DBSCAN-MFI Based Improved Clustering for Field-Road Classification in Mechanical Residual Film Recovery

Abstract

Accurate accounting of residual film recovery operation areas is essential for supporting targeted implementation of white pollution control policies in cotton fields and serves as a critical foundation for data-driven prevention and control of soil contamination. To address the reliance on manual screening during preprocessing in traditional residual film recovery area calculation methods, this study proposes a DBSCAN-MFI field-road trajectory segmentation method. This approach combines DBSCAN density clustering with multi-feature inference. Building on DBSCAN clustering, the method incorporates a convex hull completion strategy and multi-feature inference rules utilizing speed-direction feature filtering to automatically identify and segment field and road areas, enabling precise operation area calculation. Experimental results demonstrate that compared to DBSCAN, OPTICS, the Grid-Based Method, and the DBSCAN-FR algorithm, the proposed algorithm improves the F1-Score by 7.01%, 7.13%, 7.28%, and 4.27%, respectively. Regarding the impact on operation area calculation, segmentation accuracy increased by 23.61%, 25.14%, 20.71%, and 6.87%, respectively. This study provides an effective solution for accurate field-road segmentation during mechanical residual film recovery operations to facilitate subsequent calculation of the recovered area.

1. Introduction

Plastic mulch films function to increase soil temperature, retain soil moisture, maintain soil structure, and promote plant growth [1]. However, long-term improper use has led to significant accumulation of residual plastic fragments in the soil, altering land structure and causing agricultural non-point source pollution [2,3]. Researchers have conducted in-depth studies on core component design [4], whole-machine optimization [5], material parameter calibration [6], performance simulation of key components [7], and whole-machine performance trials [8,9] of residual film recovery machinery, achieving significant progress. Consequently, the operational performance of these machines has been enhanced.

Accurate delineation of field and road areas based on spatiotemporal trajectory data, along with precise calculation of operation areas, plays a crucial role in enabling regulatory authorities to monitor residual film recovery operations and formulate and implement effective guidance policies [10]. Traditional tape measurement methods can effectively measure small, regular plots but require substantial human and material resources. Additionally, measurement results are subject to significant human error, limiting their applicability [11]. While the semi-supervised recognition [12], automated annotation [13], and lightweight model deployment [14] have enhanced crop monitoring technologies, the operational field-road segmentation in agricultural machinery remains limited by complex spatiotemporal and sensor constraints. These technologies are mainly used in road recognition to assist agricultural machinery in path planning within specific areas [15,16]. Notably, Borowiec and Marmol [17] used a LiDAR sensor to determine the edges of fields, establishing cropland perimeters for subsidy verification through high-precision area computation. The field-road segmentation used for area computation mainly relies on satellite navigation technology [10,18,19].

