Wild animal population automatic counting method based on unmanned aerial vehicle multispectral imaging

Abstract

The invention discloses a wildlife population automatic counting method based on unmanned aerial vehicle multispectral imaging, and relates to the technical field of wildlife monitoring and ecological protection, and the method comprises the following steps: 1.1, imaging area planning: constructing a three-dimensional terrain model based on habitat distribution data, activity rules and terrain features of target wildlife, imaging sub-regions are divided by adopting a rasterization zoning strategy, unmanned aerial vehicle flight parameters of each sub-region are determined, the flight parameters comprise flight height, course overlapping degree, lateral overlapping degree and navigational speed, the course overlapping degree is set to be 70%-85%, the lateral overlapping degree is set to be 60%-75%, the flight height is dynamically adjusted according to the body size of a target wild animal, and the navigational speed is adjusted according to the body size of the target wild animal. And the corresponding flight height of the wild animals with the body types larger than 1m is 80-120m. According to the method, efficient and accurate counting of wild animal populations is realized through scientific imaging area planning, accurate data preprocessing, multi-dimensional feature extraction and an optimized target recognition model, and reliable technical support is provided for ecological protection work.

Description

Technical Field

[0001] This invention relates to the field of wildlife monitoring and ecological protection technology, and in particular to an automatic wildlife population counting method based on UAV multispectral imaging. Background Technology

[0002] Wildlife population size is a core indicator for assessing biodiversity and formulating ecological conservation strategies. Accurate population count data plays an irreplaceable role in endangered species rescue, ecological balance regulation, and nature reserve management. Traditional wildlife population counting methods mainly rely on manual ground patrols, infrared camera deployment, and conventional aerial remote sensing monitoring. However, these methods generally have limitations and cannot meet the demands of modern ecological monitoring for high efficiency and precision.

[0003] Ground-based manual patrol methods are significantly constrained by terrain conditions, making comprehensive monitoring difficult in complex areas such as high mountains, dense forests, and swamps. Furthermore, human intervention easily disturbs wild animals, altering their movement patterns and distorting the counting results. Additionally, this method is labor-intensive and time-consuming, making it challenging to conduct comprehensive monitoring over a 10km area. 2 Data collection and analysis in monitored areas often takes several days, resulting in extremely low efficiency. While infrared camera monitoring can achieve unattended operation, its monitoring range is limited (the coverage radius of a single device is usually no more than 50m), it is easily affected by vegetation obstruction and weather changes, and it requires manual screening of massive amounts of image data, leading to serious time lag.

[0004] Conventional aerial remote sensing monitoring mostly uses visible light imaging technology. Although it can cover a large area, in environments with high vegetation cover, the grayscale difference between wild animals and the background is not obvious, making target identification difficult and resulting in high rates of missed detections and false detections. Multispectral imaging technology can obtain the spectral response characteristics of targets in different bands, providing more dimensions of information for distinguishing wild animals from the background. Meanwhile, the development of UAV technology has solved the problems of poor mobility and high cost of traditional aerial remote sensing.

[0005] While some monitoring attempts have been made based on UAV multispectral imaging, several technical bottlenecks remain: First, imaging planning lacks specificity, and flight parameter settings do not incorporate wildlife size and habitat characteristics, resulting in insufficient image resolution or incomplete coverage. Second, simplified data preprocessing leads to low accuracy in radiometric and geometric correction, affecting subsequent feature extraction. Third, target recognition models have weak generalization capabilities and have not been optimized to address sample imbalance in wildlife monitoring, making them ill-suited for complex field environments. Therefore, developing a streamlined and reliable automatic wildlife population counting method based on UAV multispectral imaging has become an urgent need in the field of ecological monitoring. Summary of the Invention

[0006] The technical problem solved by this invention is that existing methods for counting wild animal populations are inefficient, inaccurate, and have poor adaptability. This invention provides an automatic counting method based on UAV multispectral imaging. Through scientific imaging area planning, precise data preprocessing, multi-dimensional feature extraction, and optimized target recognition model, it achieves efficient and accurate counting of wild animal populations, providing reliable technical support for ecological protection work.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an automatic wildlife population counting method based on UAV multispectral imaging, comprising the following steps:

