A method for evaluating wetland ecological status based on bird integrity

Abstract

The present invention discloses a method for evaluating the ecological status of wetlands based on bird integrity. The method provided by the present invention comprises the following steps: collecting the sounds and images of birds to be identified in a preselected time period and a preselected wetland area; inputting the sounds of the birds to be identified into a hidden Markov model, and outputting the bird species corresponding to the sound identification; and inputting the images of the birds to be identified into a target recognition model, and outputting the bird species corresponding to the image identification; merging the bird species identified by sound in the preselected wetland area and the bird species identified by image, and taking the number of bird species obtained after merging as the total number of bird species in the preset wetland area; calculating the bird integrity index according to the total number of bird species and the species conservation coefficient; and judging the ecological status level of the preselected wetland area according to the bird integrity index. The present invention can accurately and comprehensively evaluate the habitat status of the wetland area, thereby providing a rationalization indicator for ecological restoration and protection.

Description

Technical Field

[0001] The present invention relates to the technical field of ecological monitoring, and in particular to a method for evaluating wetland ecological conditions based on bird integrity. Background Art

[0002] Wetlands are a unique ecosystem with multiple functions and values. They are one of the natural environments with rich biodiversity and on which human society depends for survival and development.

[0003] Birds are the most active and eye-catching components of wetland ecosystems, and they play a vital role in wetland energy flow and maintaining the stability of the ecosystem. Changes in wetland ecological conditions will directly affect the ecology and population development of birds, so birds can be used as an objective biological indicator to monitor changes in wetland ecological conditions.

[0004] Traditional bird identification methods include video monitoring and manual identification, but the target range of video monitoring is limited; while manual identification is difficult to implement and inefficient.

[0005] Therefore, there is an urgent need to improve the comprehensiveness, accuracy and efficiency of bird identification, so as to realize the evaluation of wetland ecological status. Summary of the invention

[0006] In view of the problems existing in the above-mentioned prior art, the present invention provides a method for evaluating the ecological status of wetlands based on bird integrity.

[0007] In order to achieve the above objectives, in a first aspect, the present invention provides a method for evaluating wetland ecological status based on bird integrity, which evaluates wetland ecological status according to the total number of bird species and the conservation coefficient in a preselected wetland area.

[0008] Specifically, the steps include:

[0009] Step S101, collecting sounds and images of birds to be identified in a preselected time period and a preselected wetland area;

[0010] Step S102: input the sound of the bird to be identified into the hidden Markov model, and output the bird species corresponding to the sound identification; and input the image of the bird to be identified into the target recognition model, and output the bird species corresponding to the image identification;

[0011] Step S103, merging the bird species identified by sound and the bird species identified by image in the pre-selected wetland area, and using the number of bird species obtained after merging as the total number of bird species in the preset wetland area;

[0012] Step S104, calculating the bird integrity index according to the total number of bird species and the conservative coefficient;

[0013] Step S105: Determine the ecological status level of the pre-selected wetland area according to the bird integrity index.

[0014] The method for evaluating wetland ecological status based on bird integrity of the present invention has the following beneficial effects:

[0015] (1) The method for evaluating wetland ecological status by bird integrity provided by the present invention utilizes sound and image to jointly judge, thereby improving the recognition rate of birds in wetland areas;

[0016] (2) The method for evaluating wetland ecological status by bird integrity provided by the present invention can avoid human intervention, effectively reduce labor costs, and improve the accuracy and efficiency of bird identification;

[0017] (3) The method for evaluating wetland ecological status by bird integrity provided by the present invention can accurately obtain the bird integrity index of the wetland area, thereby accurately and comprehensively obtaining the habitat conditions of the wetland area, thereby providing rational indicators for ecological restoration and protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a process flow chart of the method for evaluating wetland ecological status based on bird integrity of the present invention. DETAILED DESCRIPTION

[0019] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0020] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "includes..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0021] Among them, birds can be used as an objective biological indicator to monitor changes in wetland ecological conditions. However, how to accurately and comprehensively identify bird species and bird integrity in wetlands is an urgent problem to be solved.

[0022] In order to solve the above problems, the present invention provides a method for evaluating wetland ecological status based on bird integrity, which evaluates the wetland ecological status according to the total number of bird species and the conservative coefficient in a preselected wetland area.

