Trajectory classification to support effective and efficient field-road classification

Overview

As shown in Fig. 2, the proposed automatic trajectory classification method consists of three parts: data preprocessing, feature extraction, and random forest (RF)-based classification.

• Data preprocessing: it provides more parameters for each trajectory to support its feature extraction.

• Feature extraction: it includes global feature extraction and local feature extraction.

  • Global feature extraction: it extracts 123 features based on the overall characteristics of a trajectory, such as direction change, time difference, and so on.

  • Local feature extraction: it extracts 52 features based on the characteristics of a segment, where a segment is part of a trajectory.

• RF-based classification: based on 175(=123+52) features, a tree-based classification method, random forest (RF), is applied to develop a trajectory classification model. Such a model can identify the category of a trajectory (“high-quality”, “medium-quality” or “low-quality”).

Data preprocessing

Based on the five recorded parameters of a GNSS point, data preprocessing was carried out to obtain more useful parameters. Specifically, each GNSS point in our dataset is represented by a fourteen-parameter vector including longitude, latitude, direction, speed, acceleration, jerk, bear rate, longitude difference, latitude difference, direction difference, time delta, sinuosity, distance, and vin. The former four parameters (longitude, latitude, direction, and speed) are directly recorded, and the latter nine parameters are derived. Moreover, jerk, bear rate, longitude difference, and latitude difference illustrate the relationship between the current point and its predecessor; time delta represents the time interval between two consecutive points; sinuosity measures the curvature between two successive paths; distance denotes the actual distance between two successive points, and vin is calculated by dividing distance by time delta (Chen et al., 2022).

Global feature extraction

Similar to the previous work (Xiao et al., 2017), 123 global features are extracted to represent a trajectory, where global features are the statistics of all points in a trajectory. As shown in Table 2, 17 statistical measures are used, seven parameters are selected, and then a statistical value is calculated per parameter and statistical measure (Mardia & Jupp, 1999). Besides the four parameters (speed, sinuosity, acceleration, and direction difference) used in Xiao et al. (2017), additional three parameters are selected: distance, time delta, and vin. Distance and time delta can reflect point-density-based characteristics, and vin is indicative of movement-based characteristics (Xiao et al., 2017). Then, as the study of Zheng et al. (2008) pointed out kinetic parameters can enhance the robustness of trajectory classification models, in addition to the three kinetic parameters (heading change rate, stop rate, and velocity change rate) used in Xiao et al. (2017), an additional kinetic parameter, total distance, is employed to capture point-density-based characteristics.

Local feature extraction

Inspired by the work of Dodge, Weibel & Forootan (2009), we designed a local feature extraction method specifically for agricultural machinery, which extracts 52 local features to enhance the feature representation of a trajectory. As illustrated in Fig. 2, the local feature extraction consists of three steps: trajectory segmentation which provides a set of segments, segment classification which automatically classifies segments into four categories, and segment-based feature extraction which extracts features based on the segments and their respective categories.

Firstly, three trajectory segmentation methods are conducted separately, and each of them slices a trajectory into a sequence of segments. As these methods have varying strengths and weaknesses in extracting characteristics, all of them are employed in this study.

• Segmentation based on fixed time interval: a trajectory is sliced into segments with a fixed time interval.

• Segmentation based on fixed point number: a trajectory is split into segments with a fixed number of points.

• Segmentation based on spare time: if the time interval between two successive points is more than a threshold (e.g., 600 s), a split is made in the middle of the two points.

Secondly, according to the two kinds of GNSS signal qualities (strong and weak/noisy) and the two distinct movement behaviors (driving straight and making a turn), the segments are classified into the following four categories to reflect the combinations of two crucial characteristics: point density and driving direction.

• Low direction change and low point density: it represents the case that agricultural machinery drives straight normally yet with weak/noisy GNSS signals.

• Low direction change and high point density: it represents the case that agricultural machinery drives straight normally with strong GNSS signals.

• High direction change and low point density: it represents the case that agricultural machinery makes a turn normally yet with weak/noisy GNSS signals.

• High direction change and high point density: it represents the case that agricultural machinery makes a turn normally with strong GNSS signals.

Finally, as detailed in Table 2, 12 local features are extracted per category. Firstly, six parameters (direction delta, distance, speed, sinuosity, bear rate delta, and vin delta) and two statistical measures (mean and standard deviation) are selected. Then, for each segment, a statistical value is calculated for each parameter and measure. Finally, for each category, the mean of the statistical values of segments belonging to that category are computed.

