A robust classification system for Southern Yellow cow behavior using 3-DoF accelerometers

2.1 Animals and farm condition

In developing countries, one of the primary sources of beef production is Vietnam Yellow Cattle. They have body weights (of an adult cow) from 180 to 300 kg. They can adapt to the local tropical environment of Vietnam.

Four cows of a farm in Ba-Vi, Hanoi, Vietnam were chosen for the experiment. These selected cows have ages ranging from 2 to 4 years. They are free to move in a large area (about 3.5 hectares as shown in Fig. 1). The experiment was done in 10 days (April 01-10, 2020).

OPEN IN VIEWER

Fig. 1

Field studied area.

The device is mounted on the front leg of the cow, as shown in Fig. 2. Wearing the device in this position makes it possible for the sensor to sense three-dimensional motion conveniently.

OPEN IN VIEWER

Fig. 2

The device is mounted on the front leg of the cow.

From the sensor’s data, the conversion formula to determine the acceleration value A k ( S i / 1024 × R - O i ) / Se i(1)

where

R: voltage reference;

O _i : offset;

S _i : value after sampling of axis k (k = X, Y, Z);

A _k : acceleration value in axis-k;

Se _i : sensitivity on the axis-k [33].

In this study, the data acquisition device has dimensions of 85×60×35 mm, and a weight of 300 g (see Fig. 3). The device uses a microcontroller ATMega328 to process the accelerometer signal from the MPU-6050 and transmit the data to the LoRawan RA-02 module. Each cow has a unique identifier to pack this data without confusion.

OPEN IN VIEWER

Fig. 3

A photo of sensor MPU6050 (a), and the bangle (b).

The steps of using the GBDT classifier (16 seconds window), giving the best overall performance, are presented in Fig. 4.

OPEN IN VIEWER

Fig. 4

Behavior recognition processing pipeline.

Training data and test data were selected randomly from all data. The training/test ratio of data at 60% /40% was used for leg-mounted acceleration data [6, 19]; the results were good. This ratio also provided the best results with the experiment data. The proposed features set (the details in Section 2.5) were calculated for all windowed records (Fig. 4) to train the classification model by ML algorithm (GBDT) based on labeled training data [17, 19]. This classification model would be used to predict the classes (behaviors) of test data. In other words, all records of test data would be labeled (as feeding, lying, standing, or walking), with only one label for each record. In the evaluation step, the labeled data from real observations were be used to evaluate the labeled data by the classification model (the details in Section 2.6) to check the effectiveness of the proposed classification model. Figure 5 shows the steps of the proposed classification.

OPEN IN VIEWER

Fig. 5

The proposed classifier constructive process.

The model was evaluated using three performance indexes (Section 2.6) based on the behavior observations log. This work was implemented using Python 3.5.

The foot-mounted device has been attached to four cows. The sensor receives twenty data samples per axis every second, then sends them to the microcontroller. The raw acceleration data is averaged to smooth and store the data at one sample per second. To exploit the statistical properties of the collected data, 16 samples were concatenated into a window with 10 overlapping samples (Section 2.3). This frame is used to extract five proposed features. Table 1 presents the raw data summary in minute unit.

Table 2 is a summary of the definitions of behavior in the data set.

Table 3 is an example of acceleration data when the cow is walking. The numbers of observations (thus, the numbers of records) and their composition are listed in Table 4. The total observations for all cows in the experiment are 12911. The total observations for individual cows are not significantly different, from 2982 to 3603. The observations of standing (4527) are greater compare to other behaviors.

The proposed method for classifying cow behavior uses five accelerometer data characteristics: mean, median, SD, RMS, and range. The reason for choosing these five types of features is based on the nature of the signal used for classification [23, 31]. These are time-series signals (acceleration values obtained vary with time). With this type of signal, engineers can exploit information in the time domain or frequency domain. However, if choosing to exploit frequency domain features will lead to significant computational complexity. Therefore, the researcher decides to use the features on the time domain, namely the 1st and 2nd order measures, to reduce the computational complexity but still ensure the correct classification. Various tests have been carried out to estimate the ability to classify behavior based on the difference of the obtained metric values. Finally, the above five features are selected. A window selection and feature extraction scheme have proposed for the recognition process. The recognition process starts with the accelerometer data extraction feature [37].

