Livestock behavior identification method based on Beidou satellite positioning and multi-parameter expansion

Abstract

The invention discloses a livestock behavior identification method based on Beidou satellite positioning and multi-parameter expansion. The method comprises the following steps: acquiring satellite positioning data and synchronous behavior tags of livestock; calculating multi-dimensional motion parameters of instantaneous speed, displacement distance and time window average speed based on the positioning data; utilizing multiple parameters and behavior labels to train and compare various machine learning models, and selecting an optimal model as a behavior recognition model; and analyzing positioning data by using the model, and outputting rest, grazing and walking behavior categories. According to the method, long-time and large-range automatic identification of key behaviors of grazing livestock is realized, the problems of endurance and interference of a traditional sensor in a field environment are solved, the reliability and precision of behavior monitoring are improved through macroscopic motion characteristics and model optimization, and an effective data basis is provided for quantitative evaluation of a grazing management strategy.

Description

Technical Field

[0001] This invention belongs to the field of animal behavior monitoring technology, and in particular relates to a method for livestock behavior identification based on BeiDou satellite positioning and multi-parameter extension. Background Technology

[0002] Scientific and rational grazing management measures are crucial for addressing grassland degradation, and livestock feeding behavior and feed intake are key indicators for evaluating the effectiveness of grazing strategies. Feed intake directly determines livestock nutrient intake, affecting weight gain, reproduction, milk production, and productivity. Monitoring feed intake to promptly assess livestock nutritional status and studying the spatiotemporal dynamics of livestock behavior patterns, feed intake, and grassland resources under different grazing intensities helps in timely adjustments to pasture management measures and effectively improves grassland resource utilization efficiency. However, due to the complexity and variability of grassland ecosystems and the high degree of randomness in individual livestock behavior, feed intake is difficult to measure directly using feeding methods. Traditional grazing management relies mainly on experience and lacks scientific monitoring tools, resulting in a lack of precision in grazing management.

[0003] Beef cattle behavior dynamics are key indicators reflecting their physiological state, feed intake, and production performance. However, traditional manual observation methods are limited by insufficient spatiotemporal coverage and subjectivity, making it difficult to meet the needs of precise monitoring. In recent years, the development of sensor technology has provided new ideas for animal behavior identification. However, in complex pasture environments, selecting efficient and stable monitoring technologies and establishing reliable behavioral classification models still faces challenges. GPS collars, with their long battery life, wide coverage, and stable spatial data acquisition capabilities, have become a potential solution for large-scale behavioral monitoring. In addition, differences in grazing intensity may indirectly affect beef cattle activity patterns through vegetation conditions. However, current quantitative research on the relationship between "environment-behavior-management" is still insufficient, and there is a lack of integrated analysis of multi-scale (temporal and spatial) data.

[0004] With the rapid development of grazing technology, grazing management needs to evolve towards greater intelligence. Sensor-based monitoring technology can provide solutions for monitoring livestock grazing behavior, ensuring continuous and uninterrupted data collection in large-scale grassland environments, and improving the accuracy and reliability of livestock feed intake estimation. Estimating the feed intake of grazing livestock often involves several key steps, including behavior recognition and model building. Currently commonly used monitoring methods include accelerometer monitoring, acoustic sensor analysis, chewing sensor measurement, and multi-sensor fusion strategies.

[0005] Mechanical sensors, especially pressure sensors, have been widely used to monitor jaw movements in livestock and have played a crucial role in estimating feed intake. These sensors are typically mounted on the animal's nose strap, identifying chewing and biting actions by monitoring pressure changes caused by jaw movements (RUTTER et al., 1997). Pressure sensors can provide accurate measurements of feeding frequency and duration, thus aiding in feed intake estimation (RUTTER, 2000). However, in practical applications, the accuracy of pressure sensors can be affected by the tightness of the nose strap. If the nose strap is too tight or too loose, fluctuations in the pressure signal can occur, thus affecting the accuracy of the data (PAHL et al., 2016).