Machinery operation trajectory data encompasses field and road segments, necessitating classification prior to area calculation. A common approach leverages the spatial aggregation characteristics of spatiotemporal trajectory points to delineate fields and roads. Li et al. [20] preliminarily applied DBSCAN clustering to identify field areas during agricultural machinery operations and calculated the operational area based on clustering results. However, its effectiveness in identifying road travel regions declines when reduced speed increases point density in those areas. Wang et al. [21] combined the linearity of road trajectories and spatial aggregation of field trajectories as spatial features with Grid-Based density clustering to analyze positioning data states for field-road segmentation. Yet, its oversimplified spatial trajectory model performs poorly under uneven density distributions. Consequently, deeper inference following initial clustering is often employed to enhance classification performance. Chen et al. [22] utilized DBSCAN to detect point categories within trajectories, then corrected clustering results based on parallel directional features between adjacent trajectory lines. This method exhibits good computational efficiency, but classification accuracy is sensitive to raw data quality. Chen et al. [23] proposed a classification method based on Graph Convolutional Neural Networks (GCNs), constructing spatiotemporal graphs from spatial and temporal relationships between trajectory points. Graph convolution then extracts spatiotemporal features for field-road classification, demonstrating robust performance. However, substantial time and computational resources are required to build spatiotemporal graphs and train complex neural networks. Poteko et al. [24] developed a decision tree-based model for agricultural vehicle operation mode recognition using only speed, ground heading, and derived parameters (acceleration, curve radius, and angular velocity), omitting field boundary and vehicle position information. While effective at stable speeds, classification errors increase significantly during stops at intersections or near field boundaries where speeds converge. Zhang et al. [25] introduced a multi-perspective density-based field-road classification method, clustering trajectory points separately using DBSCAN and an object detection model, then selecting the superior result via the Davis–Bouldin Index (DBI). To balance time consumption and accuracy, Chen et al. [26] introduced the concept of trajectory data quality, proposing a strategy to select appropriate classification methods accordingly. However, as data quality is dynamic—influenced by field conditions, operators, and signal variations—its precise quantification remains challenging. Xiao et al. [27] presented a DR-XGBoost model for identifying agricultural machinery operation modes, requiring minimal trajectory features as input and optimizing feature sets through recursive feature elimination. Nevertheless, an extensive grid search is needed to determine optimal time-window hyperparameters, incurring high computational complexity. Zhai et al. [28] employed Generative Adversarial Networks (GANs) to balance field and road trajectory data, followed by Bidirectional Long Short-Term Memory networks with attention mechanisms (Att-BiLSTM) to capture spatiotemporal relationships between adjacent points and focus on critical points. However, this approach inadequately captures dependencies within individual trajectory points. Thus, Zhai et al. [29] advanced a BiLSTM-SAGCN hybrid network: BiLSTM extracts intra-point feature correlations along feature dimensions, while the Spatiotemporal Adaptive Graph Convolutional Network (SAGCN) models inter-point dependencies for comprehensive feature representation. But the intricate model architectures and complex spatiotemporal feature extraction hinder broad adoption in practical applications. Wang and Wang [30] integrated transformer and semantic technologies to create an advanced semantic encoder. The approach imbalance issue between field-road pixels is effectively addressed by employing a pixel-wise weighted cross-entropy loss function in this study. Chen et al. [31] proposed a MultiDNN model integrating BiLSTM, Topology Adaptive Graph Convolution (TAG), and Self-Attention Networks (ATT) to effectively extract spatiotemporal features for classification. Han et al. [32] introduced an efficient classification method combining spatiotemporal clustering with semantic segmentation. These deep learning models enhance overall field-road segmentation performance by integrating visual and kinematic features. However, complex graph construction and training requirements limit their feasibility for edge deployment.

It should be noted that residual film recovery involves coordinated operations across multiple stages, characterized by the following: (1) Operational quality is influenced by soil conditions and planting practices, resulting in variable implement working speeds; (2) The process includes slow-speed phases such as bale unloading, turning maneuvers, clearing, and maintenance; (3) Operators typically adopt spiral operation patterns with repeated traversals along headland areas. These characteristics lead to clustered data point accumulation in both field interiors and headlands, resulting in non-uniform trajectory point distributions that exhibit plot-specific variations. Existing algorithms demonstrate significant limitations in handling these scenarios [22].

Therefore, this study aims to design a field-road segmentation method specifically for residual film recovery operations. By extracting characteristic features from agricultural machinery trajectory points and analyzing their informational content, this study proposes combining cluster analysis with Multi-Feature Inference (MFI) rules. This approach enables the automatic and accurate segmentation of recovery trajectories into distinct field and road areas, ultimately achieving precise computation of operational area.

2. Materials and Methods

2.1. Experimental Equipment and Conditions

This study employed a custom-developed intelligent monitoring terminal for residual film recovery operations to acquire full-process trajectory data. The GNSS module of this terminal provides a positioning accuracy of 2.5 m (Figure 1a). During field trials, the terminal was installed on a traction-type residual film recycling machine (4JMLQ-210A, Changzhou Hansen Machinery Co., Ltd., Changzhou, China). Trajectory data were collected during actual recovery operations from October to December 2024 in Korla City, Xinjiang Uygur Autonomous Region, China (Figure 1b), yielding 620,267 trajectory points. Details of the experimental background were shown in

Table 1. Experimental background.

	Parameter	Value
GNSS Module	Positioning Accuracy	2.5 m
	Recording Frequency	1 Hz
	Protocol	NMEA-0183
	Collected Data Points	620,267
Machine	Residual Film Recycling Machine	4JMLQ-210A
Field Size	Area	35 ha

.

2.2. Overview

Based on the working characteristics of residual film recovery machinery, trajectory data demonstrate three distinct features: (1) Multiple discrete high-density trajectory clusters within field plots; (2) Strip-shaped high-density trajectory clusters at both headland ends; (3) Pronounced directional distribution patterns, where consecutive operational segments exhibit near-parallel alignment.

