Wetland water bird automatic identification and statistics method and system based on unmanned aerial vehicle aerial image

Abstract

The application discloses a wetland water bird automatic identification and statistics method and system based on unmanned aerial vehicle aerial image, and belongs to the software technical field.The application aims to solve the problems that the existing technology has missed detection, false detection and uncertainty in the unmanned aerial vehicle aerial water bird identification result, and that the population size and spatial distribution statistical result is inaccurate;for this purpose, the application provides a method, which comprises: grid division and multi-frame observation effort tensor construction;water bird detection and trajectory aggregation based on multi-frame images; detection probability modeling based on observation effort, scene features and artificial checking; grid real number hierarchical inference based on the detection model; real intensity field inference fused with spatial adjacency and ecological factors; and time sequence state space updating under multi-period monitoring.Thus, the statistical inference of the real number and spatial distribution of water birds is realized, and the method has the advantages of high statistical reliability and accurate uncertainty quantification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, more particularly, the present application relates to a wetland water bird automatic identification and statistics method and system based on unmanned aerial vehicle aerial image. BACKGROUND

[0002] Wetland water birds are important indicator species for evaluating the health status of wetland ecosystems and the level of biological diversity. Many rare and migratory water birds are highly sensitive to water level conditions, vegetation types, and the degree of human disturbance in their habitats. Therefore, long-term water bird population monitoring and spatial distribution surveys are often required in the construction of wetland protection zones and nature reserves at all levels. Traditional monitoring methods rely on fixed point observation, belt transect survey, boat-based patrol, and manual telescope counting. Survey personnel identify water bird species and estimate their numbers by visual observation from the shore or boat. Although this method has high identification accuracy, it is limited in coverage, labor-intensive, time-consuming, and subjective, and it is difficult to repeat frequently, which cannot meet the demand for high spatial and temporal resolution monitoring data in large-scale wetlands and complex water network areas.

[0003] The prior art has the following disadvantages:

[0004] In existing unmanned aerial vehicle water bird automatic identification schemes, most of the work remains at the algorithm performance optimization level of detection. The main evaluation indicators are concentrated on accuracy, recall rate, mAP, etc. in single frame or single image. In engineering practice, the number of detected target boxes is often directly used as the water bird number statistics result, or simply rasterized and summarized in space to draw a species distribution heat map. For the application scenario of wetland water bird monitoring, unmanned aerial vehicle aerial tasks usually produce multiple overlapping images in the same grid cell. Under different flight routes, heights, pitch angles, and lighting conditions, the same location may be repeatedly observed in multiple frames, or completely unidentifiable due to water surface reflection, vegetation obstruction, flight posture, etc. At the same time, birds have obvious movement and aggregation during shooting, which may result in a large number of individuals appearing suddenly in one frame and significantly decreasing in another frame, making the single frame detection result itself have obvious randomness and uncertainty.

[0005] To solve the above problems, the present application provides a solution. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a wetland water bird automatic identification and statistics method and system based on unmanned aerial vehicle aerial image to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0008] The wetland water bird automatic recognition and statistical method and system based on unmanned aerial vehicle aerial image comprises the following steps:

[0009] Step S1: Regular grid division is performed on the aerial vehicle aerial area, the image frame set covered by each grid is counted for each monitoring period t, an observation effort tensor containing the number of frames, height, viewing angle, coverage rate and time span is constructed, and ecological scene features of each grid are extracted;

[0010] Step S2: Water bird target detection and classification are performed on each image, the detection box, species category, confidence and uncertainty feature information of each candidate target are obtained, and the trajectory set is formed by associating multiple frames of detection results in time and space order in each grid to obtain the number of trajectories in the grid;

[0011] Step S3: A detection probability function based on the observation effort tensor and the ecological scene features is defined, the number of trajectories in each grid is modeled as a binomial distribution with the real number and the detection probability function as parameters and is superimposed with a Poisson type false detection noise, and the model parameters are calibrated by using artificial review samples;

[0012] Step S4: A Poisson prior model of the real number of each grid is established, hierarchical Bayesian inference is performed, the posterior distribution of N is obtained by combining the trajectory observation, observation effort and scene features according to the calibrated detection probability model, and the posterior is calculated by an algorithm to estimate the real number and its variance;

[0013] Step S5: A strength field model of water bird distribution is defined, the observation results, spatial proximity smoothing constraint and ecological factors are introduced into the objective function, the objective function is solved by variational inference or iterative optimization to obtain the optimal strength estimation value of each grid.