[0006] Acoustic sensors capture the sounds of livestock chewing and biting, enabling the identification of different types of jaw movements, including chewing, biting, and chewing-biting complexes (LACA et al., 2000). For example, an automated real-time chewing-biting algorithm system developed by researchers uses a wideband microphone mounted on the animal's forehead to accurately record the sound of jaw movements and convert it into digital signals, thereby identifying feeding behavior and effectively distinguishing between chewing and biting events (CLAPHAM et al., 2011). Through acoustic analysis, not only can livestock feeding behavior be monitored, but also key data such as feeding duration and frequency can be provided (UNGAR et al., 2006). Acoustic sensors offer high accuracy in applications and are not limited by installation location; compared to pressure sensors, they can be more conveniently used for long-term monitoring in grazing environments.

[0007] An accelerometer is a device that can monitor animal movement in real time and is widely used to analyze livestock behavior and feeding patterns. By detecting changes in an animal's acceleration, an accelerometer can effectively record the animal's dynamic movement and static posture, thereby inferring its grazing behavior (GIOVANETTI et al., 2017a).

[0008] In practical applications, the above sensors still have certain limitations. Pressure sensors rely on the stability of the nose strap installation; improper tightness can lead to significant drift or noise in the pressure signal, thereby reducing the accuracy of identifying chewing and biting events. Furthermore, the comfort of long-term wear may also affect animal behavior, causing data bias. While acoustic sensors offer high accuracy, their performance is easily affected by wind, rain, environmental noise, and grazing conditions. Additionally, acoustic devices require substantial data storage and sophisticated background noise filtering algorithms, increasing processing costs.

[0009] In contrast, while accelerometers, a widely used behavioral monitoring tool in recent years, can accurately record changes in the head and body posture of livestock, their limitations are also significant, particularly in data transmission and battery life. Accelerometers typically require continuous recording of high-frequency motion signals, resulting in massive data volumes, significantly increased power consumption, and drastically shortened battery life, making it difficult to meet the demands of long-term, non-interventional monitoring under grazing conditions. Secondly, to maintain accurate signal synchronization and real-time transmission, accelerometers often rely on Bluetooth, LoRa, or other wireless communication technologies, which are prone to signal loss, delays, or incomplete data packets in mountainous, long-distance, or wirelessly congested environments. Furthermore, accelerometer data is highly dependent on algorithms; even small amounts of noise can lead to behavioral classification errors, thus degraded signal quality significantly impacts model recognition accuracy. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention proposes a livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension, thereby resolving the issues present in the existing technologies.

[0011] To achieve the above objectives, this invention provides a livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension, comprising:

[0012] Obtain satellite positioning data and corresponding behavioral tag data of livestock at the specified time.

[0013] Based on the satellite positioning data, calculate multi-dimensional motion parameters that characterize the movement status of livestock;

[0014] Using the multi-dimensional motion parameters and the corresponding behavior label data, multiple candidate machine learning classification models are trained and their performance is compared.

[0015] Based on the performance comparison results, the optimal model is selected as the behavior recognition model;

[0016] The behavior recognition model is applied to analyze the satellite positioning data of the target livestock and output the corresponding behavior category.

[0017] Optionally, obtaining livestock satellite positioning data and corresponding time-based behavioral tag data includes:

[0018] Livestock are fitted with Beidou positioning collars, which record positioning data including timestamps, latitude and longitude coordinates, and instantaneous speed at preset time intervals;

[0019] During the data collection process using the BeiDou positioning collar, the target livestock's behavior is observed synchronously, and each behavior category and its start and end times are recorded to generate the behavior tag data.

[0020] Optionally, the multi-dimensional motion parameters include instantaneous velocity, displacement distance, and time window average velocity;

[0021] The displacement distance is calculated using the Haversine formula based on the latitude and longitude coordinates of two adjacent time points;

[0022] The average speed within the time window is calculated by dividing the sum of the displacement distances of one time interval before and after the current time point by the total time.

[0023] Optionally, the behavior categories include resting, grazing, and wandering.

[0024] Optionally, training and comparing the performance of multiple candidate machine learning classification models includes:

[0025] Using the same training data, we trained random forest, support vector machine, k-nearest neighbors, backpropagation neural network, XGBoost and CatBoost models respectively; we calculated the accuracy, precision, recall and F1 score of each model on the test set; and we selected the model with the best performance based on the calculation results.