Based on these characteristics, this study proposes a field-road segmentation method specifically designed for residual film recovery machines to distinguish road segments from field segments in trajectories, as illustrated in Figure 2. The methodology comprises the following stages:

(1)

Data Preprocessing

Noise points undergo smoothing processing, with drift points filtered out to provide high-quality trajectory data for field-road segmentation.

(2)

Density-Based DBSCAN Clustering

Trajectory data are clustered using DBSCAN by setting parameters (e.g., Eps and MinPts), and classification labels are assigned to clustering results.

(3)

Application of Multi-Feature Inference Rules

Convex hull correction is applied to identify and complete low-density regions within field operation areas based on DBSCAN results. Subsequently, by combining the driving direction and speed characteristics of the field operation area in residual film recovery operations, the clustering results are further verified to check whether they conform to the typical characteristics of the field operation area, and the area is corrected based on the verification results.

2.3. Data Preprocessing

To validate the field-road segmentation method, trajectories were manually annotated using ground truth data, segmenting residual film recovery operations into distinct “field” and “road” sections based on actual operational records.

The following data processing procedures were implemented:

(1)

Noise Points Processing. Signal noise during GNSS positioning inevitably causes signal drift, generating significant localization errors. In order to reduce the impact of noise points, the noise points are detected based on the maximum traveling speed of the residual film recovery machine. If there is a large change in speed that far exceeds the maximum traveling speed of the residual film recovery machine, then remove the points [26].

(2)

Trajectory Smoothing. Discrete GNSS positioning points (2.5 m accuracy) introduce substantial errors in direction and velocity calculations. Kalman filtering was applied to smooth trajectories, reducing directional errors and velocity errors during residual film recovery operations [33].

(3)

Coordinate Transformation. Geodetic coordinates (WGS84) were converted to a two-dimensional Gauss-Krüger projection to enable accurate Euclidean distance calculations between points [34].

2.4. Density-Based DBSCAN Clustering

Residual film recovery operations exhibit distinct spatial distribution patterns where field areas demonstrate significantly higher positioning point density than road segments, thus justifying the selection of Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for clustering analysis of recovery trajectories [35]. Positioning point density in field areas substantially exceeds road areas during residual film recovery operations. This spatial distribution characteristic motivated our selection of DBSCAN for trajectory clustering analysis.

For the positioning trajectory dataset D = {x₁, x₂, …, x_m} from residual film recovery operations where each x_i ∈ R², the Eps-neighborhood defines a circular region centered at any sample x_i containing all points within this radius. A core object exists when a point x_i contains at least MinPts points within its Eps-neighborhood. A cluster M = {x_i, x_i₊₁, …, x_j} forms when density-reachable connections exist between points—specifically, when each consecutive point resides within the next point’s Eps-neighborhood. Clusters satisfying these density requirements are designated as “field clusters”.

Compared with the traditional K-Means algorithm, the most significant feature of the DBSCAN clustering method is that it does not require inputting the number of clusters k and can discover clusters of arbitrary shapes, which has great advantages in identifying the operation areas of unknown residual film recovery machines: it can maximize the discovery of field area cluster modules of arbitrary shapes and quantities [36]. After initially clustering using the density-based DBSCAN clustering method, the clustering results of residual film recovery operation trajectories may exhibit correct identification or misjudgment, mainly depending on the spatial distribution characteristics of the operation areas. The operation trajectories in field areas typically exhibit high-density characteristics, as shown in area 1 in Figure 3, while the road transition areas usually show low-density characteristics, as shown in area 3 in Figure 3. However, due to parking, changes in driving speed, etc., the road transition areas may also exhibit similar density levels, leading to DBSCAN potentially misclassifying road transition areas as field operation areas. As shown in area 4 in Figure 3, the road transition area is misclassified as a field operation area. Additionally, for strip-shaped operations in field areas, significant density differences in positioning data caused by positioning errors and other factors in some regions may also lead to misjudgment of certain trajectory points. As shown in area 2 in Figure 3, the field operation area is misclassified as a road transition area. Therefore, misjudgments primarily occur due to abnormal low-density regions within high-density areas and abnormal high-density regions within low-density areas. DBSCAN may fail to correctly distinguish these regions due to issues with the density threshold setting. During the initial clustering stage, we systematically tested various Eps and MinPts parameter combinations in DBSCAN to identify optimal values for field-road segmentation. As benchmarked in

Table 2. F1-score with different parameter settings of DBSCAN.