[0014] In a preferred embodiment, in step S1, the regular grid division comprises dividing the aerial vehicle aerial area into a plurality of grid units according to a preset spatial resolution, the observation effort tensor comprises the number of aerial frames, the average flight height, the camera pitch angle range, the proportion of the effective coverage area and the continuous observation time of each grid unit in the monitoring period, and the ecological scene features of each grid unit are extracted, including vegetation type, water depth, water level, distance from the shoreline, distance from human activity facilities, water surface reflectivity and texture complexity.

[0015] In a preferred embodiment, in step S2, when the water bird target detection and classification are performed on each image, the water bird target detection and classification output the detection box position, species category and confidence score of the candidate target based on the pre-trained or transfer learning convolutional neural network model, and only the candidate targets with a confidence greater than a preset threshold are retained to participate in the subsequent multi-frame trajectory association and trajectory number statistics in the grid unit.

[0016] In a preferred embodiment, in step S2, when performing multi-frame track association, only candidate targets with spatial position overlap exceeding a preset intersection-over-union threshold, motion direction and speed change satisfying continuous motion constraint, and time interval not exceeding a preset inter-frame interval between adjacent aerial frames in the same grid cell are merged into the same waterbird track, and each candidate target is limited to belong to at most one waterbird track.

[0017] In a preferred embodiment, in step S3, the detection probability function is represented by a parameterized function model taking the observation effort tensor and the ecological scene feature as input, the parameterized function model being a logistic regression model or a small neural network model, and a monotonicity constraint is imposed on the model parameters, so that the detection probability monotonically does not decrease with the number of aerial frames and monotonically does not increase when the flight height exceeds a preset upper limit.

[0018] In a preferred embodiment, in step S3, when calibrating the model parameters using artificial review samples, several monitoring periods are selected as artificial review grids from grid cells of different observation effort levels and different ecological scene types, the aerial images in each of the artificial review grids are manually annotated to obtain the true waterbird number, and the true number and the corresponding track number are combined to form a training sample set for maximum likelihood estimation or expectation maximization algorithm to update the model parameters.

[0019] In a preferred embodiment, in step S4, the Poisson prior model takes the intensity parameter of the grid cell as the prior mean, the detection probability function obtained in step S3, the false detection noise model, and the track number jointly constitute an observation model, the posterior distribution and its expectation value are calculated by enumeration calculation in the range of candidate true number values or approximated by expectation maximization algorithm, and the corresponding posterior variance is recorded synchronously when storing the true number estimate.

[0020] In a preferred embodiment, in step S5, the objective function includes an observation constraint term composed of the true number estimate and its variance, a spatial smoothing term that penalizes the intensity difference between adjacent grids according to the grid adjacency relationship, and an ecological factor term that predicts the intensity based on the ecological scene feature and constrains the difference between the predicted intensity and the current intensity, and a gradient descent type iterative algorithm is used to solve the objective function to update the intensity variables of each grid.

[0021] In a preferred embodiment, after step S5, step S6 is further included, in which an intensity state sequence arranged by monitoring period is established for each grid cell, the intensity states of adjacent monitoring periods are associated through a state transition equation, and the intensity state of the previous monitoring period and the change of the ecological factor are used as state prediction input, and the intensity estimate and variance obtained in step S5 are used as observation input, and Kalman filtering or particle filtering is used to complete the intensity state update.

[0022] An automatic identification and statistics system of wetland water birds based on unmanned aerial vehicle aerial image, comprising:

[0023] A grid observation construction module is used for dividing the aerial area into grid units, counting the image coverage frame set to construct an observation effort tensor, and extracting the ecological scene features of each grid;

[0024] A water bird detection and trajectory aggregation module is used for water bird target detection and classification on the aerial image, and associating multiple frame detection results to form a trajectory set and a trajectory number;

[0025] A detection probability modeling and parameter calibration module is used for establishing a water bird detection probability model based on the observation effort tensor and the ecological scene features, and calibrating the model parameters through artificial review samples;

[0026] A real number inference module is used for establishing a prior model of the real number of water birds, combining the detection probability model and the trajectory observation to calculate the posterior estimate value and its variance of the real number by using Bayesian inference;

[0027] A spatial ecological intensity field inference module is used for constructing a spatial intensity field model, optimizing inference by comprehensively considering observation data, spatial adjacency relationship and ecological factors, and obtaining the estimation result of the distribution intensity of water birds.