[0026] Optionally, before training the model, the satellite positioning data is cleaned, including: removing abnormal velocity points that exceed a preset physiological velocity threshold, and interpolating or marking data missing due to signal loss.

[0027] Optionally, after outputting the behavior categories, the behavior can be processed by month and different grazing intensities, and the occurrence time of various livestock behaviors can be statistically analyzed. Then, the analysis of variance method can be used to test the significance of the behavior time between different treatments.

[0028] Optionally, before training the candidate model using the training data, the multi-dimensional motion parameter dataset carrying the behavior label data is randomly shuffled to eliminate time series dependencies in the data.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] This invention employs BeiDou / GPS satellite positioning technology to acquire the macroscopic movement trajectory of livestock and creatively extracts multi-dimensional motion parameters such as instantaneous velocity, displacement distance, and average velocity over a time window, constructing a feature system that effectively characterizes the differences in resting, grazing, and wandering behaviors. By comparing and training various machine learning models and selecting the optimal algorithm, stable and automatic identification of key livestock behaviors in complex grazing environments is achieved. This method overcomes the limitations of traditional sensors in long-term monitoring, such as insufficient battery life, significant environmental interference, and reliance on high-frequency data. It significantly improves the reliability and scalability of practical applications in the field and provides an objective and continuous data foundation for quantitatively assessing the impact of different grazing intensities on livestock behavior patterns, thereby supporting precise grazing management decisions. Attached Figure Description

[0031] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0032] Figure 1 This is a conceptual diagram of the parameter construction mode in an embodiment of the present invention;

[0033] Figure 2 A schematic diagram illustrating the accuracy of the machine learning model construction and the contribution of each parameter in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the confusion matrix results for verifying the accuracy of the machine learning model in an embodiment of the present invention.

[0035] Figure 4 This is a statistical diagram illustrating the monitoring results of beef cattle behavior under different grazing intensities according to an embodiment of the present invention.

[0036] Figure 5 This is a schematic diagram of the location of the research sample area and the grazing experiment design in an embodiment of the present invention;

[0037] Figure 6 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0038] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0039] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0040] Example 1

[0041] like Figure 6 As shown, this embodiment provides a livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension, including:

[0042] Phase 1: Data Acquisition;

[0043] Simultaneously acquire two types of data: 1) raw motion data from GPS sensors; 2) "real behavior" tagged data with timestamps from human observation.

[0044] Step 1: Experimental Design and Equipment Deployment;

[0045] 1. Select and wear the device;

[0046] In each grazing treatment, select representative beef cattle individuals (at least 2 per plot).

[0047] Wear a GPS collar. The device automatically records the following information at a preset frequency (every 5 minutes): timestamp (local time), latitude and longitude coordinates (WGS84 coordinate system), and instantaneous speed.

[0048] 2. Parameter settings: Set the GPS collar's sampling frequency to record once every 5 minutes. This interval is a compromise between "behavior capture accuracy" and "device battery life".

[0049] Step 2: Synchronize behavioral observation (data labeling). This is the most crucial step in building a supervised learning model. The specific observation method is as follows:

[0050] Time: Continuous tracking during the core activity period (6:00-18:00) every day.

[0051] Distance and method: The observer should remain at a distance of about 20 meters from the target livestock, using the naked eye or binoculars, and avoid interfering with its natural behavior.

[0052] Recording content: Use a handheld recorder (paper form) to record the following information in real time:

[0053] Behavioral types: Grazing, Walking, Resting.

[0054] Start time: accurate to the minute.

[0055] End time: accurate to the minute.

[0056] Phase Two: Data Processing and Feature Engineering;

[0057] The goal of this stage is to transform raw GPS data and manual observation data into structured datasets that machine learning models can understand and learn from.

[0058] Step 3: Data cleaning and alignment;

[0059] GPS data cleaning: Remove obvious outliers (such as points with speeds exceeding reasonable limits or extreme coordinate drift).

[0060] Check and address data loss issues (such as temporary loss of signal from the device).

[0061] Digitizing behavioral data: Inputting manually recorded time-behavior logs into spreadsheets.