[0023] Specifically, Figure 1 As shown, the method may include the following steps:

[0024] Step S101: collecting sounds and images of birds to be identified in a preselected time period and a preselected wetland area.

[0025] In the present invention, the preselected time periods are preferably different time periods in the same season, such as every Tuesday and Sunday in summer, so as to ensure that the birds appearing in the preselected wetland area in that season can be fully identified, and then the ecological conditions of the preselected wetland area in that season can be judged.

[0026] At present, bird identification is usually based on sound or image. However, due to factors such as light, object posture, perspective, occlusion, background interference, etc., the morphology of the same bird species can vary greatly, and the sound of the same bird species can also vary. Therefore, the present invention uses sound and image to jointly judge, thereby improving the recognition rate of birds in wetland areas.

[0027] In a preferred embodiment of the present invention, before step S101, the following steps may be included:

[0028] Step S100-1: Divide the pre-selected wetland area into multiple sub-areas.

[0029] At present, the monitoring of water birds is carried out by fixed-point counting method; for other types of birds, the line sampling method is used for monitoring, that is, the observer walks along a fixed route and records the birds seen on both sides of the line sampling. The spacing between the lines should be more than 0.5km, the length of the line sampling should be 3 to 5km, and the speed of observation should be 1 to 3km / h. When a bird is found, the name, number and other information are recorded.

[0030] It can be seen that the above process is greatly affected by human factors, and the monitored data is relatively one-sided. Therefore, in order to ensure the accuracy and comprehensiveness of bird species identification, the present invention divides the pre-selected wetland area into multiple sub-areas, and preferably the areas of the multiple sub-areas are not exactly the same.

[0031] Research has found that if the wetland is mainly composed of water surface, the sub-areas can be divided into larger areas, such as 10-30m*10-30m; if the wetland is mainly composed of land covered with vegetation, the sub-areas can be divided into smaller areas, such as 2-8m*2-8m. This can reduce the impact of external factors such as vegetation obstruction, human damage and monitoring distance, thereby enabling more accurate and comprehensive identification of bird species and distribution in each sub-area.

[0032] Step S100 - 2 : Setting an image acquisition device and a sound acquisition device in each sub-area.

[0033] In order to ensure the accuracy and comprehensiveness of bird identification, image acquisition equipment and sound acquisition equipment are set in each sub-area. Since the area of ​​each sub-area is not exactly the same, the number of image acquisition equipment and / or sound acquisition equipment set in each sub-area is also not exactly the same.

[0034] According to research, there are more interference factors in the sub-areas with smaller areas, such as vegetation obstruction, etc., and setting more image acquisition devices and / or sound acquisition devices can improve the accuracy of bird species identification. Therefore, preferably, the number of image acquisition devices and / or sound acquisition devices set in the sub-area with larger area is less than or equal to the sub-area with smaller area.

[0035] Preferably, the image acquisition device includes a binocular camera and a monocular camera. In order to save costs and improve the accuracy and comprehensiveness of bird identification, the binocular cameras are mostly set in sub-areas with smaller areas, and the number of binocular cameras is proportional to the coverage of vegetation in the corresponding sub-areas, and the monocular cameras are only set in sub-areas with larger areas to facilitate monitoring of waterfowl.

[0036] For example, the image acquisition device can use a structure similar to that of patent application number 201920096153.9, which has high confidentiality, adjustable height and angle, and strong adaptability.

[0037] It has been found through research that the image acquisition process is subject to more external interference, such as weather, lighting, etc. Preferably, the number of image acquisition devices set in each sub-area is greater than or equal to the number of sound acquisition devices, so as to ensure the accuracy of image acquisition.

[0038] For example, 5 to 10 image acquisition devices and 1 to 3 sound acquisition devices are set up in sub-areas with a larger area, and 10 to 15 image acquisition devices and 1 to 5 sound acquisition devices are set up in sub-areas with a smaller area, and the equipment monitoring range must cover most of the wetland area.

[0039] In the present invention, the image acquisition device and the sound acquisition device are provided at the same time, which can avoid acquisition errors or omissions in the image acquisition device or the sound acquisition device, and ensure the accuracy and completeness of bird species identification.

[0040] Step S100-3, stitching the n-th frame image captured by the image acquisition device in each sub-area respectively, and judging whether the preset image coincidence degree is satisfied according to the stitching result; if not, adjusting the image acquisition device until the m-th frame image captured by the image acquisition device satisfies the preset image coincidence degree, where n is an integer greater than or equal to 1, and m is an integer greater than or equal to n.

