Evaluating the Reliability of Remote Sensing Techniques for Detecting the Strip Road Network in Boom-Corridor Systems

Abstract

Accurate detection of machinery-induced strip roads after forest operations is fundamental for assessing soil disturbance and supporting sustainable forest management. However, in Mediterranean pine forests where canopy openings after boom-corridor thinning are moderate, the effectiveness of different remote sensing techniques remains uncertain. Previous studies have shown that LiDAR-based methods can reliably detect logging trails in different forest stands, but their direct transfer to structurally simpler, even-aged Mediterranean stands has not been validated. This study addresses this gap by testing whether UAV-derived RGB imagery can achieve comparable accuracy to LiDAR-based methods under the canopy conditions of boom-corridor thinning. We compared four approaches for detecting strip roads in a black pine (Pinus nigra Arn.) plantation on Mount Amiata (Tuscany, Italy): one based on high-resolution UAV RGB imagery and three based on LiDAR data, namely Hillshading (Hill), Local Relief Model (LRM), and Relative Density Model (RDM). The RDM method was specifically adapted to Mediterranean conditions by redefining its return-density height interval (1–30 cm) to better capture areas of bare soil typical of recently trafficked strip roads. Accuracy was evaluated against a GNSS-derived control map using nine performance metrics and a balanced subsampling framework with bootstrapped confidence intervals and ANOVA-based statistical comparisons. Results confirmed that UAV-RGB imagery provides reliable detection of strip roads under moderate canopy openings (accuracy = 0.64, Kappa = 0.27), while the parameter-tuned RDM achieved the highest accuracy and recall (accuracy = 0.75, Kappa = 0.49). This study demonstrates that RGB-based mapping can serve as a cost-effective solution for operational monitoring, while a properly tuned RDM provides the most robust performance when computational resources are sufficient to work on large point clouds. By adapting the RDM to Mediterranean forest conditions and validating the effectiveness of low-cost UAV-RGB surveys, this study bridges a key methodological gap in post-harvest disturbance mapping, offering forest managers practical, scalable tools to monitor soil impacts and support sustainable mechanized harvesting.

1. Introduction

Machinery traffic related to ground-based forest operations primarily affect the soil through compaction, rutting, and displacement [1,2,3]. Compaction occurs when soil particles are pushed closer together, reducing pore space due to mechanical forces [4,5,6]. Repeated passes by machinery can deform the soil, especially when the soil water content is high [2,7,8]. Ground-based traffic from forest machinery can negatively impact essential soil functions by increasing soil strength and bulk density [9,10,11], decreasing water infiltration and saturated hydraulic conductivity [12], and reducing the biodiversity of soil biota [13,14]. These impacts can lead to reduced oxygen availability, increased surface water flow, soil loss, and sedimentation [3]. Ultimately, these adverse effects are likely to hinder tree and root growth [15], negatively affect site productivity [16], and potentially alter soil ecological processes such as litter decomposition rates [17].

Machine traffic while performing forest operation is concentrated along the trails [18,19,20]. These trails are usually named skid trails when the extraction operation implies the contact of the logs with the ground (skidding), while when using cut-to-length machinery (harvester + forwarder) the term strip roads is used [21,22]. The identification of the skid trail or strip road network established after forest operations is crucial for both monitoring and planning issues [23,24,25]. Knowing the extent and localization of the trails is the first step to verify if the forest operations were performed in agreement with the regulation requirements [26]. Furthermore, identifying the established trails and mapping them allows for making the skid trail/strip road network fixed, thus making possible to re-use the same trails in the following logging entry and thus limiting the soil surface affected by vehicle movement [18]. Traditionally, the area affected by machine traffic is assessed by post-harvesting field surveys through tracing the visible trails [23]. This method is obviously extremely time and manpower consuming.

Modern technologies in forestry and remote sensing have deeply changed our possibilities to effectively identifying the areas affected by machine traffic [27]. Modern forest machines are equipped with on-board computers, sensors and GNSS (Global Navigation Satellite Systems) which allow for collecting a plethora of information about the forest operations on-time, including the position of the machinery [28]. In this way we can know exactly where the machine trafficked within a given cutting block [29]. However, this applies only when modern machinery is used, which is often not the case in many parts of Europe, including the Mediterranean area [30,31,32,33]. Furthermore, in many Countries in Europe including Italy, the major part of the forests is private. Thus, without a dedicated regulatory framework which is currently missing, it is up to the owner if they want to share the data collected by the forest machinery with the controlling entities such as the local forest service for facilitating the post-harvesting monitoring.