[0062] Time synchronization: Ensure that the timestamps of GPS data match the time system of manual observation records.

[0063] Data fusion and slicing: Each 5-minute segment recorded by GPS is used as a time slice. Based on the human observation behavior corresponding to the center point or main time period of this segment, the GPS data point is labeled with behavioral tags.

[0064] The specific operations of data cleaning include:

[0065] Speed ​​anomaly filtering: Set a physiologically reasonable upper limit for speed (the maximum running speed of a cow is about 7-8 m / s). Remove records with instantaneous speeds >10 m / s or continuously negative values.

[0066] Coordinate drift handling:

[0067] Rapid jump, data interpolation, and coordinate clustering all fall under the category of coordinate drift processing.

[0068] Rapid jumps: These occur when a distance is displaced too quickly, usually due to multipath effects or signal loss followed by reconnection. The solution is to calculate the velocity and acceleration between consecutive points and eliminate points with accelerations far exceeding reasonable limits.

[0069] Coordinate clustering: For the slight drift of GPS coordinates during long periods of stillness (rest), median filtering or taking the mode of coordinates within that time period is used to represent its true location.

[0070] Data interpolation: For continuous missing data caused by brief signal loss (such as missing 1-2 points), linear interpolation can be used to complete the coordinates and velocity. Data missing for long periods of time is marked as invalid and not used for training.

[0071] Step 4: Feature calculation (constructing model input variables);

[0072] Since GPS coordinates alone are insufficient to fully reflect the behavior of beef cattle within 5 minutes, this embodiment expands the training data parameters into three categories: average speed, instantaneous speed, and distance. A schematic diagram of the parameters is shown below. Figure 1 .

[0073] (1) Instantaneous speed provided by the GPS collar:

[0074] The GPS collar records the instantaneous speed of the beef cattle every 5 minutes, which is a fundamental variable describing the cattle's movement state. However, due to the limited acquisition frequency and potential signal interference, the instantaneous speed at a single point in time is insufficient to reflect the cattle's overall movement status.

[0075] (2) This embodiment introduces a distance parameter to calculate the change in geographical location between the study time and the previous 5 minutes, in order to quantify the displacement of beef cattle over a short timescale. The distance is calculated using a variant of the Haversine Formula to determine the spherical distance between two GPS coordinate points on the Earth's surface, as follows:

[0076] (1)

[0077] in: : Spherical distance between two points (unit: kilometers); , : represents the latitude of two points (unit: radians); , : represents the longitude of the two points (in radians); 6371 is the average radius of the Earth, in kilometers.

[0078] (3) The average speed at the time of study is the sum of the distances traveled 5 minutes before and after the time (i.e., 10 minutes). The calculation formula is as follows.

[0079] (2)

[0080] This parameter further smooths out the fluctuations in the speed of beef cattle movement, further improves the stability of beef cattle movement speed data, and can reveal the characteristics of beef cattle behavior over a longer time scale.

[0081] In this embodiment, the behavior of not grazing and remaining still for a certain period of time is uniformly defined as "resting behavior," the behavior of beef cattle lowering their heads to graze is defined as grazing behavior, and the process of beef cattle raising their heads and moving their individual positions is defined as walking behavior. Therefore, the study defines three main categories of beef cattle behavior: resting, grazing, and walking.

[0082] This step does not directly use latitude and longitude, but instead extracts kinematic features that better characterize the behavior. Three features, as shown in Table 1, are calculated for each 5-minute time segment:

[0083] Table 1

[0084] Feature Name Calculation method Physical meaning Validity instantaneous speed Read directly from the GPS collar, in m / s. It represents the intensity of motion at that moment. High when roaming, low when grazing and resting. distance The Haversine formula is used to calculate the spherical distance between the latitude and longitude of the current 5-minute time point and the previous 5-minute time point. It represents the magnitude of displacement over the past 5 minutes. It directly reflects activity level. It is close to zero when resting, fluctuates slightly when grazing, and is greater when wandering. average speed Calculate the sum of the displacement distances in the 5 minutes before and 5 minutes after this time point, then divide by the total time (10 minutes). Formula: V_mean = (D_first 5min + D_last 5min) / 10min It represents the average level of motion within a 10-minute window surrounding that moment. It smooths out instantaneous speed fluctuations, better reflects continuous behavioral states, and is more robust to noise.

