Sheep unit estimation method based on unmanned aerial vehicle data

Abstract

The application provides a livestock sheep unit estimation method based on unmanned aerial vehicle data, comprising the following steps: 1, arranging an unmanned aerial vehicle aerial sample strip based on a system hierarchical sampling method, using an unmanned aerial vehicle to carry a high-resolution digital camera to sample aerial photography on a preset area, and acquiring unmanned aerial vehicle images; 2, based on a deep learning preliminary identification combined with a manual correction method, segmenting and interpreting the spliced unmanned aerial vehicle images, and extracting the number of livestock in the unmanned aerial vehicle sample strip; 3, constructing a hierarchical sampling estimation model, calculating the livestock sheep unit in the preset area according to the livestock sheep unit conversion amount in the sample strip, and realizing scale expansion from the sample strip to the area. The application develops an unmanned aerial vehicle sampling investigation method according to statistical principles in view of the characteristics that livestock has a wide distribution range and a dynamic change in position, deduces a sampling ratio / area, has the advantages of fast data acquisition, small disturbance to animals, no occlusion and unlimited movement, and thus has lower data cost, more reasonable distribution and higher precision.

Description

Technical Field

[0001] This application relates to the field of animal husbandry technology, and more specifically, to a method for estimating livestock sheep units based on unmanned aerial vehicle (UAV) data. Background Technology

[0002] Verification of grazing livestock units is a core task in the management of grassland-livestock balance in natural grasslands. Existing door-to-door survey methods suffer from low efficiency, difficulty in distinguishing between grazing and stall-fed livestock, and inaccurate estimations of grazing feed intake due to the difficulty in obtaining livestock size data. The exact number and distribution of grazing livestock on natural grasslands have long plagued grassland management departments. In recent years, drones, as a novel top-down imaging remote sensing technology, have offered new avenues for livestock surveys due to their low cost, lack of line-of-sight obstruction, and minimal disturbance to animals. However, current interpretation work still relies primarily on visual identification, which is time-consuming and hinders widespread adoption. There is an urgent need to develop high-precision automatic livestock identification methods to support high-frequency, rapid monitoring.

[0003] With the improvement of remote sensing image resolution and the development of computing technology, researchers have developed many animal identification methods, which are mainly divided into traditional target identification methods and deep learning-based identification methods.

[0004] Traditional target recognition algorithms mainly revolve around manually extracted low-level and mid-level features such as spectrum, texture, and shape. While the manually established features and rules in traditional target recognition methods generally have strong interpretability, their overall accuracy is low and their generalization ability is poor, making them only suitable for small-batch image recognition.

[0005] Deep learning recognition models are mainly divided into: 1) Object detection, which identifies the category and location of each object in an image. Object detection algorithms are generally fast, but the generated horizontal anchor boxes have low accuracy in recognizing dense / adhered objects. 2) Semantic segmentation, which performs pixel-level prediction of an image and identifies the category of each pixel. Representative algorithms include UNet and DeepLab. While it does not distinguish between individual objects and produces many false targets, it shows significant advantages in extracting features with varying scales and shapes and in land use classification. Semantic segmentation models are advantageous in recognizing adhered objects, but have low accuracy in recognizing sparse objects. Therefore, it is essential to research and improve collaborative recognition mechanisms using multiple models to construct a high-precision livestock recognition model.

[0006] Meanwhile, the data coverage obtained from UAV transect surveys is usually small. Extending the scale of transect survey data and spatially quantifying it is an effective way to study the distribution patterns of herbivores on a large spatial scale. Many scholars have explored various simulation methods to extend local survey data and achieve spatial gridding. Early studies were mostly based on the assumption of uniform animal distribution, extrapolating the density from transect surveys to the entire region to obtain a large range of wildlife populations. However, due to the strong spatial heterogeneity of animal distribution, the regional density extrapolated from local survey results based on the assumption of uniform distribution has significant uncertainty. Some scholars have adopted stratified sampling and estimation methods to improve accuracy. This involves stratifying the study area, using the mean population density obtained from each stratum as the population density of that stratum, and estimating the population size by multiplying the density by the area of ​​each stratum. Lu Feiying et al. divided the Xinjiang Altun Mountains National Nature Reserve into two layers: the eastern non-calving area and the western calving area. They used the transect method to estimate the population and distribution of three ungulate species: Tibetan antelope, Tibetan wild ass, and wild yak. However, research on how to perform stratification, how to determine the sample size, how to make more scientific stratification estimations, what data to use to verify the estimation results, and how to verify them is still lacking. Summary of the Invention

[0007] This invention provides a method for estimating livestock sheep units based on UAV data, the method comprising the following steps:

[0008] Step 1: Deploy drone aerial photography sample belts based on the system hierarchical sampling method, use drones to sample and photograph the sample belts, acquire drone images, and stitch the images together.