To identify the machine trails after harvesting in such cases, we can still rely on remote sensing technologies [25,34,35]. In this context, both Unmanned Aerial Vehicles (UAVs) and aircraft-based platforms (ALS) offer significant advantages, enabling rapid and extensive surveying compared to traditional ground-based methods [36]. These platforms can be equipped with either passive sensors (e.g., RGB cameras) or active sensors (e.g., LiDAR—Light Detection and Ranging), each suited to specific forest conditions for detecting soil disturbance caused by logging machinery. In areas where canopy openings are substantial, such as after clear-cutting or coppicing, passive sensors are generally effective in capturing the trail network, as the visibility of the ground surface is sufficient [37]. Conversely, in cases where canopy cover remains largely intact, such as after thinning or single-tree selection harvesting, active sensors like LiDAR are preferred due to their ability to penetrate the canopy and provide detailed information about the forest floor [38,39].

While airborne laser scanning (ALS) is well suited for large-scale surveys and is typically used to detect major forest road networks [24,40], effective monitoring at the forest-management-unit level, especially shortly after harvesting, requires the use of UAV-based surveys [27]. In any case, the reliability of LiDAR for delineating trail networks is well established, but both the acquisition of LiDAR data (as a consequence of high hardware cost) and their processing are substantially more costly, frequently by an order of magnitude, than surveys relying on passive sensors such as RGB cameras, and they typically require more complex analytical workflows. [41].

In the Mediterranean region, particularly in Central and Southern Italy, there is an extensive area of artificial coniferous plantations [42], predominantly composed of black pine (Pinus nigra Arn.), which were established during the first half of the 20th century [43]. Due to the generally low timber quality, partly resulting from the absence of early thinning interventions, wood harvested from these stands is typically suitable only for fuelwood production [44]. In recent years, the introduction of fully mechanized cut-to-length harvesting systems, involving harvesters and forwarders, has significantly increased the efficiency of operations in these pine forests. Additionally, the adoption of alternative thinning approaches, such as boom-corridor thinning, where trees are removed along straight corridors within the reach of the harvester’s boom, rather than through selective removal as in traditional thinning from below, has further enhanced harvesting rates and operational productivity [45,46]. Given the large size of the machinery used and the frequent presence of steep terrain, with slopes typically ranging from 30% to 40%, it is essential to monitor soil disturbance caused by forest operations in these pine stands. Addressing this issue requires careful consideration of canopy alterations resulting from the applied silvicultural practices. In particular, the boom-corridor thinning technique produces canopy openings that are less pronounced than those from clear-cutting, but still more distinct than those resulting from conventional selective thinning. This raises the question of whether UAV-based RGB imagery is sufficient for detecting the network of strip roads created during such operations, or if LiDAR-based surveys, with their superior canopy penetration capabilities, are necessary.

To explore this, we conducted a study in a black pine plantation located on Mount Amiata in Tuscany, Italy, following forest operations carried out using the boom-corridor thinning method. We performed both RGB and LiDAR surveys using UAV platforms, and compared the outputs to test the hypothesis UAV-RGB imagery can achieve comparable accuracy to LiDAR-based methods under moderate canopy openings created by boom-corridor thinning.

2. Materials and Methods

2.1. Study Area

The study area (Figure 1), corresponding to one black pine cutting block, is located in the municipality of Arcidosso (Grosseto district, Tuscany, Italy). This logging site was selected because it was highly representative of black pine forests of Central Italy, in terms of biomass stock, topography and applied harvesting system. The surface of the cutting block was 10.65 ha. In the study area the mean annual temperature is 9.3 °C with a mean annual precipitation of 708 mm [47]. The soil is a Cambisol with a loam texture [48]. The elevation of the cutting block ranged from 1119 to 1200 m a.s.l. with a mean elevation of 1150 m a.s.l. Prevalent terrain slope was 28%, but the topography of the cutting block was quite heterogeneous, ranging from almost flat zones to areas at slope higher than 50%.

The forest stand consisted of an even-aged black pine forest of about 60 years which was never thinned before. Along with the pines there were spots of chestnut (Castanea sativa Mill.) which were not subject to harvesting. Therefore, the actual intervention surface was equal to 7.85 ha. Standing volume before harvesting was 410 m³ ha⁻¹ and the harvested volume was 35% of the standing one (about 140 m³ ha⁻¹). Forest operations were carried out by the cut-to-length system with an excavator equipped with a processing head for felling and processing, and an 810D forwarder for timber extraction.