[0041] Preferably, the stitching process includes: extracting feature points from the nth frame image. Specifically, the feature points are important points in the image, such as contour points, highlight points in darker areas, dark points in brighter areas, and the like.

[0042] More preferably, the process of extracting feature points from the nth frame image includes:

[0043] Divide each n-th frame image into regions to obtain multiple regions;

[0044] Extract feature points from each region in each n-th frame of the image. For example, use the ORB algorithm and OFAST corner point detection to extract feature points (corner points) from each region;

[0045] If there are areas without feature points, the OFAST corner point detection threshold is lowered so that feature points can be extracted from the corresponding areas.

[0046] Specifically, the present invention performs region division according to the actual situation of the n-th frame image. For example, the region division can be performed according to the size of the n-th frame image or the number of feature points in the image, and the number of regions divided for each n-th frame image may be different. When the size of the n-th frame image is small or the n-th frame image has fewer feature points, fewer regions can be divided, and when the size of the n-th frame image is large or the n-th frame image has more feature points, more regions can be divided.

[0047] For example, when three n-th frames of images are collected, if the second n-th frame has fewer feature points than the first and third n-th frames, the number of divided areas in the first and third n-th frames is required to be larger than that in the second n-th frame.

[0048] In this process, it can be ensured that feature points can also be extracted from areas with weaker textures, while increasing the amount of information on each n-th frame image and improving the stitching quality.

[0049] Preferably, before extracting feature points from the n-th frame image, the process further includes: preprocessing the n-th frame image.

[0050] More preferably, camera calibration and image denoising are performed on each n-th image frame.

[0051] Specifically, the existing "Zhang Zhengyou calibration method" is used to calibrate the camera.

[0052] Specifically, the existing median filtering method is used to denoise the image, and in order to highlight the filtering result, salt and pepper noise is added to the image to be filtered.

[0053] More preferably, the process of extracting feature points from the nth frame image further includes:

[0054] The area of ​​each image is re-divided into multiple small areas; the subdivision is stopped until each small area contains at most 3 to 5 feature points.

[0055] Exemplarily, the n-th frame image is first divided into regions of size 20*20, and preliminary OFAST corner point detection is performed, and then quadtree re-division is performed in each region to perform OFAST corner point detection.

[0056] The best quality feature points are selected from each small area to ensure the uniform distribution of feature points.

[0057] Among them, the feature points are obtained by the OFAST corner point detection algorithm. When the intensity deviation between the feature point (corner point) and the adjacent pixel points is larger and the more points with deviations there are, the better the quality of the feature point is, otherwise the quality is worse.

[0058] This process can improve the uniformity of feature point distribution and prevent feature points from being too concentrated or clumped.

[0059] More preferably, the process of extracting feature points from the nth frame image further includes: using the BRIEF algorithm to create a feature vector as a descriptor for each feature point.

[0060] Preferably, the stitching process further includes: using a nearest neighbor method to perform rough matching on the feature points to obtain registration feature points.

[0061] More specifically, the Hamming distance is used for rough matching, and its principle is:

[0062] (1) If the number of identical elements in the corresponding bits of two feature codes is less than 128, they are definitely not registration feature points;

[0063] (2) The feature points on one n-th frame image and the feature points on another n-th frame image that have the largest number of identical elements in the corresponding bits of the feature code are a pair of registered feature points.

[0064] Assume that there are many feature point descriptors A to Z waiting for registration. Take A as an example. Among the remaining feature point descriptors, suppose there are 5 feature points with the same number of elements in the feature code greater than 128, and the other 18 feature points with the same number of elements less than 128, then the registration point can only be among these 5 feature points.

[0065] In addition, for these five feature points, assuming they are C, D, E, F, and G, E has the largest number of elements in common with A, so E is used as the registration feature point of A.

[0066] Due to the difference in brightness and exposure, there is a certain grayscale difference between the two n-th frame images, and there is an obvious boundary near the stitching line. In the present invention, the gradient weight method is preferably used to eliminate the gap to achieve a smooth transition between the two n-th frame images.