[0009] Step 2: Based on the improved keypoint deep learning network, segment and interpret the stitched UAV imagery to extract the number of livestock within the UAV sample strip;

[0010] Step 3: Construct a stratified sampling estimation model to estimate the number of livestock and sheep units in the preset area based on the converted livestock and sheep units within the transect, thereby achieving scale expansion from the transect to the region.

[0011] Step 4: Evaluate the accuracy of UAV image interpretation based on satellite image interpretation results, ground surveys, or statistical data.

[0012] Furthermore, step 1 also includes:

[0013] Step 11: Divide the sampling survey area into multiple regions based on the distribution density of the animals;

[0014] Step 12: Convert the total area of ​​the study area into the sampled area calculation and obtain the sampling area of ​​the i-th layer;

[0015] Step 13: Calculate the drone's flight altitude;

[0016] Step 14: Determine the overlap of aerial images and stitch together the images acquired by the drone.

[0017] Furthermore, in step 11, the sampling survey area for the stratified sampling of the system is:

[0018]

[0019] Among them, s i Let I represent the sampling area of ​​the i-th layer, and let I represent the number of regions.

[0020] Furthermore, in step 12, the sampling area of ​​the i-th layer is:

[0021]

[0022] Where n is the approximate livestock population sample size at the i-th stratum, calculated using the Cochran formula with a confidence level of 95%, N is the total area of ​​the study area, and s0 is the average pasture area occupied by each livestock population; the livestock population sample size n is:

[0023]

[0024] Among them, z α / 2 Let p be the confidence interval, p be the proportion of the sampled livestock population with the relevant attributes, and E be the error range.

[0025] Furthermore, in step 2, the improved keypoint deep learning network includes a backbone network and a prediction network, wherein the backbone network is used to generate feature maps and the prediction network is used to identify livestock.

[0026] The backbone network includes a convolutional module and an attention mechanism module. The attention mechanism module is located between the two convolutional modules and includes a channel convolutional attention unit and a spatial convolutional attention unit.

[0027] The output of the attention mechanism module is:

[0028]

[0029]

[0030] In the formula, F represents the input feature map, and M... c This represents the channel attention operation, F′ represents the output of the channel attention module, and M represents the channel attention operation. s This represents spatial attention operations. F″ represents the element-wise multiplication operation, and F″ represents the output of the convolutional attention module.

[0031] In step 2, the channel attention map M c (F) is represented as:

[0032]

[0033] in, and These represent the average pooling feature and the max pooling feature, respectively, and σ represents the sigmoid function;

[0034] The spatial attention map M s (F) is represented as:

[0035]

[0036] In the formula, two pooling operations are used to aggregate the channel information of a feature map, generating two 2D maps. The size is 1×H×W. Size is 1×H×W, f 7×7 This represents a convolution operation with a filter size of 7×7;

[0037] The prediction network includes multiple original bounding box prediction branches and rotation angle prediction branches;

[0038] The bounding box of the target input to the keypoint deep learning network is defined as (c x c y δ x δ y , w, h, θ), where, (c x c y ) represents the coordinates of the target center point, (δ) x δ y () represents the center point offset value, and (w, h) represents the width and height of the target;

[0039] The loss function of the keypoint deep learning network is a loss function that adds a rotation factor to the original loss function:

[0040] L = L k +λ size L size +λ off L off +λ angle L angle

[0041]

[0042] In the formula, λ size =0.1, λ off =λ angle =1,Lk L size L off L angle These are the loss functions for the center point, scale, center point offset, and rotation angle, respectively. a k These are the predicted and actual values ​​for the k-th angle, respectively.