[0008] 1.1 Imaging Area Planning: Based on the habitat distribution data, activity patterns, and terrain features of the target wild animals, a three-dimensional terrain model is constructed. A rasterization partitioning strategy is adopted to divide the imaging sub-regions, and the UAV flight parameters for each sub-region are determined. The flight parameters include flight altitude, forward overlap, lateral overlap, and flight speed. The forward overlap is set to 70%-85%, the lateral overlap is set to 60%-75%, and the flight altitude is dynamically adjusted according to the size of the target wild animals. Wild animals larger than 1m correspond to a flight altitude of 80-120m, wild animals with a size of 0.3-1m correspond to a flight altitude of 40-80m, and wild animals with a size smaller than 0.3m correspond to a flight altitude of 20-40m.

[0009] 1.2 Multispectral Data Acquisition: The UAV equipped with a multispectral imaging device performs flight missions according to the planned flight parameters, and simultaneously acquires multispectral image data and POS positioning data of each sub-region. The multispectral imaging device contains at least 4 band channels, namely blue band (450-500nm), green band (530-590nm), red band (630-690nm) and near-infrared band (770-890nm). During the acquisition process, the imaging attitude is corrected in real time through the IMU inertial measurement unit to ensure that the image acquisition tilt angle is less than 3°.

[0010] 1.3 Data Preprocessing: Radiometric correction, geometric correction, and image stitching are performed on the acquired multispectral image data. Radiometric correction adopts the relative radiometric correction method based on the calibration plate to eliminate the influence of uneven illumination. Geometric correction combines POS positioning data and three-dimensional terrain model to achieve accurate matching between image pixels and geographic coordinates. Image stitching adopts SIFT feature point matching algorithm combined with RANSAC robustness test to remove mismatched points and generate a complete multispectral mosaic map of the target area.

[0011] 1.4 Region of Interest Extraction: A vegetation-animal differentiation model was constructed based on multispectral features. Non-animal regions were initially removed by calculating the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI). Then, the Otsu adaptive threshold segmentation algorithm was used to binarize the remaining regions to extract candidate animal regions of interest.

[0012] 1.5 Feature Extraction and Screening: Multi-dimensional features were extracted from the regions of interest of candidate animals, including spectral features, shape features, and texture features. Spectral features included the mean gray value of each band, band ratio, and spectral angle matching degree. Shape features included roundness, rectangularity, and aspect ratio. Texture features were calculated using the GLCM matrix to calculate energy, entropy, and contrast. The features were ranked by importance using the random forest algorithm, and the top 30% of key features with the highest contribution rate were screened out.

[0013] 1.6 Target Detection and Recognition: The selected key features are input into the improved YOLOv5 network model for target detection. The improved YOLOv5 network model enhances the target feature extraction capability by introducing the attention mechanism module CBAM, and at the same time uses the FocalLoss loss function to solve the sample imbalance problem, so as to achieve accurate identification and localization of target wild animals.

[0014] 1.7 Population Counting and Verification: Identified target wild animals are counted to remove duplicates. Duplicate targets are excluded by combining UAV flight trajectories and geographic coordinate information. The count is verified by comparing the results with those from manual ground counting. The counting accuracy is calculated. If the accuracy is below 90%, the process returns to step 1.3 to readjust the preprocessing parameters and then repeats the subsequent steps.

[0015] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, the method for constructing the three-dimensional terrain model in step 1.1 is as follows: combining satellite remote sensing DEM data and UAV low-altitude photogrammetry data, the terrain data is completed using the Kriging interpolation algorithm to generate a three-dimensional terrain model with a resolution of 0.5m. The habitat distribution data is obtained through an ecological environment monitoring database, and the activity pattern data is obtained through comprehensive analysis of historical tracking records and infrared camera monitoring data.