[0028] The technical effects and advantages of the automatic identification and statistics method and system of wetland water birds based on unmanned aerial vehicle aerial image of the present application are as follows:

[0029] The present application systematically connects the multi-frame observation effort, grid ecological scene features and detection behavior by constructing a hierarchical Bayesian model, realizes the statistical inference of the real number and spatial distribution of water birds, solves the problem that the direct summation of detection frame count in the prior art leads to inaccurate results, and improves the reliability of the statistical results.

[0030] The present application can accurately calculate the number estimation variance of each grid, provide a reliability index for subsequent analysis, realize the explicit quantification of observation uncertainty, and avoid the problem of misjudging the change of observation conditions as the change of population number.

[0031] The present application also integrates spatial adjacency and ecological factors, and updates the time series state space under multi-period monitoring through a state space model, so that reasonable inference results can be obtained in the area with insufficient observation effort, and the trend of population number and spatial distribution can be effectively analyzed. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The present application is a structure schematic diagram of an automatic identification and statistics system of wetland water birds based on unmanned aerial vehicle aerial image;

[0033] Figure 2A wetland water bird automatic identification and statistics method based on unmanned aerial vehicle aerial image is provided. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments.

[0035] The present application provides a wetland water bird automatic identification and statistics method and system based on unmanned aerial vehicle aerial image, aiming at solving the problem that the existing unmanned aerial vehicle aerial water bird identification result has missed detection, false detection and uncertainty, resulting in inaccurate population size and spatial distribution statistical result. The present application systematically links multiple observation efforts, grid ecological scene characteristics and detection behaviors by constructing a hierarchical Bayesian model, realizes statistical inference of the real number and spatial distribution of water birds, and has the advantages of high statistical reliability and accurate uncertainty quantification.

[0036] Embodiment one, the present embodiment provides a wetland water bird automatic identification and statistics method based on unmanned aerial vehicle aerial image, as shown in Figure 2 , comprising the following steps:

[0037] Step S1: This step aims to spatially discretize the monitoring area and quantify the observation effort of the unmanned aerial vehicle on each spatial unit. Specifically, it includes:

[0038] Step S1.1: Grid division. Based on the projection coordinate system, the monitoring area is divided into a set of regular grid cells C with the unmanned aerial vehicle aerial area as the range, each grid is recorded as c, for example, the side length is 50m. This grid division provides a unified spatial analysis unit. For each aerial image frame, according to the unmanned aerial vehicle position, attitude, camera internal and external parameters, the image pixels are projected to geographic coordinates, and the grid set covered by the frame image is calculated This process ensures the accurate correspondence between image data and geographic grid.

[0039] Step S1.2: Multi-frame coverage relationship and observation effort statistics. For each monitoring period t and each grid c, the image frame set covering the grid is counted. On this basis, the observation effort tensor is constructed, which is the key to quantifying the observation condition. Typically includes:

[0040] Frame number : represents the number of image frames covering the grid c, and the more the frame number usually means the more sufficient observation.

[0041] Height feature: including average height and height variance The average height affects the apparent size and details of the waterbird, while the height variance reflects the consistency of the observation height.

[0042] Viewing angle features: including average pitch / yaw angle and viewing angle diversity index, from the difference of flight line directions. The pitch / yaw angle affects the shape and background occlusion of the waterbird, while the viewing angle diversity can reflect the possibility of being observed from different angles.

[0043] Coverage : represents the proportion of the grid area covered by at least one frame, reflecting the completeness of the grid coverage by images. Span of observation time : represents the time difference between the first frame and the last frame, reflecting the length of time for observation in grid c, which may be related to the movement and behavior of waterbirds. This observation effort tensor will appear repeatedly in subsequent steps S3-S6, and is one of the main variables of the entire technical solution, used to quantify the influence of observation conditions on detection results.

[0044] Step S1.3: Extraction of ecological scene features. Extract the background and ecological factors of the grid synchronously, denoted as , such as vegetation type and coverage, water depth / water level inversion value, distance from shore, distance from human activity facilities, water surface reflectivity, texture complexity, etc. These ecological scene features are closely related to the habitat preference and detection difficulty of waterbirds, and will be used as input for the subsequent detection model and intensity field model together with the observation effort tensor .