[0043] Furthermore, in step 2, the density is determined based on the spacing of the livestock borders:

[0044] If IOU≥0, it indicates that the livestock are stuck together. The semantic segmentation model based on EfficientPS is used to perform semantic segmentation on the dense region. The number of livestock and the centroid of each livestock are estimated based on the average body size of the livestock combined with the fuzzy mean clustering algorithm.

[0045] If IOU < 0, it indicates that the livestock distribution is sparse, and the number of livestock is calculated directly.

[0046] Furthermore, in step 3, the population density at a certain point on the transect and the livestock density in the preset area are estimated based on the number of livestock in the transect using the following methods:

[0047] A point (c) on the sample band x c y The population density of ) is:

[0048]

[0049] In the formula, M(c x c y ) is (c x c y The number of livestock individuals at point A is converted into sheep units; where sheep / goats are converted into 1 sheep unit, cattle / horses / mule into 5 sheep units, yaks into 4 sheep units, dairy cows into 6.5 sheep units, donkeys into 3 sheep units, and A(c x c y ) is (c x c y The territory occupied by the livestock at point )

[0050] A(c x c y The seed region is obtained through the seed region growth method, where the growth rate is proportional to the size of the herd, until it intersects with the growth regions of other seed points in the surrounding area. The growth region of each seed point is then the territory occupied by the herd.

[0051] Furthermore, in step 3, the livestock units within the transect are used to extrapolate all livestock units within the preset area, thus achieving a scale expansion of livestock units from the transect to the region:

[0052]

[0053] Among them, u i s represents the number of sheep units in the i-th layer. i Let s' be the sampling area of ​​the i-th layer. i The total area of ​​the i-th layer; the number of sheep units in the i-th layer is:

[0054] u i =∑ρ(c x c y ), (c x c y )∈W i

[0055] Among them, W i Let represent all points within the i-th layer.

[0056] Furthermore, in step 4, the interpretation bias of the flock of sheep units caused by the movement of livestock in the vertical direction of the heading is:

[0057]

[0058] In the formula, N strip For each strip of the image, after removing duplicate counts, the livestock / sheep units are stitched together. N UAV Bias refers to the livestock sheep units discovered after stitching together all the images. move Interpretation bias for sheep units in the population.

[0059] Furthermore, in step 4, the bias of the UAV imagery survey is assessed based on satellite imagery interpretation results, ground surveys, or statistical data:

[0060]

[0061] In the formula, N UAV N represents the number of livestock found after stitching together all the images. other Bias is the number of livestock obtained based on satellite imagery interpretation, ground surveys, or statistical data. UAV To assess the bias of UAV imagery surveys based on satellite imagery interpretation results, ground surveys, or statistical data.

[0062] The beneficial effects of this application are:

[0063] The present invention proposes a livestock sheep unit estimation method based on UAV data. Taking into account the characteristics of livestock's wide distribution and dynamic location changes, it develops a UAV sampling survey method based on statistical principles, derives the sampling ratio / area, and establishes a UAV transect survey method that has incomparable advantages over traditional ground transect surveys, such as faster data collection, less disturbance to animals, no obstruction, and access to areas that are difficult to reach on the ground. Therefore, the data cost is lower, the distribution is more reasonable, and the accuracy is higher.

[0064] This invention proposes a livestock and sheep unit estimation method based on UAV data, and introduces a livestock recognition network based on rotational keypoints. This transforms the target detection problem into a standard keypoint prediction problem, abandoning the concept of anchors and achieving a better balance between accuracy and efficiency in target detection. To address the issue that bounding boxes generated by the original CenterNet do not fit the target well, a rotation factor with different angles is added to the original framework, and a loss function based on the target rotation angle is added to the loss function, enabling the predicted bounding boxes to more accurately surround the target. Parallel multi-scale path extraction, dilated convolution, a category branch information enhancement module, and channel and spatial convolution attention modules are employed to increase the receptive field during convolution, improving the target mask generation capability and highlighting livestock feature information. This enhances the network's ability to recognize livestock at multiple scales. Furthermore, by combining a semantic segmentation model with traditional target extraction methods, better performance is achieved in recognizing dense livestock. Attached Figure Description

[0065] The advantages of the above and / or additional aspects of this application will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:

[0066] Figure 1 This is a schematic diagram illustrating a method for estimating and evaluating the accuracy of livestock sheep units based on UAV data.

