results.

[0010] Optionally, the training of the voiceprint recognition model includes the following steps: S1. Audio signal preprocessing: High-frequency enhancement, silent segment filtering and segmentation of the audio are performed. Abnormal samples are removed by clustering method and manually reviewed to form a standardized dataset. S2. Birdsong feature extraction: Extract the log-Mel spectrum and its first-order and second-order dynamic difference features, and fuse different features to form a composite vector; S3. Voiceprint Recognition Model Training: Based on the DenseNet-121 network framework, a two-stage training strategy is adopted. The first stage learns the general voiceprint features of all birds. The second stage combines logarithmic weighting and Focal Loss to enhance the recognition ability of rare species.

[0011] Optionally, the construction of the spatial distribution model includes the following steps: S1. Birdwatching data compilation: Based on public birdwatching data, using municipal administrative districts as the basic statistical unit, the number of reports, reporting time, number of reporters, and number of species are counted. S2. Calculation of the average observer capability index: For each observer in each city, they are arranged in chronological order of observation time. k The total number of species observed in this survey is recorded as follows: N k Using the logarithmic function form of the species accumulation curve Perform fitting, fitting slope a The observer capability index is calculated by taking the arithmetic mean of all observer capability indices in the city, resulting in the city's average observer capability index. OCI i ; S3. Calculation of the initial encounter rate: Based on the formula... Calculate birds b exist d The probability of meeting in the city, among which R b,d For birds b exist d The number of times it appeared in the city report for d The total number of reports from the city; S4. Bird distribution probability correction: using a log-normalized correction formula.

[0012] Calculate the adjusted probability of occurrence as the spatial distribution probability. P s .

[0013] Optionally, in the step of calculating the average observer capability index, when fitting the logarithmic function, data from the early stage of the species accumulation curve rising to a stable stage are used for fitting, and samples that tend to saturate in the later stage are removed.

[0014] Optionally, the bird distribution probability correction step further includes boundary condition processing: when When =1, no correction is made; when When the value is greater than 1, the correction function is a convex function, and for larger values... The result has a relatively small impact, on smaller... Compress; when When the value is less than 1, the correction function is a concave function, and the low probability value is appropriately increased.

[0015] Optionally, a real-time data preprocessing step is included before acquiring the voiceprint features of bird calls: A noise template library covering four types of noise—wind noise, vehicle noise, construction noise, and fireworks noise—was constructed. 100 samples were collected for each type of noise, and 13-dimensional MFCC features were extracted as standard templates. The real-time acquired audio stream is processed in frames, and the DTW distance is calculated frame by frame with the noise templates in the template library. The DTW distance threshold T=55 is set. If the minimum distance is lower than the threshold, it is judged as a noise segment and is removed; otherwise, it is retained as audio to be identified. A multi-threaded parallel processing mechanism is adopted to divide the audio stream into blocks for parallel matching, and a moving average strategy is used for boundary cases, combined with the judgment results of the preceding and following frames for comprehensive decision-making.

[0016] Optionally, the real-time data preprocessing step further includes a dynamic template library update mechanism: If an audio segment is repeatedly identified by a voiceprint recognition model as belonging to a species outside the region with a high confidence level, but is confirmed by manual verification as noise interference, then the feature sequence of that audio segment will be included in the template library as a new noise type. The matching hit rate of each template is calculated quarterly. Templates with a hit rate below the threshold are subject to quality review, and noise samples under different conditions are collected to enrich the original templates. Noise types that occur very infrequently are archived into the low-frequency noise library and are no longer included in real-time matching.

[0017] Optionally, the steepness parameter k The value is determined based on validation set testing: comparison k Take the model performance at times 6 to 12, when k When the value exceeds 10, the model's discriminative sensitivity becomes too strong, causing the same audio sample to output confidence scores close to 1 for multiple species. Therefore, [the value is] reduced. k The parameter is 10; a The value is set to 1 to ensure that the voiceprint recognition probability and the spatial distribution probability have equal contribution rates in the Sigmoid function.