MinPts	Eps
	1	3	5	7	9
20	31.61	45.35	49.36	39.53	34.6
22	31.61	43.62	50.61	39.67	36.08
25	31.54	43.35	51.53	41.16	37.27
28	31.54	43.08	51.38	49.05	37.53
30	31.54	42.19	51.13	49.74	37.65

, the configuration Eps = 5 and MinPts = 25 achieved peak F1-score performance.

2.5. Multi-Feature Inference Method Based on Speed and Directional Features

To better separate the preliminary clustering results of field area clusters and road area clusters derived from the DBSCAN clustering method, a screening method based on directional and speed features is proposed. The key idea of this screening method is to utilize the operational characteristics of a field area: (1) The main working mode of the residual film recovery machine is manifested as linear operation based on two parallel directions; (2) A large amount of trajectory data also exists at both ends of the field where the machine operates: due to the characteristics of its own devices, the operational mode of the recovery machine in the field is limited, and it often performs residual film recovery operations according to the operational pattern shown in Figure 4. Based on the above operational trajectory characteristics of the residual film recovery machine, its speed and directional features are comprehensively used to screen field area clusters and road area clusters.

The screening based on speed and directional features consists of the following three steps:

(1)

Based on the known field area clusters, conduct statistical analysis on the driving directions and speed characteristics within the clusters.

(2)

Complete the field area clusters. According to the driving direction and driving speed feature within the clusters, complete the points within the field area clusters that are mistakenly regarded as road area clusters.

(3)

Screen the field area clusters after noise filtering and overall completion based on driving speed and directional features. Filter out clusters strictly concentrated in only two parallel directions and regard them as road area clusters; screen out field area clusters where the speed always remains below a certain threshold.

2.5.1. Measurement Method for Driving Direction and Speed

For the preprocessed data, the speed and direction of each positioning point are calculated based on the positioning data of the next point to obtain the driving speed and direction of the global positioning data.

The driving directions θ are divided into 18 directional region groups λ (θ) using a threshold of 20 degrees, as defined in Equation (1). All points in the field area clusters are subsequently traversed through Equation (2), with point count C (λ) being counted for each directional group. The top four groups exhibiting the highest point count are designated as the cluster’s primary directional groups. A parallel directional state is established when primary groups α and β satisfy |α − β| = 9 (denoting 180° opposition). As shown in Figure 5, the left side depicts the directional feature distribution of road area clusters misclassified as “field area clusters”, while the right side shows the directional feature distribution of correctly identified field area clusters. Field area clusters conform to directional characteristics when: (1) Two parallel states exist among primary directional groups. (2) The sum of points in non-primary groups is less than the total in primary groups.

$$\lambda (\theta )=⌊{\displaystyle \frac{\theta }{20°}}⌋,\theta \in \left[0,\right.\left.360°\right),\lambda \left(\theta \right)\in \left\{0,1,2,\dots ,17\right\}$$

(1)

$$C\left(\lambda \right)={\sum }_{i=1}^N\delta \left(\lambda \left({\theta }_i\right),\lambda \right),\delta \left(i,j\right)=\left\{\begin{array}{c}1\left(i=j\right)\\ 0\left(i\ne j\right)\end{array}\right.$$

(2)

N is the number of points within the field area cluster.

Figure 5. Example of directional feature distribution judgment for field area clusters. (a) A false field cluster: points spread out across multiple regions, each with a high point count; (b) a true field cluster: points mainly located in four regions with at least one pair of directionally parallel regions and fewer points in other regions.

Figure 5. Example of directional feature distribution judgment for field area clusters. (a) A false field cluster: points spread out across multiple regions, each with a high point count; (b) a true field cluster: points mainly located in four regions with at least one pair of directionally parallel regions and fewer points in other regions.