[0085] At this point, each 5-minute sample has:

[0086] Three characteristics: instantaneous velocity, distance, and average velocity;

[0087] One label: Behavior category (Resting / Grazing / Walking).

[0088] Phase 3: Machine learning model building and validation; the goal of this phase is to train the model using a labeled dataset and evaluate its performance.

[0089] Step 5: Dataset preparation;

[0090] Data shuffling: To prevent the model from learning temporal patterns in the data, the order of all samples is completely randomized. This is usually achieved in the code by setting a fixed random number seed to ensure the reproducibility of the experiment. The goal is to break time series dependencies, prevent the model from learning time-related spurious patterns, and force the model to learn the causal relationship between behavior and motion features themselves.

[0091] Dataset partitioning: Divide the shuffled data into portions proportionally (70%-30%) as follows:

[0092] Training set: Used to train model parameters.

[0093] Test set: Used to independently evaluate the final performance of the model by simulating new data.

[0094] Sampling both the training and test sets independently and identically distributed from the overall data is a prerequisite for evaluating the model's generalization ability. SVM and BPNN are sensitive to shuffling; their accuracy decreases after shuffling, indicating that these two models may, to some extent, "memorize" the temporal patterns in the original data, resulting in weaker generalization ability.

[0095] Step 6: Model training and comparison;

[0096] In constructing the machine learning model, three key parameters (average speed, instantaneous speed, and distance) provided by GPS sensors were used as input features. These parameters directly reflect the activity status of beef cattle. Furthermore, real-world observation data was used to label the behavior of the beef cattle to ensure the high quality and accuracy of the training set. The training set, through the annotation of different behavioral categories, provides the labeling information needed for supervised learning, ensuring that the model can effectively learn behavioral patterns.

[0097] Candidate Algorithm Selection: Six representative supervised learning classification algorithms were selected: RF, SVM, KNN, BPNN, XGBoost, and CatBoost.

[0098] Training the model: Use the training set to train each algorithm separately.

[0099] Model evaluation and selection: Perform predictions on all trained models using the test set; calculate and compare accuracy, precision, recall, and F1 score. Analyze the confusion matrix to see in which specific behaviors the model is prone to confusion.

[0100] Overall decision: This study ultimately chose Random Forest (RF) because it achieved the best balance among high accuracy, fast training speed, and interpretable feature importance (discovering that "distance" is the most important).

[0101] Accuracy: Measures the proportion of the model's overall correct classifications.

[0102] (3)

[0103] Precision: Represents the proportion of samples that the model predicts as a certain behavior, but which actually belong to that behavior.

[0104] (4)

[0105] Recall: The proportion of samples that actually belong to a certain behavior that are correctly identified by the model.

[0106] (5)

[0107] F1 score: A high F1 score indicates that the prediction model can balance accuracy and comprehensiveness.

[0108] (6)

[0109] Wherein: TP represents the number of samples correctly predicted as a certain livestock behavior; FP represents the number of samples incorrectly predicted as a certain livestock behavior; FN represents the number of samples that were actually a certain livestock behavior but were not correctly predicted; TN represents the number of samples correctly predicted as not that behavior.

[0110] Phase Four: Model Application and Analysis;

[0111] Step 7: Behavior recognition and statistics;

[0112] Full Prediction: Using the selected optimal model (RF), behavioral classification predictions are performed on all GPS data (including days without human observation) for all livestock throughout the grazing season (June-September). In this way, each cow is assigned a predicted behavioral state for every 5 minutes of each day.

[0113] Time statistics: processed by month and by grazing intensity, summarizing the total time each cow spends grazing, resting, and wandering each day.

[0114] Significance analysis: Statistical methods (such as ANOVA analysis combined with post-hoc tests such as Tukey HSD) were used to test whether there were significant differences in grazing time, resting time, and wandering time of livestock under different months and different grazing intensities (p<0.05).