[0067] Specifically, the splicing process also includes:

[0068] The projection transformation matrix of the n-th frame image is calculated based on the registered feature points to obtain the spatial position relationship between the images; that is, the projection transformation matrix of the n-th frame image can be calculated based on the registered feature points using the least squares method or matrix inversion, thereby obtaining the spatial position relationship between the n-th frame images.

[0069] According to the spatial position relationship, the n-th frame image is fused using the gradient weight method.

[0070] In the present invention, the direction directly behind the lens is the positive direction of the Z axis, the horizontal rightward movement is the positive direction of the X axis, and the vertical upward movement is the positive direction of the Y axis. The collected image is an image in the XY axis direction.

[0071] More preferably, the weight is obtained by formula 1:

[0072]

[0073] Among them, δ(x i ,y i ) represents the reference image and the nth frame image to be spliced. i ,y i ) pixel weights, θ represents the angle between the reference image and the center line of the nth frame image to be spliced ​​and the X-axis direction; x min 、x max They represent the coordinate values ​​of the left and right boundaries of the overlapping area between the reference image and the nth frame image to be stitched, respectively. min ,y max They respectively represent the coordinate values ​​of the lower and upper boundaries in the Y-axis direction of the overlapping area between the reference image and the n-th frame image to be stitched, and i is an integer greater than or equal to 1.

[0074] In the present invention, the reference image may be a captured standard image, or an image obtained by splicing at least two n-th frame images, or an n-th frame image to be spliced.

[0075] More preferably, the image I obtained by fusing the reference image with the nth frame image to be spliced ​​is expressed by Formula 2:

[0076] I(x i ,y i )=δ(x i ,y i)*A(x i ,y i )+[1-δ(x i ,y i )]*B′(x i ,y i ) Formula 2

[0077] I(x i ,y i ) represents the first (x i ,y i ) pixel values ​​in the fused image I (can be grayscale values ​​or RGB colors); A(x i ,y i ) represents the first (x i ,y i ) pixel value in the reference image; B′(x i ,y i ) represents the smoothed (x i ,y i ) pixel values ​​in the smoothed image to be spliced, where B′(x i ,y i ) is expressed by formula 3:

[0078] B′(x i ,y i )=s*B(x i ,y i )+d Formula 3

[0079] B(x i ,y i ) represents the first (x i ,y i ) pixel value in the nth frame image to be stitched. s and d are both manually set parameters and can be adjusted according to the stitching effect.

[0080] For example, suppose there are three image acquisition devices (cameras A to C), the first frame image A acquired by camera A, the first frame image B acquired by camera B, and the first frame image C acquired by camera C. From left to right in the X-axis direction, they are image A, image B, and image C.

[0081] At this time, the reference image may be a standard image or image A. When the reference image is image A and the image to be stitched is image B, image A and image B are stitched according to the above formulas 1 to 3 to achieve smooth stitching of image A gradually transitioning from the left overlapping area to the right overlapping area, and then image A and image B are used as reference images, and the reference image and image C are stitched according to the above formulas 1 to 3 to achieve smooth stitching of the reference image gradually transitioning from the left overlapping area to the right overlapping area.

[0082] This method can be used to adjust the brightness of the entire spliced ​​image to ensure that the brightness difference between the overlapping parts and the remaining parts is not large, thereby effectively eliminating the splicing seams in both directions.

[0083] The δ(x i ,y i ) into the formula I(x i ,y i ), each pixel in the reference image and the n-th frame image to be spliced ​​is calculated, and finally the fused image I can be obtained.

[0084] Preferably, the preset image coincidence (overlapping area between images) corresponding to the n-th frame image captured by each sub-region image acquisition device is not the same. Preferably, the preset image coincidence is inversely proportional to the area size of the corresponding sub-region, that is, the image coincidence corresponding to the sub-region with a large area is less than the image coincidence corresponding to the sub-region with a small area. For example, the preset image coincidence corresponding to the sub-region with a large area is 3-10%, preferably 5-8%, and the preset image coincidence corresponding to the sub-region with a small area is 10-20%, preferably 14-17%. In this way, each position of the sub-region can be monitored to avoid external interference.

[0085] According to the stitching result of the nth frame image, the image overlap (image overlapping area) of the nth frame image can be obtained. When the image overlap of the nth frame image does not meet the preset image overlap, it is necessary to adjust the multi-image acquisition devices (quantity, position and angle, etc.). That is, the image acquisition devices of each sub-area are arranged in different arrays, so that each sub-area can be fully monitored while reducing the image acquisition devices as much as possible, so as to accurately obtain all the information of the birds in the preset wetland area.