[0045] It should be noted that the vegetation type and coverage are obtained by remote sensing / semantic segmentation.

[0046] Step S2: This step aims to identify waterbird individuals from the original images and aggregate multiple frames of information into stable trajectories to reduce the randomness of single-frame detection. Specifically, it includes:

[0047] Step S2.1: Frame-level detection and preliminary identification. Perform waterbird target detection and classification on each frame f of the image to obtain the candidate target set . Among them:

[0048] : Detection box position, used to locate the waterbird individual.

[0049] : Predicted species category, classifying waterbirds.

[0050] : Basic confidence, indicating the identification confidence of the detection model for the target.

[0051] : uncertainty features, e.g. prediction entropy from multi-model / multi-pass forward inference, difference of multiple prediction distributions, etc., to quantify the uncertainty of detection results. Each candidate target is attributed to a grid cell c according to its geoprojection position, to prepare for subsequent grid-level statistics.

[0052] S2.2: In-grid trajectory construction. Within each grid period , candidate targets are aggregated into trajectories according to their temporal order and spatial position. With constraints of inter-frame IOU (Intersection Over Union) and motion consistency, e.g. constant velocity / smooth trajectory, adjacent detection targets in consecutive frames are associated as the same waterbird trajectory . This trajectory aggregation leverages multi-frame overlapping information to associate detection results of the same waterbird in different frames, effectively reducing duplicate counting and missing detection. For each trajectory, the following statistics are computed:

[0053] trajectory length, i.e. number of frames: reflects the time duration of the waterbird being continuously observed.

[0054] class prediction consistency, i.e. frame-wise species distribution: to improve the accuracy of species classification.

[0055] average / variance confidence: to integrate confidence information from multiple frames.

[0056] multi-model prediction divergence: to further quantify the uncertainty of trajectory identification. Finally, each grid period obtains a set of trajectories , and the number of trajectories . Subsequent counting and inference are performed around the number of trajectories , instead of simply counting the number of bounding boxes in a single frame, thus naturally leveraging the information gain from multi-frame overlapping.

[0057] Step S3: Detection probability modeling based on observation effort, scene features and human verification. This step aims to establish a probability model to quantify the uncertainty of the detection process, and to calibrate it with a small amount of human verification data. Specifically, it includes:

[0058] Step S3.1: Definition of detection probability function. Within a grid period , the number of waterbird individuals that actually exist is , but due to factors such as shooting conditions, background, model performance, etc., only a part of them will form identifiable trajectories in multiple frames. Therefore, the detection probability function is defined, where h is the function, and p is the parameter. The function h satisfies:

[0059] as the frame number increases or the coverage As the probability of detection increases, the probability of detection rises monotonically because more observation opportunities increase the likelihood of detection.

[0060] The probability of detection decreases as the flight altitude increases, the viewing angle becomes too oblique, and the reflectivity increases, because these factors increase the difficulty of detection.

[0061] The detection probability varies depending on the habitat type, such as dense reeds versus open water, because different background complexities affect the performance of the detection algorithm.

[0062] This invention will , As a whole input, the shape is automatically learned from the data by fitting a function h using logistic regression, basis function expansion, or a small neural network, and the physical meaning is guaranteed by means of monotonic constraints, such as based on a monotonic neural network or by adding monotonic regularization.

[0063] Step S3.2: Multi-frame trajectory detection model. Given a given number of real individuals... At that time, assume that the probability of each individual forming a trajectory at least once during that period is... Then the distribution of trajectory numbers can be modeled as: ;

[0064] in, The FP noise term, representing the portion of the real individual that is detected, can be modeled as... , Similarly, it can be... , It is obtained by fitting with another function g. The FP noise term can be the number of false detection trajectories.

[0065] Step S3.3: Manually check grid calibration parameters In a portion of the sampling grid period, fine manual annotation was performed to obtain the true number of individuals. And obtain the trajectory number from S2 The parameters are analyzed using the maximum likelihood or EM algorithm. Make an estimate: Where A is the set of manually reviewed samples. This step also utilizes the multi-frame observation effort tensor. Scene characteristics The number of trajectories obtained by automatic recognition The actual number of manually labeled individuals Four types of information are used to jointly constrain the detection probability and false detection intensity through a probability model, and an observation model is solved by maximum likelihood / EM.