Since the calculation of driving speed υ is derived from positioning data, it is prone to significant noise, so the driving speed is subjected to Gaussian filtering by using Equation (3). Afterwards, the processing is similar to that of the driving direction. Using a threshold of 2 km/h, the speed υ is divided into nine groups μ (υ), as shown in Equation (4). By using Equation (5), all points in the field area clusters are traversed, and the number of points C (μ) in each of the nine regions is counted based on their speeds. As shown in Figure 6, the left side depicts the speed feature distribution of areas mistakenly identified as “field areas,” mainly concentrated in the 0–2 km/h range. This primarily results from positioning signal fluctuations when the residual film recovery machinery is stationary (on roads or fields), constrained by the positioning device’s accuracy. The right side illustrates the speed feature distribution of correctly identified field area clusters, with speed features mainly concentrated in the working speed range of 4–10 km/h, due to the characteristics of turning operations at both ends of the field. Notably, it also exhibits a secondary cluster in the 0–2 km/h range.

$$G\left(\upsilon \right)={\displaystyle \frac{1}{\sigma \sqrt{2\pi }}}{e}^{-{\displaystyle \frac{{\upsilon }^2}{2{\sigma }^2}}}$$

(3)

$\sigma$ is the standard deviation in Gaussian filtering, which mainly affects the smoothness of Gaussian filtering, primarily manifested in controlling the width of Gaussian filtering. [$\sigma$ = 0.5] is selected as the standard deviation for filtering.

$$\mu (\upsilon )=⌊{\displaystyle \frac{\upsilon }{2}}⌋,\upsilon \in \left[0,\right.\left.18\right),\mu \left(\upsilon \right)\in \left\{0,\mathrm{1,2},\dots ,8\right\}$$

(4)

$$C\left(\mu \right)={\sum }_{i=1}^N\delta \left(\mu \left({\upsilon }_i\right),\mu \right),\delta \left(i,j\right)=\left\{\begin{array}{c}1\left(i=j\right)\\ 0\left(i\ne j\right)\end{array}\right.$$

(5)

N is the number of points within the field area cluster.

2.5.2. Convex Hull Completion Based on Driving Direction Features

Due to issues such as inaccurate GPS positioning accuracy and large operation width, some points of the residual film recovery machine in the field may become relatively sparse with low point density and are classified as “road area clusters” during the initial application of the DBSCAN clustering algorithm. As shown in Figure 7, some low-density regions within the “field area clusters” are mistakenly identified as “road area clusters”.

Using the driving direction and speed features of in-field operations, the following detections are performed on the “road area clusters” located within the “field area clusters”.

(1)

Whether the “road area cluster” is within the corresponding “field area cluster”.

(2)

Whether the driving direction features of the “road area cluster” are consistent with the main directional regions of the “field area cluster”.

(3)

Whether the convex hull containing the “road area cluster” collides with other convex hulls.

Algorithm 1 is as follows:

Algorithm 1: Convex Hull Expansion Classification Algorithm

As shown in Figure 7, it is the diagram of field boundary completion. Based on the feature that both ends of the field areas have higher point density, the convex hull completion measure is implemented for positioning points in partial low-density areas within the field, which are caused by positioning errors of positioning devices, operation modes, etc.

2.5.3. Screening Based on Driving Speed and Directional Features

After preliminary density-based DBSCAN clustering, “field area clusters” with high point density are obtained. However, due to machine entry/exit from roads, turning, deceleration, and other situations, positioning points may also exhibit high density, leading to potential misidentification of “road area clusters” as “field area clusters” during preliminary clustering. Taking Figure 8 and Figure 9 as examples, Figure 8 shows the misidentification of road areas as “field area clusters,” while Figure 9 demonstrates the misidentification of road areas as “field area clusters” due to their high point density and directional feature similar to field areas.

The following rule-based screening is applied to their operation trajectory segmentation.

(1)

Detect the total number of points in field area clusters: determine the point threshold based on the minimum operation time of the residual film recovery machine, traverse the “field area clusters,” screen out those with point counts below the threshold, and classify them as “road area clusters.”

(2)

Detect the driving speed features of field area clusters: statistically analyze the driving speeds of “field area clusters” according to the speed processing method mentioned in Section 2.5.1. In the “field area clusters,” if the proportion of low driving speeds in the overall driving speed features is large, classify them as “road area clusters.” As shown in the speed feature distribution of correctly and mistakenly identified field area clusters in Figure 3, the driving speed features of correct “field area clusters” are concentrated in the working speed range, while those of mistaken “field area clusters” are concentrated in the low driving speed range of 0–1 km/h.

(3)

Detect the directional distribution features of field area clusters: Different from common agricultural machinery working modes, due to the special working mode of the residual film recovery machine, it needs to repeatedly drive in a certain directional region at both ends of the field. Therefore, based on the premise that the driving directions of the main operation area are parallel, there are also two directional regions with higher point counts.