[0115] This embodiment uses six machine learning algorithms to identify three behaviors in livestock (resting, grazing, and wandering). Accuracy verification results show that Random Forest (RF), XGBoost, and CatBoost achieve accuracy exceeding 0.85 on both the training and test sets, demonstrating the best performance. In contrast, BPNN and SVM algorithms show lower accuracy. Figure 2 As shown in 'a'.

[0116] Both XGBoost and CatBoost borrow ideas from the Random Forest algorithm, using different splitting criteria to evaluate the contribution of each feature to the performance of the decision tree model. The contribution of each parameter differs among the three algorithms. In the Random Forest model, the distance livestock travels contributes the most to livestock behavior features, while instantaneous speed contributes the least. In the XGBoost and CatBoost models, average speed contributes the least to livestock behavior features. In CatBoost, the contributions of the three parameters are more evenly distributed. Figure 2 As shown in b in the figure.

[0117] Because XGBoost and CatBoost employ gradient boosting, the result of each round of training depends on the output of the previous round, thus lacking full parallelization, which makes their training time relatively long. Random Forest, on the other hand, is relatively simple and has a shorter training time. Figure 2 As shown in c in the figure.

[0118] The confusion matrix results show the recognition accuracy of six algorithms for three behaviors in livestock, such as... Figure 3As shown, Random Forest (RF), XGBoost, and CatBoost algorithms can all effectively identify livestock resting, grazing, and walking behaviors. All three algorithms are particularly accurate in identifying walking behavior. There is some confusion between resting and grazing behaviors. Support Vector Machine (SVM) is only effective at identifying grazing behavior, while Backpropagation Neural Network (BPNN) performs poorly in distinguishing different livestock activity behaviors. SVM and BPNN algorithms are significantly affected by data shuffling (i.e., disrupting the GPS data time series, preventing the algorithms from learning the regularity of the time series). Without data shuffling, these two algorithms can achieve an accuracy of over 70%.

[0119] Considering both the accuracy and training efficiency of the algorithm, this study selected the Random Forest (RF) algorithm as the main method to identify the behavioral dynamics of livestock in different months and conduct further analysis and research. The parameters of the Random Forest algorithm are shown in Table 2 below.

[0120] Table 2

[0121] Parameter name Parameter value Data Splitting 0.161s Data Splitting 0.7 Data Shuffling yes Cross-Validation (CV) no Criterion for Node Splitting gini Number of Trees 100 Bootstrap Sampling TRUE Bootstrap Sampling FALSE Max Features auto Minimum Samples for Node Splitting 2 Minimum Samples per Leaf 1 Minimum Leaf Weight 0 Maximum Depth of the Tree 10 Maximum Depth of the Tree 50 Threshold for Node Impurity 0

[0122] Prediction results:

[0123] In June, the grazing time of moderately grazed cattle (G0.46) was significantly longer than that of other grazing intensities (G0.23, G0.34, G0.92, p<0.05), with the shortest grazing time observed under heavy grazing, showing a trend of first increasing and then decreasing with grazing intensity. In July, the grazing time of beef cattle under heavier grazing (G0.69) was significantly longer than that under light grazing (G0.23), also showing a trend of increasing and decreasing with grazing intensity, with the maximum grazing time occurring at G0.69. In August... The grazing time of beef cattle showed an increasing trend with the increase of grazing intensity, with heavy grazing (G0.92) > moderate grazing (G0.46) > light grazing (G0.23), showing a significant difference (p<0.05). In the later stage of grazing (September), the grazing time of livestock generally increased, showing a significant difference (p<0.05), with moderate grazing (G0.46) > very heavy grazing (G0.92) > relatively light and heavy grazing (G0.34, G0.69). Overall, the grazing time of beef cattle showed a trend of first decreasing and then increasing, reaching its lowest point in August. The changes in grazing time of beef cattle under different grazing intensities showed that in June and July, it first increased and then decreased with the increase of grazing intensity, and then continued to increase with the increase of grazing intensity in August.

[0124] No significant differences were observed in rest time for beef cattle under different grazing intensities in June. However, in July, moderate grazing (G0.46) and light grazing (G0.23) significantly exceeded other grazing intensities (p<0.05). In August, rest time under light and heavy grazing (G0.23, G0.34, G0.69) was significantly longer than under very heavy grazing (G0.92). In September, rest time under moderate grazing (G0.46) was significantly longer than other grazing intensities (p<0.05).