[0066] Step S4: This step aims to infer the true number of waterbirds from grid cells that have not been manually inspected, using the detection model calibrated in S3. Specifically, it includes:

[0067] Step S4.1: True number prior model. For the majority of grid periods that are not manually reviewed , only the observed number of trajectories , and . Assume the true number follows a Poisson distribution: where is the intensity parameter of the grid period, which is further modeled in steps S5, S6.

[0068] Step S4.2: Observation distribution. Given and , the trajectory number probability is determined by the detection probability function of step S3 and the false detection model.

[0069] Step S4.3: Posterior inference. The posterior distribution of the true number is obtained by Bayes' theorem: ; for each grid period, the expectation of can be computed by the EM algorithm, or the posterior distribution is directly enumerated in a limited range of , e.g., 0-N_max, to obtain the true number estimate and its variance . This step uses both the detection probability determined by of step S3 and the number prior in a joint probability model; since the posterior variance is computed rather than subjectively set, the credibility of the grid observation information can be controlled in steps S5, S6.

[0070] Step S5: This step aims to utilize spatial adjacency and ecological factor information to jointly infer the grid intensity parameter to compensate for the limited data of individual grids. Specifically, it includes:

[0071] Step S5.1: Spatial-ecological-observation joint objective. Define the intensity field to be inferred, and use the following objective function to optimize the inference: ;

[0072] where:

[0073] is the posterior variance of S4, which depends on and the detection model. The more sufficient the observation and the more frames, the smaller the variance, and the observation constraint naturally has a larger weight.

[0074] N is the set of grid adjacency pairs, The adjacency relationship determines the weights of the observation terms, which are used to encourage adjacent grids to have similar intensity.

[0075] To describe the constraints of the ecological factors on the intensity, we can use a regression model to construct the residual, which is used to introduce the ecological prior knowledge. The empirical relationship between water depth and vegetation type can be used to construct the residual of the regression model, which is used to introduce the ecological prior knowledge.

[0076] Note that the observation term weights in this step are derived from the statistical inference results from S4 , and the spatial smoothing and ecological gradient are used as regularization terms, which are naturally balanced by the optimization process of minimizing .

[0077] Step S5.2: Intensity field estimation. The intensity field estimate is obtained by variational inference or iterative optimization. In this stage, the role of the observation term is implicitly reflected by : grids with high observation effort have stronger constraints on ; grids with low observation effort are more dependent on adjacent grids and ecological priors.

[0078] Step S6: This step aims to integrate multi-period monitoring data and use state-space models for temporal updating to obtain more stable population distribution estimates and trend interpretation. Specifically, it includes:

[0079] Step S6.1: State evolution equation, i.e., ecological process. Wetland monitoring is usually conducted across multiple periods, and the observation conditions and true population in different periods are changing. The state evolution equation is established as follows: ; where is the intensity of the previous period, can be a function based on the changes in ecological factors to describe the impact of water level rise, vegetation succession, etc. on intensity; is the process noise, representing random fluctuations in ecological processes.

[0080] Step S6.2: Observation equation. The obtained in S5 is considered as a noisy observation of the true state: where , the covariance of is determined by the fitting residual in S4 and the fitting residual in S5, which is still affected by .

[0081] Step S6.3: Filtering and smoothing. Use Kalman filtering, extended Kalman filtering, or particle filtering methods to perform temporal updating for each grid.

[0082] When a period is poor, i.e., the observation effort is weak, is larger, the filter relies more on the state evolution equation, i.e., it trusts the historical trend and the ecological model more.

[0083] When a period is good, i.e., the observation effort is sufficient, is smaller, the filter relies more on the current period observation inference, i.e., it trusts the current observation data more. Finally, the posterior state of each grid period is obtained, which constitutes a multi-period stable population intensity field sequence. The total number of areas with confidence intervals can be further output over time, and the trend comparison and anomaly detection of different areas, such as a sudden and substantial decrease in the intensity of a protected area, are performed. At the time series level, the present application unifies the multi-period observation inference (controlled uncertainty) with the ecological evolution model and the historical state into a state space framework; the filter algorithm automatically adjusts whether to trust the historical prediction or the current observation according to the observation variance.

[0084] In Embodiment Two, the present embodiment provides a wetland waterbird automatic identification and statistical system based on unmanned aerial vehicle aerial images, as shown in the figure, comprising: Figure 1 The grid observation construction module 100 is used to perform the S1 step in Embodiment One described above. The module further comprises:

[0085] The grid division sub-module 110 is used to divide the aerial vehicle aerial area into a set of regular grid units C, and calculate the grid set covered by each image frame.