Algorithm 2 is as follows:

Algorithm 2: Field Area Verification

The detection rules for driving direction distribution are as follows.

(1)

Statistically analyze the point distribution of “field area clusters” in 36 directional regions, and select the top four directional region groups with the highest point counts.

(2)

Detect the selected top four directional region groups to ensure that at least one group consists of directionally parallel regions with similar point counts.

2.6. Evaluation Metrics

2.6.1. Precision, Recall, and F1-Score

The density-based DBSCAN clustering algorithm is an unsupervised learning algorithm in machine learning. When evaluating the performance of this algorithm, precision, recall, and F1-score are commonly used evaluation metrics [37]. In the classification problem of evaluation metrics, the field area is defined as the “positive” class, and the road area is defined as the “negative” class. TP represents the number of positioning points in the field area that are correctly classified; FP represents the number of positioning points in the field area that are mistakenly classified as the road area; FN represents the number of positioning points in the road area that are mistakenly classified as the field area; TN represents the number of positioning points in the road area that are correctly classified. According to Equations (6)–(8), calculate the precision, recall, and F1-score for field-road trajectory segmentation, and evaluate the model performance.

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(6)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(7)

$$\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2\times {\displaystyle \frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

(8)

2.6.2. Operational Area

The segmented operational area is used for assessment to better assess the impact of the proposed method on field-road trajectory segmentation. The area evaluation metrics include the correct segmentation area, over-segmentation area, and under-segmentation area.

(1)

Correct Segmentation Area. The area of the actual field regions that is correctly segmented as fields.

(2)

Over-segmentation Area. The area of actual road regions that is mistakenly segmented as fields (i.e., road areas misclassified as fields).

(3)

Under-segmentation Area. The area of actual field regions that is mistakenly segmented as roads (i.e., field areas misclassified as roads).

The simplest method for area calculation is using the product of operational trajectory and width to compute the field area. However, due to the large number of repeated paths at both ends of the field during residual film recovery operations, using only the trajectory × width method would include repeated operational areas in the total. Therefore, the buffer zone generation method is adopted: taking the operational trajectory as the base, establishing an operational buffer with the operation width ω as the buffer zone width, and calculating the overall operational area by using the polygonal area calculation formula (Equation (9)) [38].

$$S={\displaystyle \frac{1}{2}}{\sum }_{\lambda =1}^{\infty }(x_{\lambda }{y}_{\lambda +1}-x_{\lambda +1}{y}_{\lambda })$$

(9)

3. Results and Discussion

3.1. Segmentation Results

The trajectories of residual film recovery operations and the identification results of different algorithms for field operation areas and road transition areas in the recovery operation trajectories are shown in Figure 10.

By comparing with the classification results of the DBSCAN algorithm (

Table 3. The overall trajectory sample confusion matrix.

Method	Trajectory Type	Number of Identified Field	Number of Identified Road
DBSCAN	Field	433,549	40,023
	Road	82,777	63,918
DBSCAN-MFI	Field	466,413	7,159
	Road	45,968	100,727
OPTICS	Field	429,574	43,998
	Road	79,046	67,649
Grid-Based Method	Field	439,808	33,764
	Road	94,030	52,665
DBSCAN-FR	Field	459,436	14,136
	Road	84,160	62,535

), the algorithm in this study successfully identified 466,413 positioning points in field areas, an increase of 7.6% compared with DBSCAN, which identified 433,549 positioning points; meanwhile, the number of positioning points misjudged as road areas decreased from 40,023 to 7159, a reduction of 82%. This improvement is reflected in the under-segmentation area in Figure 7a and

Table 4. The overall segmented area calculation results of different methods.

Method	Area Accuracy (%)	Correct Segmentation Area (ha)	Over-Segmentation Area (ha)	Under-Segmentation Area (ha)
DBSCAN	71.74	25.11	1.03	9.24
DBSCAN-MFI	95.33	33.36	0.88	1.27
OPTICS	70.19	24.56	0.75	9.75
Grid-Based	74.62	26.12	1.57	8.13
DBSCAN-FR	88.47	30.96	1.27	3.57

, which mainly focuses on the positioning point data in field areas.