2.2. Field and UAV Surveys

Field surveys were carried out at the end of 2024, three months after the end of the forest operations, which were performed in September 2024, where the soil moisture was about 30%–35%. We performed a ground-based survey by a handheld GNSS receiver by tracing and recording all the strip roads established by the forwarder (largely corresponding with those of the felling excavator as usual in cut-to-length operations). This strip roads dataset represented our control strip road network (Figure 2). The GNSS data were collected using a Spectra Precision SP20 handheld receiver (Spectra Geospatial, Westminster, CO, USA) operating in DGPS mode, which applies real-time satellite-based differential corrections (SBAS). Under these conditions, the SP20 provides a typical horizontal positional accuracy of approximately 30–50 cm. No post-processing or RTK corrections were applied.

The day after the GNSS survey, we performed the RGB-UAV survey with a DJI Air 2S (SZ DJI Technology Co., Ltd., Shenzen, China) UAV equipped with a 1” CMOS” image sensor with FOV 88° and 5.4 K resolution. Flight altitude for the UAV-RGB survey was 80 m, with 80% overlap and GSD (Ground Sampling Distance) of 2.3 cm per pixel. One week later we outsourced a professional company to perform a LiDAR-UAV survey by a DJI Matrice 300 equipped with a 5-return LiDAR DJI L2 sensor (SZ DJI Technology Co., Ltd., Shenzen, China). The main flight settings included an altitude of 80 m above ground level, a speed of 6.5 m s⁻¹, a scan angle of ±25° from nadir, a front overlap of about 70%, and a side overlap of about 40%. The density of the obtained LiDAR point cloud was about 600 points per square meter. The percentage of ground points in relation to the total number of points was about 40%. Point cloud density was 686 N m⁻² in non-strip roads area and 531 N m⁻² on the strip roads.

2.3. Detection of the Strip Roads

We identified the strip road network using four distinct methods (Figure 3): one based on RGB data and three derived from LiDAR data. Among the LiDAR-based methods, one operates directly on the point cloud, while the other two require the generation of a 0.5 m resolution Digital Elevation Model (DEM) from the point cloud as a preliminary step.

The method based on passive remote sensing (RGB) involved manually digitizing the strip roads visible in the high-resolution orthophoto produced from the UAV survey. This approach leverages the canopy openings resulting from harvesting activities, which are evident in the RGB imagery, to delineate the trail network.

The first of the LiDAR-based methods is the Hillshading method (Hill). Hillshading enhances terrain visualization by simulating the effects of light and shadow from a defined illumination source [49]. Specifically, The Hillshading technique is a terrain visualization model based on the geometric relationship between light source direction and surface orientation. It simulates illumination across a digital terrain model (DTM) by computing the cosine of the angle between the illumination vector (defined by azimuth and altitude) and the surface normal. The resulting grayscale image enhances relief by representing illuminated slopes as bright and shadowed slopes as dark. In this context, skid trails appear as elongated depressions or flattened surfaces that generate characteristic shading contrasts. Hillshading does not alter elevation values but increases visual interpretability by emphasizing microtopographic variation. This technique, widely available in GIS software (QGIS software version 3.40) under the “hillshade” function, creates a 3D-like effect that accentuates topographic variations. For this method, the DEM is first processed using hillshading, after which the strip roads are manually digitized from the resulting image.

The second method, the Local Relief Model (LRM), enhances the visibility of small-scale terrain features by removing large-scale topographic patterns. The Local Relief Model isolates small-scale topographic variation by removing large-scale terrain trends. The theoretical foundation of LRM lies in spatial frequency filtering. A low-pass filter (typically Gaussian) is applied to the DTM, acquired after the establishment of the strip roads, to capture broad terrain forms such as ridges and valleys. This smoothed DTM is then subtracted from the original DEM, producing a residual model that highlights local elevation deviations. Positive residuals indicate small elevations (e.g., logs, mounds), while negative ones represent depressions such as strip roads. Thus, LRM enhances subtle geomorphological features that are otherwise obscured by the dominant terrain shape. This technique has proven particularly effective for identifying linear depressions indicative of strip roads [38,50].