[0086] The multi-frame observation effort tensor construction sub-module 120 is used to, for each monitoring period t and each grid c, count all image frame sets

[0087] covering the grid, construct the observation effort tensor , and extract the ecological scene features of each grid c . The identification result correction and hierarchical statistical inference module 200 is used to perform the S2, S3 and S4 steps in Embodiment One described above. The module further comprises:

[0088] The multi-frame detection and trajectory generation sub-module 210 is used to perform waterbird target detection and classification on each image frame to obtain a candidate target set

[0089] , and, in each grid period , aggregate the candidate targets according to time sequence and spatial position to obtain a trajectory set and the number of trajectories .

[0090] ​​Detection Probability Modeling and Parameter Calibration Submodule 220: Used to define the detection probability function , number of trajectories Distribution modeling is And parameters were verified through manual sample inspection. Perform calibration.

[0091] Grid True Quantity Hierarchical Inference Submodule 230: Used to establish true quantities Poisson prior model The true quantity is obtained using Bayes' theorem. The posterior distribution is obtained, and the posterior distribution is calculated using the EM algorithm or finite enumeration method to obtain the true quantity estimate. and its variance .

[0092] Spatial-ecological-temporal joint inference module 300: used to execute steps S5 and S6 in Embodiment 1 above. This module further includes:

[0093] Spatial-ecological intensity field inference submodule 310: used to define the intensity field And adopt the objective function Optimization inference is performed, and the intensity field estimate is obtained through variational inference or iterative optimization. .

[0094] Multi-stage state space update submodule 320: used to establish state evolution equations and observation equations The time series update of each grid is performed using methods such as Kalman filtering, extended Kalman filtering, or particle filtering to obtain the posterior state of each grid period. .

[0095] The statistical results visualization output submodule 330 is used to visualize the final statistical results, such as the curve of the total number of regions changing over time (with confidence intervals), trend comparison of different regions, and anomaly detection.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended 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. Such 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 method for automatic identification and statistical analysis of wetland waterbirds based on UAV aerial imagery, characterized in that, Includes the following steps: Step S1: Divide the drone aerial photography area into regular grids. For each monitoring period t, count the set of image frames covered by each grid, construct an observation effort tensor containing the number of frames, altitude, viewpoint, coverage and time span, and extract the ecological scene features of each grid. The ecological scene features include vegetation type, water depth, water level, distance from shoreline, distance from human activity facilities, water surface reflectivity and texture complexity. Step S2: Perform waterbird target detection and classification on each frame of image, and obtain the detection box, species category, confidence and uncertainty feature information of each candidate target. The uncertainty feature information is the prediction entropy obtained by multiple models or multiple forward inferences. In each grid, the detection results of multiple frames are associated in time and space order to form a trajectory set, and the number of trajectories in the grid is obtained. Step S3: Define a detection probability function based on the observation effort tensor and ecological scene characteristics. Model the number of trajectories in each grid as a binomial distribution with the actual number and the detection probability function as parameters and superimpose Poisson false detection noise. Then, use manually checked samples to calibrate the model parameters. Step S4: Establish a Poisson prior model of the true number of waterbirds in each grid, perform hierarchical Bayesian inference, combine trajectory observation, observation effort and scene features according to the calibrated detection probability model to obtain the posterior distribution of the true number N of waterbirds in the grid, and calculate the posterior through the algorithm to estimate the true number and its variance. Step S5: Define the intensity field model of waterbird distribution, incorporate the observation results, spatial proximity smoothing constraints and ecological factors into the objective function, and solve the objective function through variational inference or iterative optimization to obtain the optimal intensity estimate of each grid. In step S5, the objective function includes an observation constraint term consisting of the true quantity estimate and its variance weighted, a spatial smoothing term that penalizes the intensity difference between adjacent grids based on grid adjacency, and an ecological factor term that predicts the intensity based on ecological scene features and constrains the difference between the predicted intensity and the current intensity. The objective function is solved using a gradient descent-type iterative algorithm to update the intensity variables of each grid.