[0016] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, the acquisition parameters of the multispectral imaging device in step 1.2 are set as follows: bandwidth ≤ 10nm, image resolution ≥ 5 million pixels, frame rate 1-3fps. During the acquisition process, the exposure parameters are automatically adjusted through the UAV flight control system to ensure the brightness consistency of images in different sub-regions. The sampling frequency of POS positioning data is 10Hz, and the positioning accuracy is ≤ 0.5m.

[0017] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, the specific steps of radiometric correction in step 1.3 are as follows:

[0018] 4.1 Acquire multispectral images of the standard calibration plate under the same illumination conditions to obtain standard reflectance data for each band;

[0019] 4.2 Calculate the ratio of the gray value of each band of the image to be corrected to the gray value of the corresponding band of the calibration plate to obtain the radiometric correction coefficient;

[0020] 4.3 The image to be calibrated is corrected pixel by pixel using the radiometric correction coefficient. The formula is: R = R × (R / G), where R is the reflectance after correction, R is the gray value of the original image, R is the standard reflectance of the calibration plate, and G is the gray value of the calibration plate image.

[0021] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, wherein the formula for calculating the Normalized Difference Vegetation Index (NDVI) in step 1.4 is:

[0022] NDVI = (NIR - Red) / (NIR + Red)

[0023] Where NIR is the gray value of the near-infrared band and Red is the gray value of the red band. The NDVI threshold is set to 0.3, and vegetation areas with NDVI > 0.3 are removed.

[0024] The formula for calculating the Normalized Difference Water Index (NDWI) is as follows:

[0025] NDWI=(Green-NIR) / (Green+NIR)

[0026] Where Green represents the grayscale value of the green band, and the NDWI threshold is set to 0.2 to remove water areas with NDWI > 0.2.

[0027] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, the feature selection process of the random forest algorithm in step 1.5 is as follows:

[0028] 6.1 Construct a random forest model containing 100 decision trees, using the extracted multi-dimensional features as input variables and manually labeled "animal / non-animal" as output variables;

[0029] 6.2 Calculate the reduction in the Gini coefficient of each feature across all decision trees, as an indicator of feature importance;

[0030] 6.3 Sort the features from highest to lowest importance, select the top 30% of features to form a key feature set, and remove redundant features to reduce the computational complexity of the model.

[0031] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, the specific improvements to the improved YOLOv5 network model in step 1.6 include:

[0032] 7.1 Insert the CBAM attention mechanism module after the C3 module of the backbone network. First, calculate the weight coefficients of each feature channel through the channel attention submodule, and then focus on the spatial features of the target region through the spatial attention submodule to enhance the distinction between the target and the background.

[0033] 7.2 The FocalLoss loss function is used to replace the original Cross-EntropyLoss. By introducing the modulation coefficient (1-p), where p is the model's prediction probability of the sample and γ is the focusing parameter, which is set to 2, the weight of easily classified background samples is reduced and the training weight of difficult-to-classify animal target samples is increased.

[0034] 7.3 The anchor box size of the network output layer is adaptively adjusted. Based on the size statistics of the target wild animals, the K-means clustering algorithm is used to redetermine the anchor box parameters to improve the matching accuracy of the target box.

[0035] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, the specific method for deduplication counting in step 1.7 is as follows:

[0036] 8.1 Extract the geographic coordinates and bounding box parameters of each identified target, and determine whether targets in adjacent images belong to the same region based on the UAV flight trajectory;

[0037] 8.2 The IOU (Intersection over Union) algorithm is used to calculate the overlap of target bounding boxes in adjacent images. If IOU > 0.5 and geographic coordinate distance < 1m, they are determined to be the same target and are deduplicated.

[0038] 8.3 Count the number of targets after deduplication, classify and count them according to wild animal species, and generate a population count report.

[0039] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, the automatic counting method further includes a model optimization and update step: periodically collecting new multispectral image data and manually labeled samples, incrementally training the improved YOLOv5 network model, updating the model parameters, and improving its adaptability to different environmental conditions and wildlife species.