[0067] Figure 2 A schematic diagram of the structure of the improved keypoint deep learning network provided for an example of the present invention;

[0068] Figure 3 A flowchart illustrating the improved keypoint deep learning network provided as an example of the present invention. Detailed Implementation

[0069] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this application can be combined with each other.

[0070] In the following description, many specific details are set forth in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below.

[0071] As attached Figure 1 As shown in the example, this invention proposes a method for estimating livestock sheep units based on UAV data, including the following steps:

[0072] Step 1: Deploy drone aerial photography sample belts based on the system hierarchical sampling method, use drones to sample and photograph the sample belts, acquire drone images, and stitch the images together.

[0073] Step 1 also includes the following steps:

[0074] Step 11: Divide the sampling survey area into multiple regions based on the distribution density of the animals;

[0075] Systematic stratified sampling is a method based on statistical principles that samples livestock based on their distribution density, such as stratified sampling based on grazing zones and grassland yields in different regions. Given that livestock locations are not fixed and the sample size is difficult to determine directly, this invention transforms sample size calculation into sampling area calculation, facilitating the layout of transects. The sampling area for systematic stratified sampling is calculated using the following formula (1):

[0076]

[0077] Among them, s i Let I represent the sampling area of ​​the i-th stratum, calculated according to formula (2). I represents the number of strata in the system, and the stratification type is determined based on the distribution density of animals in different areas of the study area.

[0078] Step 12: Convert the total area of ​​the study area into the sampling survey area calculation, and obtain the sampling area of ​​the i-th layer;

[0079]

[0080] Where n is the number of livestock groups approximated at the i-th layer with a confidence level of 95% calculated using the Cochran formula, N is the total area of ​​the study area, s0 is the average grassland area occupied by each livestock group, and n is calculated according to the following formula (3).

[0081]

[0082] Among them, z α / 2 For example, when the confidence level is 95%, z α / 2=1.96; p is the proportion of the sampled livestock population with relevant attributes. When the value of p cannot be determined, the maximum possible value of (1-p) is used to replace the actual p(1-p). This approximate sample size is generally larger than the actual required sample size. The calculation result of the livestock population sample size is accurate when p is close to 0.5; E is the error margin, such as an allowable error of 5%.

[0083] For example, the sample size of the livestock population is expected to be estimated with a 95% confidence interval and an error margin of no more than 5%. It is calculated according to the following formula (4):

[0084]

[0085] If the population size of the study is small, such as N = 1000, the newly adjusted livestock population sample size m is calculated according to the following formula (5):

[0086]

[0087] That is, if there are 1000 herds of livestock in the survey area, we only need to survey 278 herds, with each herd occupying an average of 1 km². 2 For grasslands, we only need to survey 278km. 2 The survey area decreased significantly. Table 1 shows the sample size required for different population sizes at a 95% confidence level.

[0088] Table 1. Sample size estimation for different population sizes at a 95% confidence level.

[0089]

[0090] As can be seen from the above, assuming a single herd of livestock occupies 1 km² 2 Grassland, with a survey area of ​​100 km² 2 With 100 herds of livestock, and assuming an error margin of less than 5%, approximately 80 herds (80km) need to be sampled. 2 This represents 80% of the total livestock population (grassland area). The larger the livestock population and the larger the grassland area surveyed, the higher the accuracy requirements, the larger the required sampling size, and the higher the cost. Considering both accuracy and cost, this standard stipulates that the survey area should be less than 100 km². 2 When necessary, a comprehensive survey method should be adopted to investigate the entire livestock herd; the survey area should be greater than 100 km². 2 When sampling livestock, it is advisable to use the transect survey method to conduct a sampling survey of the livestock population. In order to ensure the efficiency of the survey and the integrity of the livestock population, the width of each transect should be more than 200m. The sampling area is calculated according to formula (1) to estimate the total number.

[0091] Step 13: Calculate the drone's flight altitude;

[0092] The ground resolution of the image should be 3-5cm.

[0093] Flight altitude is calculated using the following formula (6):

[0094]

[0095] Where H is the flight altitude relative to the average reference plane; GSD is the ground resolution; a is the pixel size; and f is the lens focal length.