[0086] Preferably, step S101 may further include: using a sound acquisition device to acquire the sound of the bird to be identified in each sub-area; and using an image acquisition device that meets a preset image overlap to acquire the image of the bird to be identified in each sub-area.

[0087] In the present invention, the n-th frame image captured by the image acquisition device is spliced ​​to ensure that a complete video image of each sub-area is obtained, thereby obtaining more comprehensive bird information, quantity and more accurate bird morphology in each sub-area as much as possible.

[0088] Step S102: input the sound of the bird to be identified into the hidden Markov model, and output the bird species corresponding to the sound identification; and input the image of the bird to be identified into the target recognition model, and output the bird species corresponding to the image identification.

[0089] In the present invention, the training process of the hidden Markov model and the target recognition model is:

[0090] (1) Collect the sounds and images of birds in pre-selected wetland areas at different time periods.

[0091] (2) Use sound to train the hidden Markov model and use images to train the target recognition model.

[0092] Since different frequency bands in the sound reflect different information, the importance of feature parameters extracted from signals in different frequency bands is also different. Research has found that decomposing the original sound into the sum of signals in each frequency band, extracting feature parameters of signals in different frequency bands respectively, and then fusing the feature parameters of signals in different frequency bands using weighted values, and finally using the fused feature parameters for training and recognition can improve the recognition rate.

[0093] The sounds of birds have corresponding highest and lowest frequencies. According to the difference between the highest and lowest frequencies, the sounds of birds are divided into 5 to 10 segments, and WT-MFCC feature parameters are extracted for each frequency band.

[0094] For example, the singing frequency of the lark in the early morning is between 2076 and 5167 Hz. If the frequency is divided into segments of 500 Hz, we can get six frequency bands: 2076-2576 Hz, 2576-3076 Hz, 3076-3576 Hz, 3576-4076 Hz, 4076-4576 Hz, and 4576-5167 Hz.

[0095] It has been found through research that if the weighted value corresponding to the low frequency band (the first 3 to 5 bands) is set larger, the recognition rate will decrease; if the weighted value corresponding to the high frequency band (the last 2 to 5 bands) is set larger, a better recognition effect can be obtained. Therefore, in the present invention, the weighted value corresponding to the high frequency band (the last 2 to 5 bands) is set to be greater than or equal to the weighted value corresponding to the low frequency band (the first 3 to 5 bands). For example, the weighted value corresponding to the high frequency band is set to 3 to 6, and the weighted value corresponding to the low frequency band is set to 1 to 3, and the sound recognition rate can reach more than 90%.

[0096] In the present invention, the existing hidden Markov model is used, and the training process is consistent with the existing training process. The hidden Markov model can refer to the document "Lu Jian, Chen Yisong, Sun Zhengxing, et al. Automatic audio classification based on hidden Markov model [J]. Journal of Software, 2002, 13 (8): 1593-1597." Its main process is: directly determine the corresponding bird species according to the input fused feature parameters, and output the bird species as the bird recognition result, so as to realize the automatic recognition of bird species, avoid manual intervention, effectively reduce labor costs, and improve the accuracy and efficiency of bird sound recognition.

[0097] Preferably, the position of the bird in the image is identified by a target recognition model to remove the background other than the bird. The target recognition model is preferably a CNN model.

[0098] Among them, the CNN model is selected from an existing CNN model, such as an existing Mask R-CNN convolutional neural network, and its training method is consistent with the existing training method. Those skilled in the art can train the CNN model according to the existing training method.

[0099] For example, the model structure of the Mask R-CNN convolutional neural network can refer to the document “He K, Gkioxari G, PDollár, et al. Mask R-CNN[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 42(2): 386-397.”

[0100] Step S103: merging the bird species identified by sound and the bird species identified by image in the pre-selected wetland area, and using the number of bird species obtained after merging as the total number of bird species in the preset wetland area.

[0101] Specifically, corresponding bird sound recognition results and bird image recognition results can be obtained in each sub-area, and the results are merged. The merged number of bird species is used as the total number of bird species in the pre-selected wetland area, avoiding missed detection or false detection, thereby obtaining bird information in the pre-selected wetland area more comprehensively and accurately.