[0040] As a preferred embodiment of the automatic wildlife population counting method based on UAV multispectral imaging described in this invention, step 1.1 further includes flight path obstacle avoidance planning: combining real-time meteorological data and obstacle monitoring information, an AI algorithm is used to plan the optimal flight path to ensure the flight safety of the UAV. The obstacle monitoring information is acquired in real time by the lidar sensor carried by the UAV.

[0041] The beneficial effects of this invention are:

[0042] 1. Significantly Improved Counting Efficiency: Rapid data acquisition over large areas is achieved through UAV multispectral imaging, combined with automated data processing and target identification workflows, significantly shortening the population counting cycle. (For a 20km radius...) 2 In the monitoring area, this invention only needs 1-2 hours to complete data collection and 4-6 hours to complete automatic counting. Compared with the traditional manual counting method (which takes several days), the efficiency is more than 10 times higher, and it is especially suitable for routine monitoring of large-area wildlife habitats.

[0043] 2. Significantly Improved Counting Accuracy: Through multi-dimensional feature extraction and an improved YOLOv5 model, the problem of low target recognition accuracy in traditional methods is effectively solved. The CBAM attention mechanism enhances the target feature extraction capability, the Focal Loss loss function solves the sample imbalance problem, and the adaptive anchor box improves the localization accuracy, enabling the target recognition accuracy to reach over 92% and the population counting accuracy to remain stable at over 90%. This represents a qualitative leap compared to traditional visible light remote sensing monitoring methods (accuracy of approximately 70%).

[0044] 3. Strong environmental adaptability: By dynamically adjusting flight parameters (based on the size of wild animals), incrementally training models (based on new sample data), and scientific regional planning (based on terrain and habitat characteristics), this invention can adapt to wild animals of different sizes (from small birds to large mammals), different terrain environments (forests, grasslands, mountains), and different seasonal conditions, thus solving the bottleneck problem of weak adaptability of traditional methods.

[0045] 4. Minimal disturbance to wildlife: Drone flight monitoring avoids direct interference with wildlife caused by manual ground patrols, reduces stress responses in wildlife, and ensures that the counting results are closer to their natural population size; at the same time, multispectral imaging technology can penetrate some vegetation cover, reducing the impact of vegetation cover on target identification and expanding the monitoring range.

[0046] 5. High comprehensive value of data: The collected multispectral image data can not only be used for population counting, but also for further analysis of wildlife habitat utilization (based on differences in spectral characteristics), health status (based on changes in spectral response), and other information, providing multi-dimensional data support for wildlife conservation work and enhancing the comprehensive utilization value of monitoring data. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the basic process of an automatic wildlife population counting method based on UAV multispectral imaging, provided as an embodiment of the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0049] Example 1: Population counting of red deer

[0050] 1. Overview of the test area

[0051] The experimental area is located in a national nature reserve, with a core monitoring area of ​​25 km. 2 It is a transitional zone between mixed coniferous and broad-leaved forests and mountain grasslands, with an altitude range of 900-1500m and a vegetation coverage of about 65%. It includes a small number of streams (water area of ​​about 3%). Red deer mainly live in areas with an altitude of 1100-1300m, with peak activity in the early morning and evening.

[0052] 2. Imaging region planning

[0053] Distribution data of red deer habitat was obtained from the ecological environment monitoring database, and combined with infrared camera monitoring data from the past year (a total of 50 infrared cameras were deployed, acquiring 320 valid video clips), to determine the core monitoring area. Satellite remote sensing DEM data at a resolution of 30m and low-altitude photogrammetry data from UAVs were collected for this area, and a 0.5m resolution 3D terrain model was generated using the Kriging interpolation algorithm.