[0096] Step 14: Determine the overlap of aerial images and stitch together the images acquired by the drone;

[0097] When collecting flight data, the overlap of flight paths should be 65%–75%, and 80%–85% in steep mountainous areas with large elevation differences. The spacing between flight paths should be more than 200 meters to avoid duplicate or missed counts of livestock due to movement during the survey period. The camera lens should be perpendicular to the ground, and the solar altitude angle should be greater than 20° to ensure that livestock are clearly identifiable in the images. The operation should be completed within 10 days to avoid excessive changes in the number and location of livestock.

[0098] After the drone data collection is completed, the images are stitched together without color mixing to avoid difficulties in subsequent interpretation.

[0099] Step 2: Based on the improved keypoint deep learning network, segment and interpret the stitched UAV imagery to extract the number of livestock within the UAV sample strip;

[0100] In step 2, the improved keypoint deep learning network is a dense livestock deep learning recognition model that integrates object detection and semantic segmentation. The model consists of a livestock recognition model based on rotation keypoints and a semantic segmentation model based on EfficientPS. Specifically, it also includes a backbone network and a prediction network. The backbone network is used to generate feature maps, and the prediction network is used to identify livestock.

[0101] The livestock recognition model based on rotation keypoints is built upon technologies such as rotation keypoints, and the following improvements are made to the keypoint deep learning network (CenterNet):

[0102] The prediction network adds a rotation angle prediction branch to the existing bounding box prediction branches. Both the existing bounding box prediction branches and the rotation angle prediction branches are connected to the backbone network to obtain feature maps. The existing bounding box prediction branches and the rotation angle prediction branches work together in parallel to identify individual livestock.

[0103] To enable the predicted bounding boxes to represent the target more accurately, a rotation factor that generates the predicted boxes at different angles is added to the prediction network branch, and a loss function representing the target rotation angle is added to the loss function to improve accuracy.

[0104] The bounding box of each target in the network is defined as (c x c y δ x δ y , w, h, θ), where, (c x c y ) represents the coordinates of the target center point, (δ) x δ y () represents the center point offset value, and (w, h) represents the width and height of the target.

[0105] The backbone network employs a parallel multi-scale extraction path using a Feature Pyramid Network (FPN) to extract features at multiple scales, thereby improving the ability to identify small targets such as young animals. The backbone network includes convolutional modules and an attention mechanism module, which is positioned between the two convolutional modules and includes channel convolutional attention units and spatial convolutional attention units.

[0106] Convolutional attention modules are added between feature channels at different pooling scales. By concatenating channel convolutional attention units and spatial convolutional attention units, weight allocation is optimized from both channel and spatial dimensions, highlighting livestock feature information and addressing the problem that traditional pyramid pooling fusion does not consider the contribution rate of feature maps at different pooling scales to target recognition at different scales. The convolutional attention modules sequentially infer a 1D channel attention map M. c A Cx1x1 spatial attention map of size M. s The size is 1xHxW. The specific implementation of the residual network is calculated according to the following formula (7).

[0107]

[0108] In the formula, F represents the input feature map, and M... c This represents the channel attention operation, F′ represents the output of the channel attention module, and M represents the channel attention operation. s This represents spatial attention operations. F″ represents the element-wise multiplication operation, and F″ represents the output of the convolutional attention module.

[0109] Channel attention focuses on identifying which input image channels are meaningful. To efficiently compute channel attention, the spatial dimension of the input feature map needs to be compressed. Common methods for aggregating spatial information include average pooling and max pooling. The generated average-pooled and max-pooled features are forwarded to a shared network to produce our channel attention map M. c (F) is calculated using the following formula (8).

[0110]

[0111] in, and These represent the average pooling feature and the max pooling feature, respectively. The shared network consists of a multilayer perceptron (MLP) with one hidden layer. To reduce parameter overhead, the activation size of the hidden layer is set to R / C = r × 1 × 1, where R is the descent rate. After applying the shared network to each descriptor, the output feature vectors are merged using element-wise summation. σ represents the sigmoid function.