[0102] Taking the birds in the river and lake habitats within the Beijing Economic and Technological Development Zone as an example, it is divided into five sub-areas: Sanhaizi East Road, No. 3 Rubber Dam, Maju Bridge, Jinghai Road, and Exit Section. A total of 25 bird species were identified in the five sub-areas using images, with 14, 18, 9, 9, and 7 species in each sub-area; a total of 24 bird species were identified using sound, with 14, 18, 8, 9, and 7 species in each sub-area. At this time, the two were combined to obtain 25 bird species, which was taken as the total number of bird species in the area.

[0103] Step S104: Calculate the bird integrity index based on the total number of bird species and the conservative coefficient.

[0104] Specifically, the Bird Integrity Index in It represents the average conservative coefficient of bird species in the pre-selected wetland area, which is the ratio of the sum of conservative coefficients of all bird species in the pre-selected wetland area to the total number of bird species, that is, q represents the total number of bird species in the pre-selected wetland area.

[0105] As can be seen from the formula, the total number of bird species and the conservative coefficient jointly determine the bird integrity index of the pre-selected wetland area. The more bird species there are in the wetland and the higher the conservative coefficient, the higher the bird integrity index, indicating that the ecological condition of the pre-selected wetland area is better and closer to the original habitat.

[0106] The assignment rules of the conservative coefficients of bird species are shown in Table 1:

[0107] Table 1

[0108]

[0109]

[0110] Among them, the conservative coefficient of bird species is assigned by 5 to 7 experts according to Table 1, field survey conditions and bird species characteristics, and then the highest and lowest values ​​are removed, and the average of the remaining values ​​is taken as the conservative coefficient C of the corresponding bird species. j .

[0111] Step S105: Determine the ecological status level of the pre-selected wetland area according to the bird integrity index.

[0112] In the present invention, a correspondence is established between the bird integrity index ABI and the ecological status level of the preselected wetland area, so that the ecological status of the preselected wetland area can be evaluated based on the bird integrity index ABI, thereby providing corresponding technical support for protecting or restoring the preselected wetland area.

[0113] Specifically, the corresponding relationship between the bird integrity index ABI and the habitat is established as shown in Table 2:

[0114] Table 2

[0115] ABI Habitat level Habitat characteristics 0-10.00 Difference There are basically no natural features 10.01-20.00 Poor Preserve some natural features 20.01-30.00 medium Preserve some natural features 30.01-40.00 good High-quality habitat ≥40.01 excellent Close to native habitat

[0116] Therefore, the present invention uses the bird integrity index of the wetland area obtained according to Table 2, so as to accurately and comprehensively evaluate the ecological status of the wetland area, thereby providing a rationalization indicator for ecological restoration and protection.

[0117] Example

[0118] Taking the birds in the river and lake habitats within the Beijing Economic and Technological Development Zone as an example, they are divided into five sub-areas: Sanhaizi East Road (approximately 12m*12m in area), No. 3 Rubber Dam (approximately 10m*10m in area), Majuqiao (approximately 11m*11m in area), Jinghai Road (approximately 10m*10m in area), and exit section (approximately 13m*13m in area). Ten image acquisition devices and two sound acquisition devices are first set up in each sub-area.

[0119] The first frame of image captured by the image acquisition device in each sub-area is stitched separately, and it is determined whether the image overlap reaches 7% based on the stitching result. Then, the number of image acquisition devices in the Sanhaizi East Road and exit section of the sub-area is adjusted to 8, and the number of image acquisition devices in other sub-areas remains unchanged, so that the image overlap of the image acquisition devices reaches 7%.

[0120] The specific process of splicing is as follows:

[0121] Divide each first frame image into regions to obtain multiple regions;

[0122] The ORB algorithm is used and OFAST corner point detection is used to extract feature points (corner points) in each area;

[0123] If there are areas without feature points, the OFAST corner point detection threshold is lowered so that feature points can be extracted from the corresponding areas.

[0124] Re-divide the area of ​​each first frame image so that there are at most 3 feature points in each small area;

[0125] Select the best quality feature points from each small area, and use the BRIEF algorithm to create a feature vector as a descriptor for each feature point;

[0126] The nearest neighbor method is used to roughly match the feature points to obtain the registration feature points.

[0127] The projection transformation matrix of the first frame image is calculated using the least square method according to the registered feature point pairs, thereby obtaining the spatial position relationship between the images.