[0125] The overall trend for the wandering time of beef cattle showed a decrease followed by an increase. Specifically, the wandering time was lowest under moderate grazing, while it was higher under heavy grazing (G0.92) in June, August, and September, and showed significant differences compared to other grazing intensities. Figure 4 As shown.

[0126] The experimental plots in the study area were single-factor grazing control experiments. The experiment setup included six different grazing intensities, with three replicates for each intensity, totaling 18 experimental plots of 5 hectares each. Grazing was conducted annually from June 1st to October 1st, lasting 120 days. Different grazing intensities were controlled by introducing 0, 2, 3, 4, 6, and 8 head of beef cattle weighing 250–300 kg. The corresponding stocking rates were 0.00, 0.23, 0.34, 0.46, 0.69, and 0.92 cow·AU Ha. -1 (One standard beef cattle unit (AU) equals 500 kg). Based on carrying capacity, the six grazing intensities can be categorized into the following four types: Grazing prohibition areas (0.00 cow·AU Ha). -1 ), lightly grazing areas (0.23 and 0.34 cow·AU Ha) -1 ), moderately grazing areas (0.46 cow·AU Ha) -1 ) and heavily grazing areas (0.69 and 0.92 cow·AU Ha) -1 The study included the location, landscape, and plot design of the residential area. Figure 5 As shown.

[0127] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for livestock behavior recognition based on BeiDou satellite positioning and multi-parameter extension, characterized in that, Includes the following steps: Obtain satellite positioning data and corresponding behavioral tag data of livestock at the specified time. Based on the satellite positioning data, calculate multi-dimensional motion parameters that characterize the movement status of livestock; Using the multi-dimensional motion parameters and the corresponding behavior label data, multiple candidate machine learning classification models are trained and their performance is compared. Based on the performance comparison results, the optimal model is selected as the behavior recognition model; The behavior recognition model is applied to analyze the satellite positioning data of the target livestock and output the corresponding behavior category.

2. The livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension according to claim 1, characterized in that, Obtaining livestock satellite positioning data and corresponding time-based behavioral tag data includes: Livestock are fitted with Beidou positioning collars, which record positioning data including timestamps, latitude and longitude coordinates, and instantaneous speed at preset time intervals; During the data collection process using the BeiDou positioning collar, the target livestock's behavior is observed synchronously, and each behavior category and its start and end times are recorded to generate the behavior tag data.

3. The livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension according to claim 1, characterized in that, The multidimensional motion parameters include instantaneous velocity, displacement distance, and average velocity over a time window; The displacement distance is calculated using the Haversine formula based on the latitude and longitude coordinates of two adjacent time points; The average speed within the time window is calculated by dividing the sum of the displacement distances of one time interval before and after the current time point by the total time.

4. The livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension according to claim 1, characterized in that, The behavioral categories include resting, grazing, and wandering.

5. The livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension according to claim 1, characterized in that, Training and performance comparison of multiple candidate machine learning classification models includes: Using the same training data, we trained random forest, support vector machine, k-nearest neighbors, backpropagation neural network, XGBoost and CatBoost models respectively; we calculated the accuracy, precision, recall and F1 score of each model on the test set; and we selected the model with the best performance based on the calculation results.

6. The livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension according to claim 2, characterized in that, Before training the model, the satellite positioning data is cleaned, which includes: removing abnormal velocity points that exceed a preset physiological velocity threshold, and interpolating or marking data missing due to signal loss.

7. The livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension according to claim 4, characterized in that, After outputting the behavior categories, the occurrence times of various livestock behaviors were statistically analyzed by month and different grazing intensities, and the significance of the behavior times between different treatments was tested using analysis of variance.

8. The livestock behavior recognition method based on BeiDou satellite positioning and multi-parameter extension according to claim 5, characterized in that, Before training the candidate model using the training data, the multi-dimensional motion parameter dataset carrying the behavior label data is randomly shuffled to eliminate time series dependencies in the data.

Images