[0112] Spatial attention focuses on "where" the most information-rich part, complementing channel attention. The output spatial attention map encodes the locations that need attention or suppression. To compute spatial attention, average pooling and max pooling operations are applied along the channel axis, and then concatenated to generate a valid feature descriptor. A convolutional layer is then applied to generate a spatial attention map M of size R×H×W. s (F) is calculated using the following formula (9).

[0113]

[0114] In the formula, two pooling operations are used to aggregate the channel information of a feature map, generating two 2D maps: The size is 1×H×W. Its size is 1×H×W. 7×7 This represents a convolution operation with a filter size of 7×7.

[0115] The loss function is based on the original CenterNet loss function, with the addition of a rotation factor:

[0116] L = L k +λ size L size +λ off L off +λ angle L angle (10)

[0117]

[0118] In the formula, λ size =0.1, λ off =λ angle =1,L k L size L off L angle These are the loss functions for the center point, scale, center point offset, and rotation angle, respectively. a k These are the predicted and actual values ​​for the k-th angle, respectively.

[0119] After identifying individual livestock using a livestock recognition model based on rotational key points, the density is determined based on the spacing of the livestock borders. If IOU ≥ 0, it indicates that the livestock are clustered together. A semantic segmentation model based on EfficientPS is used to perform semantic segmentation on the dense region. The number of livestock and the centroid of each livestock are estimated based on the average body size of the livestock combined with a fuzzy mean clustering algorithm. If IOU < 0, it indicates that the livestock are sparsely distributed, and the number of livestock is directly calculated.

[0120] Step 3: Construct a stratified sampling estimation model to estimate the number of livestock and sheep units in the preset area based on the converted livestock and sheep units within the transect, thereby achieving scale expansion from the transect to the region.

[0121] In step 3, the population density at a certain point on the transect and the livestock density in the preset area are calculated based on the number of livestock in the transect using the following methods:

[0122] A point (c) on the sample band x c y The population density of ) is calculated using the following formula (12):

[0123]

[0124] In the formula, M(c x c y ) is (c x c y The number of livestock individuals at point A is converted to sheep units (sheep / goat is converted to 1 sheep unit, cattle / horse / mule is converted to 5 sheep units, yaks are converted to 4 sheep units, dairy cows are converted to 6.5 sheep units, and donkeys are converted to 3 sheep units), A(c x c y ) is (c x c y The herd occupies the area at point A(c). x c y The seed region is obtained through the seed region growth method, where the growth rate is proportional to the size of the herd, until it intersects with the growth regions of other seed points in the surrounding area. The growth region of each seed point is then the territory occupied by the herd.

[0125] Using the livestock and sheep units within the transect, the total number of livestock and sheep units within the pre-defined area is extrapolated, thus achieving the scale expansion of livestock and sheep units from the transect to the area. The calculation is performed using the following formula (13):

[0126]

[0127] Among them, s i Let s' be the sampling area of ​​the i-th layer. i Let u be the total area of ​​the i-th layer. i This represents the number of sheep units in the i-th layer, which is the sum of sheep units at all points in the i-th layer.

[0128] u i =∑ρ(c x c y ), (c x c y )∈W i (14)

[0129] Among them, W i Let represent all points within the i-th layer.

[0130] Step 4: Evaluate the accuracy of UAV image interpretation based on satellite image interpretation results, ground surveys, or statistical data;

[0131] This invention also utilizes the interpretation results of UAV imagery by flight strip to calculate the interpretation errors caused by UAV image stitching and livestock movement, and evaluates the accuracy of UAV image interpretation based on satellite image interpretation results, ground surveys or statistical data.

[0132] The accuracy of UAV image interpretation is evaluated using the following methods:

[0133] The interpretation bias of sheep units caused by the directional movement of livestock is negligible. The interpretation bias of sheep units caused by the movement in the vertical direction of the directional movement is calculated according to the following formula (15):

[0134]

[0135] In the formula, N strip For each strip of the image, after removing duplicate counts, the livestock / sheep units are stitched together. N UAV The livestock sheep unit was discovered after stitching together all the images.

[0136] Based on satellite image interpretation results, ground surveys, or statistical data, the bias of UAV image surveys is calculated using the following formula (16):

[0137]

[0138] In the formula, N UAV N represents the number of livestock found after stitching together all the images. other The number of livestock is obtained based on satellite imagery interpretation, ground surveys, or statistical data.