[0128] According to the spatial position relationship, the first frame image is fused using the gradient weight method to obtain the splicing result.

[0129] Using sound collection equipment and image collection equipment in the five sub-areas, the sounds and images of birds in the five sub-areas were collected continuously every Monday and Friday from 2015 to 2019.

[0130] The hidden Markov model (described in the reference "Lu Jian, Chen Yisong, Sun Zhengxing, et al. Automatic audio classification based on hidden Markov model [J]. Journal of Software, 2002, 13(8): 1593-1597.") and the Mask R-CNN convolutional neural network (described in the reference "He K, Gkioxari G, P Dollár, et al. Mask R-CNN [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 42(2): 386-397.") were trained using the sounds and images of birds in the five sub-regions.

[0131] The sounds and images of birds to be identified were collected in five sub-areas throughout the day on Tuesdays and Sundays from May to June 2020, and the sounds of the birds to be identified were input into the trained hidden Markov model to output the bird species corresponding to the sound identification; the images of the birds to be identified were input into the trained target recognition model to output the bird species corresponding to the image identification;

[0132] The number of bird species identified by sound in the pre-selected wetland areas was counted (24 bird species were identified, and 14, 18, 8, 9 and 7 bird species were identified in the five sub-areas respectively); the number of bird species identified by image in the pre-selected wetland areas was counted (25 bird species were identified, and 14, 18, 9, 9 and 7 bird species were identified in the five sub-areas respectively).

[0133] The bird species identified by sound and the bird species identified by image are combined to obtain 25 bird species in the wetland area. At this time, 25 species are taken as the total number of bird species in the wetland area.

[0134] The bird integrity index was calculated based on the total number of bird species and the conservation coefficient.

[0135] Bird Integrity Index in It represents the average conservative coefficient of bird species in the region, which is the sum of the conservative coefficients of all bird species in the region divided by 25, that is, q represents the total number of bird species in the area.

[0136] Based on the bird integrity index, the ecological status level of the pre-selected wetland area is determined.

[0137] The specific experimental results are shown in Table 3:

[0138] Table 3

[0139]

[0140]

[0141] As can be seen from Table 3, the bird integrity index of the wetland area is low, that is, the ecological condition is poor. It is known that the ecological condition of the wetland area is at a poor level, which shows that the level judgment here is accurate, and the wetland area needs to be restored in time.

[0142] The present invention is described in detail above in conjunction with specific implementation methods and exemplary examples, but these descriptions cannot be understood as limiting the present invention. Those skilled in the art should understand that, without departing from the spirit and scope of the present invention, various equivalent replacements, modifications or improvements can be made to the technical solution of the present invention and its implementation methods, all of which fall within the protection scope of the present invention.