Experimental settings

Baseline methods: to demonstrate the effectiveness of our feature extraction methods, a comparison was made between our proposed feature extraction features and two existing feature extraction methods that have been commonly used in the transportation mode classification studies: original global feature extraction and original global+local feature extraction (Dodge, Weibel & Forootan, 2009; Xiao et al., 2017; Meng et al., 2020; Hechifa et al., 2023).

1. Original global feature extraction (Original GF): it follows the work of Dodge, Weibel & Forootan (2009) to extract 71 features based on four parameters (speed, sinuosity, acceleration, and direction_diff), as shown in the global feature part of Table 2;

2. Original global+local feature extraction (Original GF+LF): it follows the work of Xiao et al. (2017) to extract 119 features including the 71 original global features and 48 local features in Table 2;

3. Overall global feature extraction (Overall GF): it is our global feature extraction including the 123 global features in Table 2.

4. Overall global+local feature extraction (Overall GF+LF, our feature extraction): it extracts 175 global+local features including the 123 global features and 52 local features in Table 2.

Model training and evaluation: the experiment of trajectory classification consists of three steps: data split, model training, and model test. The TC data was randomly split into 80% and 20% for training and testing, the training data was used to train a trajectory classification model, and the test data was used to evaluate the performance of the developed model. Similar to the previous work on a classification problem (Chen et al., 2021; Chen et al., 2022; Zhang et al., 2022), four metrics, i.e., precision, recall, F1-score, and accuracy, were used to evaluate model performance.

Method comparisons

Table 3 lists the performances of trajectory classification using the four feature extraction methods. Compared to the baselines, our proposed feature extraction (i.e., overall global+local features) performs best, achieving 86.24% in F1-score and 86.67% in accuracy. This indicates that the design of our feature extraction is effective for the quality classification of agricultural machinery trajectories.

Furthermore, performance improvement mainly comes from incorporating global features specific to agricultural machinery. As illustrated in Table 3, compared to “original global features”, the adding of the specific global features in “overall global features” yields 4.88% and 1.9% increase in F1-score and accuracy, respectively. This confirms the effectiveness of additional global parameters in our feature extraction. Moreover, the incorporation of local features also can enhance trajectory classification. For example, compared to “overall global features”, “overall global+local features” yield 0.74% and 0.7% improvement in F1-score and accuracy, respectively. This shows that certain characteristics of agricultural machinery may be discernible only within particular segments of a trajectory. In summary, the effective extraction of global and local features specific to agricultural machinery leads to a promising trajectory classification model.

Finally, Table 4 lists the performances of these trajectory classification methods for each quality category. We can observe that our feature extraction method improves the performance of the “medium-quality” and “low-quality” classes, but hurts the performance of the “high-quality” class. For example, comparing “overall global features” to “original global features”, the F1-score increases by 12.87% on the “medium quality” class and 4.94% on the “low-quality” class, but decreases by 3.19% on the “high-quality” class. The results demonstrate that the additional features used in our global feature extraction method significantly enhance the representation of medium-quality data, and moderately improve low-quality data, but bring a negative impact on the representation of high-quality data. Comparing “overall global+local features” to “overall global features”, the F1-score increases by 0.75% on the “high-quality” class, 0.64% on the “medium quality” class, and 0.80% on the “low-quality” class. This demonstrates that the additional local feature in our method can contribute to a consistent improvement in data representation. In summary, these additional features enhance the overall performance, which validates the effectiveness of our feature extraction method.

Experimental settings

Baseline methods: to show the efficiency and effectiveness of the proposed TCModel-based DBSCAN+GCN, it was compared with four state-of-the-art field-road classification methods, DBSCAN + Rules (Chen et al., 2021), DT (Poteko, Eder & Noack, 2021), DBSCAN + OD +DBI (Zhang et al., 2022), and GCN (Chen et al., 2022), and another hybrid field-road classification method, TCRule-based DBSCAN+GCN (a hybrid field-road classification using the DBSCAN + Rules and GCN method according to the result from rule-based trajectory classification). TCRule-based DBSCAN+GCN uses rule-based trajectory classification as follows: if the majority of the logging intervals in a trajectory are not greater than 5 s, DBSCAN + Rules is selected; otherwise, the GCN method is used. Among the four baseline field-road classification methods, in terms of effectiveness, the GCN method demonstrates outstanding performance in both the wheat harvester trajectory dataset and the paddy harvester trajectory dataset, surpassing the other three methods. In terms of efficiency, the DBSCAN method proves to be significantly more time-efficient, compared to the other three methods.

Model training and evaluation: the following two experiments were carried out to evaluate the effectiveness and efficiency of six field-road classification methods, respectively.