2. The automatic identification and statistical method for wetland waterbirds based on UAV aerial imagery according to claim 1, characterized in that: In step S1, the regular grid division includes dividing the drone aerial photography area into multiple grid units according to a preset spatial resolution. The observation effort tensor includes the number of aerial frames, average flight altitude, camera pitch angle range, effective coverage area ratio, and continuous observation time of each grid unit during the monitoring period. Ecological scene features such as vegetation type, water depth, water level, distance from shoreline, distance from human activity facilities, water surface reflectivity, and texture complexity are extracted for each grid unit.

3. The automatic identification and statistical method for wetland waterbirds based on UAV aerial imagery according to claim 2, characterized in that: In step S2, when performing waterbird target detection and classification on each frame of image, the waterbird target detection and classification outputs the detection box position, species category and confidence score of candidate targets based on a pre-trained or transfer learning convolutional neural network model, and only retains candidate targets with a confidence score greater than a preset threshold to participate in the subsequent multi-frame trajectory association and trajectory number statistics within the grid cell.

4. The automatic identification and statistical method for wetland waterbirds based on UAV aerial imagery according to claim 2, characterized in that... ； When performing multi-frame trajectory association in step S2, only candidate targets whose spatial position overlap between adjacent aerial frames within the same grid cell exceeds a preset intersection-union ratio threshold, whose motion direction and speed change meet continuous motion constraints, and whose time interval does not exceed a preset inter-frame interval are merged into the same waterbird trajectory, and each candidate target is limited to belonging to at most one waterbird trajectory.

5. The automatic identification and statistical method for wetland waterbirds based on UAV aerial imagery according to claim 4, characterized in that: In step S3, the detection probability function is represented by a parameterized function model that takes the observation effort tensor and the ecological scene features as input. The parameterized function model is a logistic regression model or a small neural network model, and a monotonicity constraint is applied to the model parameters so that the detection probability increases monotonically with the number of aerial frames and does not increase monotonically when the flight altitude exceeds a preset upper limit.

6. The automatic identification and statistical method for wetland waterbirds based on UAV aerial imagery according to claim 5, characterized in that, In step S3, when calibrating model parameters using manual survey samples, several monitoring periods are selected from grid cells of different observation effort levels and different ecological scene types as manual survey grids. The aerial images in each manual survey grid are manually labeled frame by frame to obtain the actual number of waterbirds. The actual number and the corresponding number of trajectories are combined to form a training sample set for updating model parameters using maximum likelihood estimation or expectation-maximization algorithm.

7. The automatic identification and statistical method for wetland waterbirds based on UAV aerial imagery according to claim 1, characterized in that, In step S4, the Poisson prior model uses the intensity parameter of the grid cell as the prior mean. The detection probability function obtained in step S3, the false detection noise model, and the number of trajectories are used to form the observation model. The posterior distribution and its expected value are solved by enumerating within the range of candidate true number values ​​or by using the expectation-maximization algorithm. The corresponding posterior variance is recorded synchronously when storing the estimated true number value.

8. The automatic identification and statistical method for wetland waterbirds based on UAV aerial imagery according to claim 1, characterized in that, Step S5 is followed by step S6, in which an intensity state sequence is established for each grid cell according to the monitoring period, the intensity state of adjacent monitoring periods is correlated through the state transition equation, and the intensity state and ecological factor changes of the previous monitoring period are used as state prediction inputs. The intensity estimate and variance obtained in step S5 are used as observation inputs, and Kalman filtering or particle filtering is used to complete the intensity state update.

9. An automatic wetland bird identification and statistics system based on UAV aerial imagery, used to implement the method described in any one of claims 1-8, characterized in that, include: The grid observation construction module is used to divide the aerial photography area into grid cells, statistically analyze the set of image coverage frames to construct the observation effort tensor, and extract the ecological scene features of each grid cell. The waterbird detection and trajectory aggregation module is used to detect and classify waterbird targets in aerial images, and associate the detection results of multiple frames to form a trajectory set and the number of trajectories; The detection probability modeling and parameter calibration module is used to establish a waterbird detection probability model based on the observation effort tensor and ecological scene characteristics, and to calibrate the model parameters through manual sample inspection. The true number inference module is used to establish a prior model of the true number of waterbirds, and combine the detection probability model with trajectory observation to use Bayesian inference to calculate the posterior estimate of the true number and its variance. The spatial ecological intensity field inference module is used to construct a spatial intensity field model, and to perform optimization inference by integrating observation data, spatial proximity relationships and ecological factor constraints to obtain the estimated results of waterbird distribution intensity.

Images