[0018] Optionally, in the birdwatching data processing step, the fields collected include report number, observer, observation date, observation location, bird name, and number of observations; and observation data with incorrect geographic coordinates are removed.

[0019] Optionally, the joint identification probability P f The value range is [0,1]. When the voiceprint recognition result is consistent with the spatial distribution probability, the joint probability is significantly increased; when there is a contradiction between the two, the joint probability is significantly reduced.

[0020] Compared with the prior art, the present invention has the following beneficial effects: First, this invention introduces spatial distribution information as an ecological constraint. Traditional voiceprint recognition models classify based solely on sound features, easily misidentifying species with similar calls but non-overlapping geographical distributions. This invention constructs a spatial distribution probability model, incorporating prior knowledge of geographical distribution into the recognition decision. It utilizes a sigmoid function for joint probability fusion. When the species identification result exceeds its known distribution range, the overall confidence level is significantly reduced, effectively decreasing false positives and improving species identification accuracy in complex wild environments.

[0021] Second, this invention can improve the problem of imbalanced data and the presence of rare species. While existing methods improve voiceprint recognition performance by modifying network structures, they still fail to address the imbalanced nature of field data, where common species have a large number of samples and rare species have a small number. This invention, through two-stage training combined with spatial distribution probability, ensures that even if rare species have low initial confidence in the voiceprint model, their spatial distribution confidence will be positively weighted if they appear within their historical distribution area. This prevents rare species from being masked by common species and improves the monitoring capabilities for endangered and elusive species.

[0022] Third, the identification results are more ecologically interpretable. The spatial distribution model of this invention is built upon large-scale birdwatching data and corrects for observer differences using an average observer ability index, making the distribution probability calculation results more scientific and comparable. This design ensures that the identification results not only output species names but also implicitly convey the ecological rationale for the species' presence in specific areas, providing more reliable basic data for ecological research, habitat assessment, and conservation planning. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a model configuration diagram provided for an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The purpose of this invention is to provide a method that can combine a voiceprint recognition model and a spatial distribution model, and output ecologically interpretable bird call recognition results through nonlinear fusion.

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Example 1: This embodiment provides a bird call recognition method based on a combination of voiceprint and spatial distribution. It incorporates a two-stage training method to overcome the problems of insufficient integration of geographical distribution information and imbalanced data during training in existing technologies, thereby improving the scientific accuracy of the recognition results. Specifically, it includes a voiceprint recognition model, a spatial distribution model, and a joint recognition model. A diagram of the model structure is shown below. Figure 1 .

[0029] 1. Voiceprint recognition model: Training a voiceprint recognition model includes the following steps: S1. Audio signal preprocessing. High-frequency enhancement, silent segment filtering, and segmentation are performed on the audio. Outlier samples are removed using clustering methods, and the data is manually reviewed to create a standardized dataset.

[0030] S2. Birdsong Feature Extraction. Log-Mel spectrum and its first and second-order dynamic difference features are extracted. Different features are fused to form a composite vector, preserving time-frequency structure information, providing input for deep learning model training.

[0031] S3. Voiceprint Recognition Model Training. Based on the DenseNet-121 network framework, a two-stage training strategy is adopted: the first stage learns the universal voiceprint features of all birds; the second stage combines logarithmic weighting and Focal Loss to enhance the model's ability to recognize rare species.

[0032] 2. Spatial distribution model: Traditional voiceprint recognition relies solely on sound features, neglecting the geographical heterogeneity and niche constraints of bird distribution, leading to some recognition results that violate biogeographical principles. This invention addresses this limitation by innovatively constructing a spatial distribution model based on publicly available scientific data. This model is not merely a simple correction to voiceprint recognition results; it incorporates factors such as regional observation intensity and observer capability into the probability distribution framework, giving the final recognition results stronger ecological interpretability and realistic plausibility. Compared to the Maxent habitat model, it requires fewer parameters and is easier to construct. The spatial distribution model includes the following steps: S1. Birdwatching data processing: S101. Data Collection: Obtain public birdwatching data. The collected fields include report number, observer, observation date, observation location, bird name, number of observations, etc.