All the formulae (2–6) are illustrated for X-axis acceleration: range ( X j ) = [ min i = 1 N { x i } , max i = 1 N { x i } ](2)

where X are acceleration data;

X _j : record j;

N: number of samples in a data frame;

x _i : the i^th sample of the record X _j ;

min: X _j minimum value;

max: X _j maximum value;

range(X _j ): distance from min to max.

m ( X j ) = 1 N ∑ i = 1 N x i(3)

median ( X j ) = x [ # N / 2 + 1 ] + x [ # N / 2 ] 2(4) where m(X _j ) is the mean of X _j

In formular (4), x _i values in X _j is sorted. σ ( X j ) = 1 N ∑ i = 1 k ( x i - m ) 2(5) where σ(X _j ) is the standard deviation of X _j . RM S X j = 1 N ∑ i = 1 N x i 2(6) where RMS _Xj is the root mean square of X _j .

The algorithm’s performance was evaluated based on accuracy, sensitivity, and positive predictive value. Table 5 is an example confusion matrix to use in classifying cow behavior.

Formulae of evaluation indicators: Accuracy = TP + TN TP + TN + FP + FN(7) Sensitivity = TP FN + TP(8) PPV = TP FP + TP(9) where PPV is the positive predictive value.

3.1 Performance of GBDT for behavioral classification

The performance of classification in terms of accuracy, sensitivity, and PPV was calculated by Equations (7)–(9) for four behaviors. Table 6 presents the confusion matrix for four cows using GBDT classifier. The predicted behavior using the classification are compared to the true behavior. The bold values are the number of TP (when the predicted behavior and true behavior are the same). The other values (not bold) are FP, FN, and TN (see Section 2.6 for the details). The results of the individual cow are presented in four tables: Table 6(a-d) for cow ID01, ID02, ID03, and ID04, respectively.

The performance of the proposed model was examined using three indicators as shown in Equations (7)–(9). Table 7 shows the performance of the GDBT algorithm. It was calculated for all individual cows. From Table 7, lying and walking classification prediction were excellent (0.99-1 accuracy, 0.96-1 sensitivity and 0.98-1 PPV). Feeding and standing resulted in lower performance, the sensitivity of standing for cow ID03 was less than 0.50.

In the dataset of this work, the proportion of behavior data was an imbalance. For example, Table 4.B&4.C show that feeding has 158 records while average is about 298 records for all behaviors. This problem is challenging, and difficult to be solved. So, the evaluation classification (Table 8) based on the micro-average method because this method computes average values by aggregating all classes’ contributions. Wang et al. [19] used this evaluation approach for the same reason.

Table 9 shows that the average performance of the proposed method was good: 0.91 accuracy, 0.85 sensitivity, and 0.85 PPV. The results proved that the proposed system provided good classification performance in the case of Southern Yellow cattle. The performance of this work (bold values) is better than Wang et al. [19] and P. Martiskainen and M. Jarvinen [30] for all evaluation indicators.

Grazing time was estimated using accelerometers in the work [12]. The performance of the estimation was 0.82 PPV and 0.7 sensitivity, not high, even with limited behaviors. It showed that for the behavior classification problem, working with grazing cows is more challenging than free-stall barn cows.

Accelerometer thresholds were proposed for the classifier of [3, 4]. That excellent approach saves the model training time, and still archive high accuracy. However, it may not work well if the cows have not similar conditions such as body size, age, sex, etc. In a real scenario, cows should have different parameters.

Four cows in the experiments were separated into two groups. ID01 and ID03 cows belonged to group 1, the rest (ID02 ID04) belonged to group 2. Table 8 shows that group 1 obtained better classification performance than group 2. The reason is cows in group 1 are accustomed to wearing devices.

Using leg tag sensors has the limit in the classification of standing and feeding. This was shown in this paper and [6, 19]. The problem of data imbalance should also be considered because the collected behavior data will be difficult to balance. This is also a common problem that all machine learning problems need to be solved [38].

The future work combines both leg sensor and collar sensor on the same animal in a sensor network to design an IoT based system for monitoring cow. In the future, there will have many improvements for the sensor network, such as self-organization, multi-protocol compatibility, high throughput [5], and prolonging lifetime [39]. It will support IoT (Internet of Things) technology. IoT is considered one of the keys to smart agriculture and many other fields [5].