The misjudgments of DBSCAN mainly originate in GNSS positioning signal errors during residual film recovery machine operations and the problem of non-uniform density within field areas. To address this issue, the algorithm proposed in this study introduces inference rules based on DBSCAN, accurately distinguishing field operation areas from road transition areas by combining speed and directional features of different clusters, overcoming the limitation of traditional methods relying solely on density. DBSCAN-MFI introduces a speed-direction feature inference mechanism in global region recognition. The direction feature inference is determined by the heading angle groups, which are obtained by dividing the heading angles according to manually set intervals. It requires the presence of two parallel states in the overall directional groups of the field area. The speed feature inference is determined through speed grouping, which divides the heading speed in the same way. It requires the existence of low-speed points, but not exceeding a certain proportion. At the same time, DBSCAN-MFI introduces a convex hull completion strategy in the recognition of internal areas in the field.

From the overall performance comparison (

Table 5. The overall performance of different methods.

Method	Precision (%)	Recall (%)	F1-Score (%)
DBSCAN	91.55	83.96	87.59
DBSCAN-MFI	98.48	91.02	94.60
OPTICS [39]	90.71	84.46	87.47
Grid-Based [40]	92.87	82.39	87.32
DBSCAN-FR [22]	97.02	84.52	90.33
ATRNet [41]	88.44	97.43	92.72
GCN [23]	88.72	98.81	93.49

), it can be seen that the introduction of inference rules significantly improves the relevant machine learning metrics of the DBSCAN clustering algorithm. The algorithm in this study improves the precision by 6.87% compared to the DBSCAN algorithm, which means that the proportion of correct predictions of positioning points in field areas by this algorithm has increased; the recall has increased by 7.06%, indicating that the algorithm can better cover the positioning points in actual field areas; and the F1-score has also increased by 7.01%, reflecting the overall performance optimization of the algorithm. Meanwhile, in terms of the F1-score, the DBSCAN-MFI algorithm increases by 7.13%, 7.28%, and 4.27% compared to the OPTICS algorithm, the Grid-Based algorithm, and the DBSCAN-FR algorithm, respectively.

The precision of the DBSCAN-FR algorithm is similar to that of DBSCAN-MFI, differing by only 1.46%, but the recall and F1-score are significantly lower than those of the DBSCAN-MFI algorithm, indicating that although the DBSCAN-FR algorithm has a similar capability in identifying field operation areas to DBSCAN-MFI, its overall performance is still insufficient. The OPTICS algorithm has a precision of 90.71%, a recall of 84.46%, and an F1-score of 87.47%, which are similar to the results of the DBSCAN algorithm. Although the Grid-Based algorithm has a precision of 92.87%, its recall and F1-score are lower, 82.39% and 87.32%, respectively, and it cannot identify low-density regions in field operation areas. The precision of DBSCAN-MFI is 10.04% and 9.76% higher than that of ATRNet and GCN. While the proposed algorithm exhibits 6.44% and 7.79% lower recall than these methods, it demonstrates superior F1-scores. Higher recall would enable identification of more actual field operation areas, but at the cost of increased misidentifications. Such false positives would adversely impact subsequent area calculations by introducing substantial over-segmentation.

To evaluate the computational complexity of the proposed algorithm, experiments were conducted using trajectory point datasets with varying sample sizes, where the algorithm’s runtime performance was measured. A comparative analysis was then performed against existing DBSCAN-based methods, with the quantitative results presented in Figure 11. It can be seen from Figure 11 that the proposed method maintains comparable computational efficiency to DBSCAN-based approaches, with only marginal increases in computational overhead. Processing datasets on the order of 10⁶ requires approximately 1000 s using our method. By comparison, Zhai et al. [41] report processing times of 4718 s (GCN) and 1650 s (ATRNet) for equivalently sized datasets, demonstrating the proposed algorithm’s significant computational efficiency advantage.

3.2. Operational Area Calculation

In terms of the impact on operational area calculation, the DBSCAN-MFI algorithm demonstrates significant advantages in handling the segmentation of operational areas, especially in the control of correct segmentation area and under-segmentation area, outperforming other algorithms.