[0033] S102. Data cleaning: Remove observation data with incorrect geographic coordinates and retain valid records.

[0034] S103. Indicator Statistics: Using municipal-level administrative regions as the basic statistical unit, the following indicators are statistically analyzed: Number of reports: The total number of all observation reports in the city; Reporting time: The time when the city uploaded the report; Number of reporters: The number of personnel involved in the observation; Number of species: The number of species observed by each observer in each instance.

[0035] S2. Calculation of the average observer's ability index: While public science data boasts a large sample size, it suffers from significant observer experience bias and uneven regional monitoring intensity. Directly using raw report frequencies overestimates the contributions of popular birdwatching spots or experienced birdwatchers, while underestimating the actual distribution of remote areas or common species. This invention introduces a species accumulation curve and uses an observer capability index as an intermediate correction factor. This index quantifies the species discovery efficiency per unit of observation capability, effectively reducing data discrepancies caused by differences in observer skill levels. It allows subsequent distribution probability calculations to reflect the ecological attributes of the species themselves, representing a crucial step in the deep cleaning and ecological value extraction of public science data.

[0036] S201, Individual Observer Ability Fitting: For a certain city, the... The observers are arranged in chronological order of their observations. The total number of species observed in this survey is recorded as follows: Logarithmic function fitting is used: The formula is as follows:

[0037] in, To fit the slope (the larger the slope, the stronger the ability); The intercept; The number of observations.

[0038] S202. Fitting conditions: Data from the early stage of the curve rising to a stable stage are used for fitting, and samples that tend to saturate in the later stage are removed to truly reflect the observer's discovery efficiency.

[0039] S203. Calculation of the average observer ability index: The fitting slope of all observers in the city... The arithmetic mean is taken to obtain the city's average observer capability index:

[0040] in, for i City average observation capability index n for i Number of people observed in the city.

[0041] S3. Calculation of the original encounter rate: The initial encounter rate is used as the target for subsequent adjustments, as shown in the following formula:

[0042] in, For birds b exist d The probability of meeting in the city, For birds b exist d The number of times it appeared in the city report for d The total number of reports from the city.

[0043] S4. Bird distribution probability correction: To avoid overestimating or underestimating the distribution probability due to differences in observer ability and uneven observation intensity in different regions when directly using the original encounter rate, this invention introduces a log-normalization correction method based on the average observer ability index: S401 Correction Principle: In areas with strong observation capabilities (higher values), the original encounter rate better reflects the true distribution, and the correction magnitude is smaller; in areas with weak observation capabilities (lower values), the original data may contain randomness and needs to be appropriately compressed to avoid overestimating the species distribution probability due to a single accidental record. This processing method makes the final distribution probability comparable between areas with different observation intensities, which is more in line with the continuity and gradient characteristics of species distribution.

[0044] S402, Logarithmic Normalization Correction Formula:

[0045] in, For birds b exist d The adjusted probability of the city's occurrence. for i The city's average observer capability index, This represents the original encounter rate.

[0046] S403, Boundary Condition Handling: when When =1, that is, when the observation capability is at the average level, no correction is made; when When >1, the correction function is a convex function. Larger results have a smaller impact. Smaller results are compressed. when When the value is less than 1, the correction function is a concave function, which appropriately increases the low probability value to avoid species being missed due to insufficient observation capabilities.

[0047] 3. Joint Recognition Model The joint identification model described in this invention includes two stages: real-time data preprocessing and joint probability calculation. The specific steps are as follows: S1. Real-time data preprocessing.

[0048] S101. Noise Template Library Construction: For common environmental noise types found at monitoring stations, clean noise samples are pre-collected to construct a noise template library. The template library covers the following four noise categories: Wind sound: The continuous sound of airflow at different wind speeds; Vehicle noise: traffic noise such as vehicles moving and honking; Construction noise: Noise from piling, cutting, and machinery operation; Fireworks sound: the crackling sound and the trailing sound when fireworks are set off.

[0049] One hundred samples were collected for each type of noise, covering different intensities, durations, and variation patterns. After preprocessing, 13-dimensional MFCC features were extracted from all samples as standard templates.