[0054] The core monitoring area was divided into 25 imaging sub-regions of 1km × 1km. Given the size of the red deer (1.5-2.0m), the flight altitude was set at 100m, with 80% forward overlap, 70% lateral overlap, and a flight speed of 5m / s. Combining real-time meteorological data (wind speed 2.5m / s, visibility 10km) and obstacle information from lidar monitoring (mainly tall larch trees, ≤25m in height), the A* algorithm was used to plan the flight path, avoiding areas with concentrated obstacles.

[0055] 3. Multispectral data acquisition

[0056] A multispectral imaging device from a certain brand was selected, containing four channels: blue (470nm), green (550nm), red (670nm), and near-infrared (850nm), with a bandwidth of 8nm, an image resolution of 12 megapixels, and a frame rate of 2fps. It was mounted on a DJI Matrice 300RTK drone, with the flight control system automatically adjusting the exposure parameters to a shutter speed of 1 / 800s and ISO 200.

[0057] The flight mission was conducted during the peak activity period of red deer, from 6:00 AM to 8:00 AM. The IMU (Inertial Measurement Unit) corrected the attitude in real time to ensure that the imaging tilt angle was ≤2°. The POS (Positioning System) recorded coordinate data at a frequency of 10Hz, with a positioning accuracy of 0.3m. A total of 980 multispectral images were acquired, and POS and attitude data were recorded simultaneously, all stored in RAW format.

[0058] 4. Data Preprocessing

[0059] Radiometric correction uses a standard grayscale calibration board (with reflectivities of 10%, 30%, 50%, 70%, and 90%). Images of the calibration board are acquired under the same illumination conditions, and correction coefficients for each band are calculated. Taking the red band as an example, the grayscale value corresponding to a standard reflectivity of 50% on the calibration board is 128, and the average grayscale value of the red band image to be corrected is 112. The correction coefficient is 50% / 128×112≈43.75%, and the red band image is corrected pixel by pixel based on this.

[0060] Geometric correction, combining POS data with a 3D terrain model, employs bilinear interpolation to match pixels with geographic coordinates, achieving a geolocation error of ≤0.4m for the corrected image. Image stitching utilizes the SIFT algorithm to extract feature points, extracting approximately 2200 feature points per pair of adjacent images. RANSAC is then used to remove approximately 90 mismatched points, achieving a matching accuracy of 96%. The stitched image generates a complete 25km sequence. 2 Multispectral mosaic.

[0061] 5. Region of Interest Extraction

[0062] NDVI values ​​were calculated, and vegetation areas (62%) with NDVI > 0.3 were removed. NDWI values ​​were calculated, and water areas (3%) with NDWI > 0.2 were removed. The remaining 35% of the area was binarized using the Otsu algorithm, and the optimal segmentation threshold was calculated to be 125, resulting in 218 candidate regions of interest for animals.

[0063] 6. Feature Extraction and Filtering

[0064] Multidimensional features were extracted from 218 candidate regions. Spectral features included a mean gray value of 105 in the blue band, 118 in the green band, 102 in the red band, and 135 in the near-infrared band, with a blue-green ratio of 0.889, a red-green ratio of 0.864, a red-near-infrared ratio of 0.756, and a spectral angle matching degree of 0.93. Shape features included a circularity of 0.71, a rectangularity of 0.84, and an aspect ratio of 1.4. Texture features included an energy of 0.36, an entropy of 1.85, and a contrast of 88.

[0065] The random forest algorithm was used to select features. The top 30% of the key features in terms of importance were: mean gray value in the near-infrared band, spectral angle matching degree, circularity, entropy, and red-near-infrared ratio, which formed the key feature set.

[0066] 7. Target Detection and Recognition

[0067] Key features were input into an improved YOLOv5 model. The model's CBAM module enhanced the near-infrared band feature weights, and the Focal Loss function focused on training on deer target samples. Adaptive anchor boxes were determined by K-means clustering to [150,180], [160,190], and [170,200]. The model output identified 165 valid targets, including 158 deer targets and 7 misidentified targets (rocks, dead trees), achieving an accuracy of 95.8%.