The total working area for this collected trajectory is 35 ha. The correct segmentation area of the DBSCAN-MFI algorithm is 33.36 ha, which is an increase of 23.61%, 25.14%, 20.71%, and 6.87% compared to the DBSCAN algorithm, OPTICS algorithm, Grid-Based algorithm, and DBSCAN-FR algorithm, respectively. This gap indicates that the DBSCAN-MFI algorithm can more accurately identify field operation areas in residual film recovery operation trajectories, especially under conditions of non-uniform density. The DBSCAN-MFI algorithm can effectively correct misjudgments in low-density regions within field operation areas through the convex hull algorithm, thereby enhancing the ability to distinguish between field and road areas and reducing the proportion of under-segmentation area. In contrast, other algorithms, especially DBSCAN and OPTICS, although performing well in terms of over-segmentation area, show poor performance in the correct segmentation of field operation areas, indicating that they adopt a more conservative identification strategy in regions with large density differences, focusing only on identifying parts with optimal density features.

Compared with other algorithms, DBSCAN-MFI demonstrates better segmentation capability, especially in handling non-uniform density in field operation areas and identifying road transition areas and field operation areas. This is mainly attributed to the introduction of inference rules, which combine speed and directional features with completion measures for density differences in field areas. Compared with DBSCAN, OPTICS, and Grid-Based algorithms, it effectively optimizes the distinction between fields and roads, reducing over-segmentation and under-segmentation issues. Compared with the DBSCAN-FR algorithm, it uses the convex hull algorithm to focus on optimizing the identification of low-density regions within field operation areas, significantly reducing the area of under-segmentation.

3.3. Discussion

The trajectory dataset generated by residual film recovery machinery contains multiple discrete high-density clusters. This presents new challenges for density-based clustering methods, as discrete misclassified areas may occur within fields. While existing studies predominantly focus on trajectory density non-uniformity between fields and roads, solutions for intra-field misclassification remain unexplored.

To address this, we introduce a completion algorithm based on convex hull to locate misclassified intra-field areas. However, limitations persist:

(1)

Parameters derived from residual film recovery machinery’s operational characteristics improve segmentation accuracy but fail to resolve the inherent algorithmic sensitivity problem. Furthermore, the impact of GPS noise and signal loss requires additional verification.

(2)

Machinery-specific knowledge enhances accuracy but exhibits regional/crop dependencies, limiting immediate transferability to other equipment without adaptation.

Notably, increasing intelligence in agricultural machinery enables adaptive algorithm optimization by acquiring and mining equipment-specific operational features. Current research already demonstrates this paradigm, such as trajectory classification methods developed for efficient field-road segmentation, validating the approach’s technical feasibility [26].

4. Conclusions

Due to the presence of multiple discrete high-density trajectory clusters in residual film recovery fields, strip-shaped terminal clusters, and near-parallel continuous operational segments in the field, conventional density-based clustering algorithms exhibit suboptimal performance in field-road segmentation. To address this limitation, this study proposes DBSCAN-MFI, a trajectory segmentation method integrating DBSCAN density clustering with multi-feature inference, to identify field operation areas in residual film recovery.

Based on four trajectory types derived from DBSCAN clustering, misclassified data points in field areas are first retrieved using convex hull algorithms. These points are then rectified through feature inference leveraging driving speed and direction characteristics. To validate the optimization of DBSCAN by multi-feature inference rules, field trajectories were collected by using a custom-developed intelligent monitoring terminal during residual film recovery operations in Korla, Xinjiang Uygur Autonomous Region, China. Experimental results demonstrate that compared to conventional DBSCAN, OPTICS, Grid-Based Methods, and DBSCAN-FR clustering, the DBSCAN-MFI algorithm achieves F1-score improvements of 7.01, 7.13, 7.28, and 4.27 percentage points, respectively, while increasing correctly segmented operational areas by 23.61%, 25.14%, 20.71%, and 6.87%.

The enhanced DBSCAN-MFI algorithm enables effective field-road segmentation for residual film mechanical recovery. Given the non-uniform distribution of trajectory data points, real-time computation faces challenges as trajectory density may vary during the operational phase. Calculating at intervals of multiple days/weeks can reduce the impact of natural density changes. Utilizing edge devices can alleviate sudden computing loads and avoid intensive short-term data processing.

Future work will validate the algorithm’s generalizability across diverse geographical regions and crop types. Specifically, we will investigate machine learning approaches to extract equipment-specific operational features from agricultural machinery trajectory data for adaptive algorithm optimization. For agricultural practitioners, we recommend that operators actively adopt high-precision and reliable positioning devices and maintain stable operating speeds to avoid non-uniform distribution of trajectory point data.