[0050] S102, Template Matching and Noise Removal: The real-time acquired audio stream is processed frame by frame, and the DTW distance is calculated frame by frame against various noise templates in the template library. To reduce computational load, a multi-threaded parallel processing mechanism is adopted to match the audio stream in blocks in parallel. A DTW distance threshold T=55 is set. If the minimum distance is lower than the threshold, it is judged as a noise segment and removed; otherwise, it is retained as audio to be recognized and input into the subsequent speaker recognition model. For boundary cases (DTW distance near the threshold), a moving average strategy is adopted, combining the judgment results of the previous and next frames to make a comprehensive decision and reduce misjudgments.

[0051] S103. Dynamic Update of Template Library: To adapt to changes in the monitoring environment, a dynamic update mechanism for the template library is designed. If an audio segment is repeatedly identified by the voiceprint recognition model as belonging to a species outside the designated area with high confidence, but is manually verified as noise interference, then the feature sequence of that audio segment is included in the template library as a new noise type. The matching hit rate of each template is calculated quarterly. Templates with a hit rate below a threshold undergo quality review, and noise samples under different conditions are collected to enrich the original templates. Noise types with extremely low frequency of occurrence are archived in the low-frequency noise library and no longer participate in real-time matching.

[0052] S2, Joint probability calculation.

[0053] S201. Input Acquisition: Input the voiceprint feature vector extracted in S1 into the voiceprint recognition model to obtain the recognition confidence score for each species. Simultaneously, based on the geographical location information of the audio collection points, the spatial distribution model is input to obtain the spatial distribution probability of the species in the corresponding city. .

[0054] S202. Joint Probability Fusion: This invention employs the Sigmoid function to nonlinearly fuse voiceprint confidence and spatial distribution probability to construct a joint recognition probability. Unlike traditional linear weighting, this fusion mechanism introduces ecological constraints: when the voiceprint recognition result is consistent with the spatial distribution probability, the joint probability is significantly increased; when the two contradict each other (such as identifying a species not distributed in the area), the joint probability will be significantly reduced, thereby effectively eliminating false identifications due to geographical anomalies. The fusion formula is as follows:

[0055] in, The joint recognition probability ranges from [0,1]. The probability range for the voiceprint recognition model is [0,1]. The spatial distribution probability model takes values ​​in the range [0,1]. k The steepness parameter is set to 10. The value is 1.

[0056] S2-3. Basis for parameter determination: To ensure that the probability of voiceprint recognition and the probability of spatial distribution have equal contribution rates, we set parameters... a It equals 1. k The value is determined based on validation set testing, and compared with... k The model performance was measured when the kurtosis parameter was between 6 and 12. k As the confidence level increases, the model can still maintain a high recognition accuracy in the high confidence interval; however, when... k When the value exceeds 10, the model's discrimination sensitivity becomes too strong, easily resulting in the same audio sample outputting confidence scores close to 1 for multiple species. This leads to low discriminative power and affects the reliability of the recognition results. Therefore, a lower confidence score is adopted. k It is 10.

[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0058] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for bird call recognition based on a combination of voiceprint and spatial distribution, characterized in that, include: The voiceprint features of bird calls are obtained, and these features are input into a pre-trained voiceprint recognition model to obtain the recognition confidence score for each species. p i ; Based on the geographical location information of the collected bird calls, a pre-built spatial distribution model is input to obtain the spatial distribution probability of each species in the corresponding urban area. P s ; The recognition confidence level is determined using the Sigmoid function. P i With the spatial distribution probability P s Perform nonlinear fusion to calculate the joint recognition probability. P f The joint identification probability P f The calculation formula is: in, k The steepness parameter is set to 10. a The value is 1; Based on the joint identification probability P f Output the bird song classification results.

2. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, The training of the voiceprint recognition model includes the following steps: S1. Audio signal preprocessing: High-frequency enhancement, silent segment filtering and segmentation of the audio are performed. Abnormal samples are removed by clustering method and manually reviewed to form a standardized dataset. S2. Birdsong feature extraction: Extract the log-Mel spectrum and its first-order and second-order dynamic difference features, and fuse different features to form a composite vector; S3. Voiceprint Recognition Model Training: Based on the DenseNet-121 network framework, a two-stage training strategy is adopted. The first stage learns the general voiceprint features of all birds. The second stage combines logarithmic weighting and Focal Loss to enhance the recognition ability of rare species.

3. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, The construction of the spatial distribution model includes the following steps: S1. Birdwatching data compilation: Based on public birdwatching data, using municipal administrative districts as the basic statistical unit, the number of reports, reporting time, number of reporters, and number of species are counted. S2. Calculation of the average observer capability index: For each observer in each city, they are arranged in chronological order of observation time. k The total number of species observed in this survey is recorded as follows: N k Using the logarithmic function form of the species accumulation curve Perform fitting, fitting slope a The observer capability index is calculated by taking the arithmetic mean of all observer capability indices in the city, resulting in the city's average observer capability index. OCI i ; S3. Calculation of the initial encounter rate: Based on the formula... Calculate birds b exist d The probability of meeting in the city, among which R b,d For birds b exist d The number of times it appeared in the city report for d The total number of reports from the city; S4. Bird distribution probability correction: using a log-normalized correction formula. Calculate the adjusted probability of occurrence as the spatial distribution probability. P s .

4. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, In the step of calculating the average observer capability index, when fitting the logarithmic function, data from the early stage of the species accumulation curve rising to a stable stage are used for fitting, and samples that tend to saturate in the later stage are removed.

5. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 3, characterized in that, The bird distribution probability correction step also includes boundary condition processing: when When =1, no correction is made; when When the value is greater than 1, the correction function is a convex function, and for larger values... The result has a relatively small impact, on smaller... Compress; when When the value is less than 1, the correction function is a concave function, and the low probability value is appropriately increased.

6. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, Before acquiring the voiceprint features of bird calls, a real-time data preprocessing step is also included: A noise template library covering four types of noise—wind noise, vehicle noise, construction noise, and fireworks noise—was constructed. 100 samples were collected for each type of noise, and 13-dimensional MFCC features were extracted as standard templates. The real-time acquired audio stream is processed in frames, and the DTW distance is calculated frame by frame with the noise templates in the template library. The DTW distance threshold T=55 is set. If the minimum distance is lower than the threshold, it is judged as a noise segment and is removed; otherwise, it is retained as audio to be identified. A multi-threaded parallel processing mechanism is adopted to divide the audio stream into blocks for parallel matching, and a moving average strategy is used for boundary cases, combined with the judgment results of the preceding and following frames for comprehensive decision-making.

7. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, The real-time data preprocessing step also includes a dynamic template library update mechanism: If an audio segment is repeatedly identified by a voiceprint recognition model as belonging to a species outside the region with a high confidence level, but is confirmed by manual verification as noise interference, then the feature sequence of that audio segment will be included in the template library as a new noise type. The matching hit rate of each template is calculated quarterly. Templates with a hit rate below the threshold are subject to quality review, and noise samples under different conditions are collected to enrich the original templates. Noise types that occur very infrequently are archived into the low-frequency noise library and are no longer included in real-time matching.

8. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, The steepness parameter k The value is determined based on validation set testing: comparison k Take the model performance at times 6 to 12, when k When the value exceeds 10, the model's discriminative sensitivity becomes too strong, causing the same audio sample to output a confidence level close to 1 for multiple species. Therefore, [the value is] reduced. k The parameter is 10; a The value is set to 1 to ensure that the voiceprint recognition probability and the spatial distribution probability have equal contribution rates in the Sigmoid function.

9. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, In the birdwatching data processing step, the fields collected include report number, observer, observation date, observation location, bird name, and number of observations; and observation data with incorrect geographic coordinates are removed.

10. The bird call recognition method based on a combination of voiceprint and spatial distribution according to claim 1, characterized in that, The joint identification probability P f The value range is [0,1]. When the voiceprint recognition result is consistent with the spatial distribution probability, the joint probability is significantly increased; when there is a contradiction between the two, the joint probability is significantly reduced.

Images