[0068] 8. Population Counting and Verification

[0069] After deduplication of 158 red deer targets, based on geographic coordinates and the IOU algorithm, 23 targets were found to be duplicates detected in adjacent images, leaving 135 targets after deduplication. The ground-based manual counting team used the transect method (deploying 5 transects 10km long) combined with infrared camera data to obtain an actual red deer population of 142. The counting accuracy of this invention is 135 / 142×100%≈95.1%, meeting the requirement of ≥90%, and the output population count result is 135.

[0070] 9. Model Optimization

[0071] We collected 158 red deer target samples and 7 misidentified samples from this monitoring, merged them with historical samples (600 samples), and incrementally trained the improved YOLOv5 model. After training, the model's recognition accuracy improved to 96.2%.

[0072] Example 2: Wild rabbit population counting

[0073] This embodiment uses population counting of wild rabbits (approximately 0.4-0.6m in size) in a grassland reserve as an example to verify the applicability of the present invention to small wild animals. The core difference from Embodiment 1 is:

[0074] 1. Imaging area planning: The flight altitude is set to 60m, the forward overlap is 85%, the lateral overlap is 75%, the flight speed is 3m / s, and the three-dimensional terrain model is generated based on grassland terrain features.

[0075] 2. Multispectral data acquisition: Imaging equipment frame rate 3fps, exposure parameters shutter speed 1 / 1000s, ISO 100, POS positioning accuracy 0.4m.

[0076] 3. Feature selection: Key features are mean gray value of green band, spectral angle matching degree, rectangularity, contrast, and blue-red ratio.

[0077] 4. Target identification: The adaptive anchor boxes were determined by K-means clustering as [45,55], [50,60], and [55,65].

[0078] The final population counting accuracy rate was 92.3%, which meets the requirements of practical applications, indicating that the invention has good applicability to wild animals of different body sizes.