Effectiveness evaluation experiment: except TCRule-based DBSCAN+GCN, other five field-road classification methods (DBSCAN + Rules, DT, DBSCAN + OD + DBI, GCN, TCModel-based DBSCAN+GCN) were conducted independently on the FRCEFfct data. Firstly, the wheat harvester trajectory dataset (or the paddy harvester trajectory dataset) was randomly split into 80% and 20% for training and testing. Then, the field-road classification method was used to train a model with the training data, and the effectiveness of the developed model was evaluated on the test data. Notice that the trajectory classification model used in TCModel-based DBSCAN+GCN already had been trained with all trajectories in the TC data. Consistent with prior research (Chen et al., 2021; Chen et al., 2022; Zhang et al., 2022), four metrics (i.e., precision, recall, F1-score, and accuracy) were used to evaluate model performances. Moreover, the ignoring of TCRule-based DBSCAN+GCN is attributed to the characteristics of the FRCEFfct data. Specifically, for the wheat harvester trajectories, the majority of the logging intervals are less than 5 s, and only the DBSCAN + Rules method is used. Meanwhile, for the paddy harvester trajectories, the majority of the logging intervals are greater than or equal to 5 s, only GCN is used.

Efficiency evaluation experiment: all six field-road classification methods were conducted independently using both the FRCEFfct data and the FRCEFfcy data. In general, for each field-road classification method, there are two steps: training and testing. In the training phase, a field-road classification model was developed using the wheat harvester dataset within the FRCEFfct data, and in the testing phase, the model’s time consumption was evaluated on the FRCEFfcy data, where “time consumption” is defined as the total time interval from the initiation to the completion of the testing phase, including processes such as data loading, execution of field-road classification, and storage of classification results. Moreover, for the TCModel-based DBSCAN+GCN, the time consumption associated with trajectory classification was taken into account, which refers to the total time required for predicting trajectory quality.

Method comparisons

Effectiveness evaluation: Tables 5 and 6 list the performances of the five field-road classification methods on the wheat and paddy harvester trajectory datasets in the FRCEFfct data, respectively. When the dataset is altered, such as switching from the wheat harvester trajectory dataset to the paddy harvester trajectory dataset, there is a notable change in the performance of the same field-road classification method. Among the five methods, the performance of GCN decreases slightly, and the performance of DBSCAN + Rules drops a lot, which confirms the robustness of the GCN method. The observed variations in the performance of the same field-road classification method can be primarily attributed to the alterations in logging interval settings of GNSS receivers used during data collection (Chen et al., 2022). For example, TCModel-based DBSCAN+GCN achieves 86.79% accuracy on the wheat harvester trajectory dataset whose logging interval setting is 5 s, and obtains 84.82% accuracy on the paddy harvester trajectory dataset whose logging interval setting is 30 s. The prolonged logging interval typically results in low-quality trajectories, which further affect the performance of field-road classification. So, the quality of the trajectories is a critical factor in determining the effectiveness of field-road classification methods.

Moreover, the proposed TCModel-based DBSCAN+GCN achieves acceptable performance. Specifically, on the wheat harvester trajectory dataset, TCModel-based DBSCAN+GCN surpasses DBSCAN + Rules, DT, DBSCAN + OD +DBI, and by 2.14%, 2.56% and 5.17% in accuracy, respectively. However, compared to GCN, it has a 0.17% drop in accuracy. Similarly, on the paddy harvester trajectory dataset, compared to DBSCAN + Rules, DT, DBSCAN + OD + DBI, and GCN, TCModel-based DBSCAN+GCN achieves 13.22%, 8.41%, 12.58%, and 0.27% increase in accuracy, respectively. These results confirm the effectiveness of our TCModel-based DBSCAN+GCN, which is a hybrid approach that integrates the DBSCAN + Rules and GCN methods.

Efficiency evaluation: Figure 3 shows the time consumption of the six field-road classification methods inputted with a different number of trajectories, where the input trajectories were randomly selected from the FRCEFfcy dataset.

As the number of input trajectories increases, the time consumption increases continuously for all methods, where DBSCAN+Rules increases most slowly, and DT increases fastest. Among the whole process of a given field-road classification method, feature extraction is the most time-consuming step. However, the time does not increase linearly with the number of input trajectories due to varying points per trajectory, resulting in a nonlinear time cost in feature extraction.

Additionally, the TCModel-based DBSCAN+GCN displays a distinct advantage when processing over 5,000 trajectories. In a scenario where all trajectories in the FRCEFfcy dataset are utilized, the time consumption of TCModel-based DBSCAN+GCN amounts to approximately 83% of that required by TCRule-based DBSCAN+GCN. This finding substantiates the superiority of our trajectory classification model over the methods relying solely on logging interval settings in terms of field-road classification method selection.