Claims

1. A method for evaluating wetland ecological status based on bird integrity, characterized in that: Evaluate the ecological status of wetlands based on the total number of bird species and conservative coefficients in pre-selected wetland areas; The following steps are involved: Step S101, collecting sounds and images of birds to be identified in a preselected time period and a preselected wetland area; Step S102: input the sound of the bird to be identified into a hidden Markov model, and output the bird species corresponding to the sound identification; and input the image of the bird to be identified into a target recognition model, and output the bird species corresponding to the image identification; Step S103, merging the bird species identified by sound and the bird species identified by image in the pre-selected wetland area, and using the number of bird species obtained after merging as the total number of bird species in the preset wetland area; Step S104, calculating the bird integrity index according to the total number of bird species and the conservative coefficient; Step S105, judging the ecological status level of the pre-selected wetland area according to the bird integrity index; The following sub-steps are also included before step S101: Step S100-1, dividing the pre-selected wetland area into multiple sub-areas; Step S100-2, setting an image acquisition device and a sound acquisition device in each sub-area; The areas of multiple sub-regions are not exactly the same. If the water surface is the main area, the sub-region is divided into a large area; if the land covered with vegetation is the main area, the sub-region is divided into a small area; The number of the image acquisition devices and / or sound acquisition devices provided in each sub-area is not completely the same, and the number of the image acquisition devices and / or sound acquisition devices provided in the sub-area with a larger area is less than or equal to the number of the sub-area with a smaller area; Step S100-3, respectively performing stitching processing on the n-th frame image of the image acquisition device in each sub-area, and judging whether the preset image coincidence degree is satisfied according to the stitching result, if not satisfied, adjusting the image acquisition device until the m-th frame image acquired by the image acquisition device satisfies the preset image coincidence degree, where m≥n≥1; The process of stitching includes: extracting feature points from the nth frame image; The projection transformation matrix of the n-th frame image is calculated according to the registration feature points to obtain the spatial position relationship between the images. According to the spatial position relationship, the n-th frame image is fused using the gradient weight method; The process of extracting feature points from the nth frame image includes: Divide each n-th frame image into regions to obtain multiple regions; Extract feature points from each region in each n-th frame of the image. Use the ORB algorithm and OFAST corner point detection to extract feature points from each region. If there are regions without feature points, lower the OFAST corner point detection threshold so that feature points can be extracted from the corresponding regions. The process of extracting feature points from the n-th frame image also includes: Re-divide the area of ​​each image into multiple small areas; stop dividing until each small area contains at most 3 to 5 feature points; The step of fusing the n-th frame image by using a gradient weight method according to the spatial position relationship includes: The direction directly behind the lens is the positive direction of the Z axis, the horizontal right movement is the positive direction of the X axis, and the vertical upward movement is the positive direction of the Y axis. The collected image is the image in the XY axis direction. The weights are obtained by formula 1: Among them, δ(x i ,y i ) represents the reference image and the nth frame image to be spliced. i ,y i ) pixel weights, θ represents the angle between the reference image and the center line of the nth frame image to be spliced ​​and the X-axis direction; x min 、x max They represent the coordinate values ​​of the left and right boundaries of the overlapping area between the reference image and the nth frame image to be stitched, respectively. min ,y max Respectively represent the coordinate values ​​of the lower boundary and the upper boundary of the Y-axis direction of the overlapping area between the reference image and the n-th frame image to be spliced, and i is an integer greater than or equal to 1; The reference image is a standard image or an image obtained by splicing at least two n-th frame images; The image I is obtained by fusing the reference image with the nth frame image to be spliced ​​using formula 2: I(x i ,y i ) = δ(x i ,y i ) * A(x i ,y i ) + [1 - δ(x i ,y i )] * B′(x i ,y i ) Equation 2 I(x i ,y i ) represents the first (x i ,y i ) pixel gray value or RGB color in the fused image I; A(x i ,y i ) represents the first (x i ,y i ) pixel value in the reference image; B′(x i ,y i ) represents the smoothed (x i ,y i ) pixel values ​​in the smoothed image to be spliced, where B′(x i ,y i ) is expressed by formula 3: B′(x i ,y i )=s*B(x i ,y i )+d Formula 3 B(x i ,y i ) represents the first (x i ,y i ) pixel value in the nth frame image to be stitched, s and d are adjustment parameters adjusted according to the stitching effect; In step S102, the original sound is decomposed into signals of each frequency band, and the signals of different frequency bands are added, and characteristic parameters are extracted respectively, and then the characteristic parameters of the signals of different frequency bands are fused using weighted values, and finally the fused characteristic parameters are used for training and recognition; The sounds of birds have corresponding highest and lowest frequencies. According to the difference between the highest and lowest frequencies, the sounds of birds are divided into 5 to 10 segments, and the weighted value corresponding to the high-frequency segment, i.e., the last 2 to 5 segments, is set to be greater than or equal to the weighted value corresponding to the low-frequency segment, i.e., the first 3 to 5 segments.

2. The method for evaluating wetland ecological status based on bird integrity according to claim 1 is characterized in that: The training process of the hidden Markov model and the target recognition model is: Collecting sounds and images of birds in pre-selected wetland areas at different time periods; The hidden Markov model is trained using the sound, and the target recognition model is trained using the image.

3. The method for evaluating wetland ecological status based on bird integrity according to claim 1 is characterized in that: The target recognition model is a CNN model.

4. The method for evaluating wetland ecological status based on bird integrity according to claim 1 is characterized in that: The bird integrity index is expressed as follows: in, represents the average value of the conservative coefficients of bird species in the pre-selected wetland area, C j represents the conservative coefficient of the j-th bird species; q represents the total number of bird species in the pre-selected wetland area.

5. The method for evaluating wetland ecological status based on bird integrity according to claim 1 is characterized in that: The ecological status levels of wetland areas include poor, relatively poor, medium, good and excellent.

Images