[0079] In this invention, rapid data acquisition over a large area is achieved through multispectral imaging by unmanned aerial vehicles (UAVs). Combined with automated data processing and target identification procedures, the population counting cycle is significantly shortened. (For a range of 20km...) 2In the monitored area, this invention only requires 1-2 hours to complete data acquisition and 4-6 hours to complete automatic counting, which is more than 10 times more efficient than the traditional manual counting method (which takes several days). It is especially suitable for routine monitoring of large-area wildlife habitats. Through multi-dimensional feature extraction and an improved YOLOv5 model, it effectively solves the problem of low target recognition accuracy in traditional methods. The CBAM attention mechanism enhances the target feature extraction capability, the Focal Loss loss function solves the sample imbalance problem, and the adaptive anchor box improves the positioning accuracy, making the target recognition accuracy reach more than 92% and the population counting accuracy stable at more than 90%. Compared with the traditional visible light remote sensing monitoring method (accuracy of about 70%), this is a qualitative leap. By dynamically adjusting flight parameters (based on wildlife size), incrementally training the model (based on new sample data), and scientific regional planning (based on terrain and habitat characteristics), this invention can adapt to wildlife of different sizes (from small birds to large mammals), different terrain environments (forests, grasslands, mountains), and different seasonal conditions, solving the problem of the limitations of traditional methods. Overcoming the bottleneck of weak responsiveness, drone-based aerial monitoring avoids direct interference with wildlife by manual ground patrols, reduces stress responses in wild animals, and ensures that the counting results are closer to their natural population sizes. Simultaneously, multispectral imaging technology can penetrate some vegetation cover, reducing the impact of vegetation cover on target identification and expanding the monitoring range. The collected multispectral image data can not only be used for population counting but also for further analysis of wildlife habitat utilization (based on spectral feature differences) and health status (based on spectral response changes), providing multi-dimensional data support for wildlife conservation efforts and enhancing the comprehensive utilization value of monitoring data.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for automatic counting of wildlife population based on unmanned aerial vehicle multispectral imaging, characterized in that, The method comprises the following steps: 1.1 Imaging area planning: based on habitat distribution data, activity rules and terrain characteristics of target wild animals, a three-dimensional terrain model is constructed, a grid partition strategy is adopted to divide imaging sub-regions, and flight parameters of each sub-region are determined, including flight height, heading overlap, lateral overlap and airspeed, wherein the heading overlap is set to 70%-85%, the lateral overlap is set to 60%-75%, and the flight height is dynamically adjusted according to the size of the target wild animal, the flight height corresponding to the wild animal with a size greater than 1m is 80-120m, the flight height corresponding to the wild animal with a size of 0.3-1m is 40-80m, and the flight height corresponding to the wild animal with a size less than 0.3m is 20-40m; 1.2 Multispectral data acquisition: the unmanned aerial vehicle carrying the multispectral imaging equipment executes the flight task according to the planned flight parameters, and synchronously acquires multispectral image data and POS positioning data of each sub-region, wherein the multispectral imaging equipment at least includes four wave band channels, namely blue wave band (450-500nm), green wave band (530-590nm), red wave band (630-690nm) and near-infrared wave band (770-890nm), and the imaging posture is corrected in real time by an IMU inertial measurement unit during the acquisition process to ensure that the image acquisition inclination is less than 3°; 1.3 Data preprocessing: the acquired multispectral image data is subjected to radiation correction, geometric correction and image stitching processing, the radiation correction adopts a relative radiation correction method based on a calibration plate to eliminate the influence of uneven illumination, the geometric correction realizes accurate matching of image pixels and geographic coordinates in combination with POS positioning data and a three-dimensional terrain model, and the image stitching adopts a SIFT feature point matching algorithm combined with RANSAC robustness test to remove mis-matching points, thereby generating a complete multispectral mosaic image of the target region; 1.4 Extraction of regions of interest: a vegetation-animal distinguishing model is constructed based on multispectral characteristics, preliminary elimination of non-animal regions is realized by calculating normalized vegetation index (NDVI) and normalized difference water index (NDWI), and binary processing of the remaining regions is carried out by using an Otsu adaptive threshold segmentation algorithm, thereby extracting candidate animal regions of interest; 1.5 Feature extraction and screening: multi-dimensional features are extracted from the candidate animal regions of interest, including spectral features, shape features and texture features, the spectral features include gray mean values of each wave band, wave band ratios and spectral angle matching degrees, the shape features include circularity, rectangularity and Aspect ratio, the texture features adopt GLCM matrix to calculate energy, entropy and contrast, the features are sorted in importance by using a random forest algorithm, and key features with a contribution rate of the top 30% are screened out; 1.6 Target detection and identification: the screened key features are input into an improved YOLOv5 network model for target detection, the improved YOLOv5 network model enhances the target feature extraction capability by introducing a CBAM attention mechanism module, and solves the sample imbalance problem by using a FocalLoss loss function, thereby realizing accurate identification and positioning of target wild animals. 1.7 Population counting and verification: De-duplicate the identified target wildlife, exclude repeated detection targets combined with the UAV flight trajectory and geographic coordinate information, and verify by comparing with the ground manual counting results to calculate the counting accuracy. If the accuracy is less than 90%, return to step 1.3 to adjust the preprocessing parameters and execute the subsequent steps again.

2. The method of claim 1, wherein the method comprises: The construction method of the three-dimensional terrain model in step 1.1 is: combining satellite remote sensing DEM data and unmanned aerial photogrammetry data, using Kriging interpolation algorithm to complete the terrain data, generating a three-dimensional terrain model with a resolution of 0.5m, the habitat distribution data is obtained from the ecological environment monitoring database, and the activity rule data is obtained based on historical tracking records and infrared camera monitoring data comprehensive analysis.

3. The method of claim 2, wherein the method further comprises: The acquisition parameter setting of the multi-spectral imaging device in step 1.2 is: wave band width ≤10nm, image resolution ≥5 million pixels, frame rate 1-3fps, automatic adjustment of exposure parameters is realized through the UAV flight control system during the acquisition process to ensure the consistency of the brightness of different sub-region images, the sampling frequency of POS positioning data is 10Hz, and the positioning accuracy is ≤0.5m.