[0139] The following section uses a set of UAV imagery data to test the effectiveness and efficiency of the proposed UAV-based livestock sheep unit estimation and accuracy evaluation method, and compares it with traditional ground-based survey methods. In this example, the UAV data was collected in the Longbao Wetland National Nature Reserve in Yushu County. Longbao National Nature Reserve, established in 1986, is the first national nature reserve in Qinghai Province. It is located in Longbao Town, Yushu County, Yushu Tibetan Autonomous Prefecture, Qinghai Province, approximately 75 kilometers from Jiegu Town, the capital of Yushu Prefecture. The main protected objects of Longbao National Nature Reserve are waterfowl such as the black-necked crane and their habitats. The main vegetation types within the reserve are meadows and freshwater marshes, providing ample food and a good ecological environment for migratory waterfowl. It has become a concentrated breeding ground for black-necked cranes and is one of the highest-altitude nature reserves in the world, with a total area of ​​approximately 107.1 km², including a core area of ​​76.4 km², a buffer zone of 15.8 km², and an experimental zone of 14.9 km². The average temperature is -0.4°C, and the average annual rainfall ranges from 480.5 to 526.1 mm, with rainfall concentrated between June and September. The drone data was acquired by an assembled fixed-wing drone equipped with a standard digital camera, covering four flight paths and capturing a total of 1094 drone images. Specific parameters are shown in Table 2.

[0140] Table 2 Parameters of Electric Fixed-Wing Unmanned Aerial Vehicles

[0141]

[0142]

[0143] Before the flight, the flight strip was laid out using a systematic stratified sampling method. Then, the proposed livestock identification model based on rotation key points was used for identification. Combined with manual interpretation, it was found that there were 2,684 yaks and 490 sheep in the transect. According to formula (13), there were 1,092 sheep and 5,510 yaks in the protected area, totaling 23,131 sheep units. The sheep density in the core area, buffer area and experimental area was 3.56 sheep / km2, 47.88 sheep / km2 and 6.83 sheep / km2, respectively. The cattle density was 52.55 cattle / km2, 33.43 cattle / km2 and 121.43 cattle / km2, respectively. The sheep unit density was 213.75 / km2, 181.61 / km2 and 492.55 / km2, respectively.

[0144] Using Equation (15), it was calculated that the number of yaks in the UAV imagery was underestimated by 1.14% and the number of sheep was overestimated by 1.66% due to livestock movement. Based on Equation (16), we compared the UAV estimation results with the results of the 2011 ground survey and the 2010 high-resolution satellite image interpretation. We found that the UAV image estimation results deviated by 2.85% from the 2011 ground survey results and by 0.17% from the 2010 satellite imagery results. The overall accuracy is high, which can provide technical support for the identification of livestock and sheep units and the management of grass-livestock balance in remote areas.

[0145] The steps in this application can be rearranged, combined, or deleted according to actual needs.

[0146] Although this application has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of this application. The scope of protection of this application is defined by the appended claims and may include various variations, modifications, and equivalents of the invention without departing from the scope and spirit of this application.