4. The method of claim 3, wherein the method further comprises: The specific steps of the radiation correction in step 1.3 are: 4.1 Collect the multi-spectral images of the standard calibration board under the same lighting conditions to obtain the standard reflectivity data of each wave band; 4.2 Calculate the ratio of the gray value of each wave band of the image to be corrected to the corresponding wave band gray value of the calibration board to obtain the radiation correction coefficient; 4.3 Use the radiation correction coefficient to correct the image to be corrected pixel by pixel, the formula is: R=R×(R / G), where R is the corrected reflectivity, R is the original image gray value, R is the standard reflectivity of the calibration board, and G is the image gray value of the calibration board.

5. The method of claim 4, wherein: The calculation formula of the normalized vegetation index (NDVI) in step 1.4 is: NDVI=(NIR-Red) / (NIR+Red) Where NIR is the near-infrared wave band gray value, Red is the red wave band gray value, and the NDVI threshold is set to 0.3, and the vegetation area with NDVI>0.3 is removed; The calculation formula of the normalized difference water index (NDWI) is: NDWI=(Green-NIR) / (Green+NIR) Where Green is the green wave band gray value, and the NDWI threshold is set to 0.2, and the water area with NDWI>0.2 is removed.

6. The method of claim 5, wherein: The feature selection process of the random forest algorithm in step 1.5 is: 6.1 Build a random forest model containing 100 decision trees, and use the extracted multi-dimensional features as input variables and the "animal / non-animal" label annotated by humans as output variables; 6.2 Calculate the Gini coefficient reduction of each feature in all decision trees as a feature importance evaluation index; 6.3 Sort the features from high to low according to the feature importance, select the top 30% of the features to form a key feature set, and remove redundant features to reduce the model calculation complexity.

7. The method of claim 6, wherein the method further comprises: The specific improvements of the improved YOLOv5 network model in step 1.6 include: 7.1 Insert CBAM attention mechanism module after C3 module of backbone network, first calculate weight coefficient of each feature channel through channel attention submodule, then focus on spatial features of target region through spatial attention submodule, enhance the distinguishability of target and background; 7.2 Use FocalLoss loss function instead of original Cross-EntropyLoss, by introducing modulation coefficient(1-p), where p is the prediction probability of the model for the sample, and γ is the focus parameter, set to 2, reduce the weight of easy-to-classify background samples, and improve the training weight of difficult-to-classify animal target samples; 7.3 Adaptive adjustment of anchor box size of network output layer, based on the size statistics of target wild animals, use K-means clustering algorithm to re-determine the anchor box parameters, improve the matching accuracy of target box.

8. The method of claim 7, wherein: The specific method of de-duplication counting in step 1.7 is: 8.1 Extract the geographic coordinate information and boundary box parameters of each identified target, and judge whether the target in the adjacent image belongs to the same area based on the flight trajectory of the unmanned aerial vehicle; 8.2 Calculate the overlap of the target boundary box in the adjacent image using the IOU(intersection over union) algorithm, if IOU>0.5 and the distance between geographic coordinates is<1m, it is determined as the same target, and it is de-duplicated; 8.3 Count the number of targets after de-duplication, classify and count by wild animal species, and generate a population count report.

9. The method of claim 8, wherein: The automatic counting method also includes a model optimization update step: regularly collect new multispectral image data and manually annotated samples, incrementally train the improved YOLOv5 network model, update the model parameters, and improve the adaptability to different environmental conditions and wild animal species.

10. The method of claim 9, wherein: The step 1.1 also includes flight path obstacle avoidance planning: combine real-time weather data and obstacle monitoring information, use AI algorithm to plan the optimal flight path, ensure the safety of unmanned aerial vehicle flight, and the obstacle monitoring information is obtained in real time by the laser radar sensor carried by the unmanned aerial vehicle.

Images