Claims

1. A method for estimating livestock sheep units based on UAV data, characterized in that, The method for estimating livestock sheep units based on UAV data includes the following steps: Step 1: Deploy drone aerial photography sample belts based on the system hierarchical sampling method, use drones to sample and photograph the sample belts, acquire drone images, and stitch the images together. Step 2: Based on the improved keypoint deep learning network, segment and interpret the stitched UAV imagery to extract the number of livestock within the UAV sample strip; Step 3: Construct a stratified sampling estimation model to estimate the number of livestock and sheep units in the preset area based on the converted livestock and sheep units within the transect, thereby achieving scale expansion from the transect to the region. Step 4: Evaluate the accuracy of UAV image interpretation based on satellite image interpretation results, ground surveys, or statistical data; Step 1 also includes: Step 11: Divide the sampling survey area into multiple regions based on the distribution density of the animals; Step 12: Convert the total area of ​​the study area into the sampled area calculation and obtain the sampling area of ​​the i-th layer; Step 13: Calculate the drone's flight altitude; Step 14: Determine the overlap of aerial images and stitch together the images acquired by the drone; The sampling area of ​​the i-th layer is: ； in, This is the approximate livestock population sample size for the i-th stratum, calculated using the Cochran formula with a confidence level of 95%. It is the total area of ​​the study region. This represents the average pasture area occupied by each livestock herd; the sample size n of the livestock herd is: ； in, Let be the confidence interval. E represents the proportion of the sampled livestock population with relevant attributes, and E represents the error margin. In step 2, the improved keypoint deep learning network includes a backbone network and a prediction network, wherein the backbone network is used to generate feature maps and the prediction network is used to identify livestock. The backbone network includes a convolutional module and an attention mechanism module. The attention mechanism module is located between the two convolutional modules and includes a channel convolutional attention unit and a spatial convolutional attention unit. The output of the attention mechanism module is: ； In the formula, Indicates the input feature map, This indicates channel attention operations. This indicates the output of the channel attention module. This represents spatial attention operations. This represents a convolution operation that multiplies element by element. This represents the output of the convolutional attention module; In step 2, the channel attention map Represented as: ； in, and These represent the average pooling feature and the max pooling feature, respectively. Represents the sigmoid function; The spatial attention map Represented as: ； In the formula, two pooling operations are used to aggregate the channel information of a feature map, generating two 2D maps. The size is 1×H×W. The size is 1×H×W. This represents a convolution operation with a filter size of 7×7; The prediction network includes multiple original bounding box prediction branches and rotation angle prediction branches; The bounding box of the target input to the keypoint deep learning network is defined as... ,in, Represents the coordinates of the target's center point. This represents the center point offset value. Indicates the width and height of the target.

2. The livestock sheep unit estimation method based on UAV data according to claim 1, characterized in that, In step 11, the sampling area for the stratified sampling of the system is: ； in, Let I represent the sampling area of ​​the i-th layer, and let I represent the number of regions.

3. The livestock sheep unit estimation method based on UAV data according to claim 1, characterized in that, In step 2, The loss function of the keypoint deep learning network is a loss function that adds a rotation factor to the original loss function: ； ； In the formula, , , , , , These are the loss functions for the center point, scale, center point offset, and rotation angle, respectively. , These are the predicted and actual values ​​for the k-th angle, respectively.

4. The livestock sheep unit estimation method based on UAV data according to claim 1, characterized in that, In step 3, the population density at a certain point on the transect and the livestock density in the preset area are calculated based on the number of livestock in the transect using the following methods: Sample band at a certain point The population density is: ； In the formula, for The number of livestock individuals at each point is converted into sheep units; sheep / goats are converted into 1 sheep unit, cattle / horses / mule into 5 sheep units, yaks into 4 sheep units, dairy cows into 6.5 sheep units, and donkeys into 3 sheep units. for The territory occupied by the livestock herd at a point; The seed region growth method is used to obtain the growth rate, which is proportional to the size of the herd. When the growth regions of each seed point intersect with the growth regions of other seed points in the surrounding area, the growth region of each seed point is the territory occupied by the herd.

5. The livestock sheep unit estimation method based on UAV data according to claim 1, characterized in that, In step 3, the livestock units within the transect are used to extrapolate all livestock units within the preset area, thus achieving a scale expansion of livestock units from the transect to the region: ； in, Indicates the first Number of sheep units per layer For the first Sampling area of ​​the layer For the first Total area of ​​the layer; the first The number of sheep units in each layer is: ； in, For the first All points within the layer.

6. The livestock sheep unit estimation method based on UAV data according to claim 1, characterized in that, In step 4, the interpretation bias of the flock sheep units caused by the movement of livestock in the vertical direction of the heading is: ； In the formula, The livestock sheep units were removed by stitching the images strip by strip and removing duplicate counts. The livestock sheep unit was discovered after stitching together all the images. Interpretation bias for sheep units in the population.

7. The livestock sheep unit estimation method based on UAV data according to claim 1, characterized in that, In step 4, the bias of the UAV imagery survey is assessed based on satellite imagery interpretation results, ground surveys, or statistical data: ； In the formula, The number of livestock found after stitching together all the images. The number of livestock is based on satellite imagery interpretation, ground surveys, or statistical data. To assess the bias of UAV imagery surveys based on satellite imagery interpretation results, ground surveys, or statistical data.

Images