The third method, the Relative Density Model (RDM) [51], differs from the previous two by using raw LiDAR point cloud data rather than a DEM. The Relative Density Model is based on the vertical distribution of LiDAR returns within the point cloud. The theoretical assumption is that areas affected by machine traffic exhibit reduced vegetation density and fewer LiDAR returns in the lower height strata. RDM quantifies this by calculating the ratio between the number of LiDAR returns within a specified height range (e.g., 1–5 m) and the total number of returns within each grid cell. Lower RDM values correspond to disturbed zones or bare soil, while higher values indicate undisturbed forest floor or dense understory. The model therefore acts as a proxy for ground-cover continuity and vegetation structural integrity, making it particularly effective for detecting narrow linear disturbances such as strip roads. When visualized as a raster, strip roads appear as narrow, linear features with noticeably low RDM values, distinguishing them from surrounding undisturbed forest.

While the Hillshading method is relatively straightforward and readily applicable, both LRM and RDM required specific adaptations to account for the structural characteristics of Mediterranean forests. For LRM, two key parameters of the Gaussian filter must be carefully chosen: the kernel size and the standard deviation. The kernel size determines the spatial extent of smoothing, while the standard deviation controls its intensity. These parameters must be tailored to suppress large-scale topography (e.g., ridges and gullies) while retaining minor terrain features like strip roads. Based on the scale of the forestry machinery used and the typical dimensions of the resulting trails, we selected a kernel size of 7 m and a standard deviation of 2 m. A comparative view of strip road detection using different kernel sizes and standard deviations is provided in Figure S1 of Supplementary Materials. As shown, the selected parameters enable the most effective identification of the strip roads.

Regarding the RDM, the method was originally developed for tropical rainforests characterized by complex, multi-layered canopies. In those contexts, the commonly used height interval for return density calculation is 1–5 m [51]. However, this range is not appropriate for recently harvested, even-aged Mediterranean pine stands. To select a more suitable interval, we considered the appearance of strip roads in these forests (see Figure 3 RGB). Post-harvest, vegetation is largely absent from the trail surface, and litter cover is significantly reduced. To reflect this, we modified the RDM calculation to assess return density in the 1–30 cm height range (bare soil or only herbaceous cover), aiming to highlight areas of bare soil associated with vehicle tracks. A comparative view of strip road detection using different RDM parameters is provided in Figure S2 of Supplementary Materials. As shown, the chosen parameters proved optimal for achieving clear and reliable identification of the strip roads.

In both the LRM and RDM methods, once the respective rasters were generated, strip roads were manually digitized based on the enhanced visualizations provided by each model. To ensure the robustness and reproducibility of the interpretation, the digitization procedure was independently performed by four different operators. The resulting four shapefiles were then merged to obtain a consensus strip road pattern. Specifically, the final strip road network was defined by overlaying all individual digitizations and retaining only those trail segments that were identified by at least three out of the four operators. This majority-based approach minimizes subjectivity and operator bias, ensuring that the final map represents the most consistently interpreted features across multiple observers. A vision of the same portion of the cutting block with the rasters obtained with the four methods is given in Figure 3.

2.4. Data Analysis

To assess the accuracy of different methods for identifying strip roads, we compared four classification approaches, RGB, RDM, LRM, and Hill, against a reference dataset obtained through a GNSS survey. The classification results and reference data were represented as binary raster layers obtained from the rasterization of the strip roads shapefile for each method. The shapefile of the strip roads was rasterized with a pixel of 4 m resolution (about the width of the strip roads), where pixel values of 1 indicated the presence of strip roads, and 0 indicated their absence. The entire analysis was performed using QGIS 3.40 [52] for data preprocessing and Rstudio (R version 4.3.3) [53] for accuracy assessment and statistical analysis. Because the number of trail pixels was considerably smaller than the number of background pixels, the dataset exhibited strong class imbalance. To mitigate the bias that such imbalance introduces in global accuracy measures, we adopted a balanced case–control evaluation. For each method, all positive pixels (strip roads pixels identified in the control layer) were retained, while an equal number of background (non-strip roads) pixels were randomly sampled from the logging site area. This balanced subsampling was repeated 1000 times, ensuring robust estimation of performance metrics while maintaining equal representation of classes in each iteration. Each iteration therefore consisted of a balanced dataset composed of all trail pixels and an equal number of randomly sampled background pixels, on which the accuracy metrics were recalculated.

The confusion matrix for each classification method was computed by comparing each method’s raster against the control raster. From these confusion matrices, we derived nine accuracy metrics: overall accuracy, precision, recall, Cohen’s kappa, Intersection over Union (IoU), Dice Similarity Coefficient (DSC), true positive rate (TPR), false positive rate (FPR), and specificity.

Each metric (all dimensionless) was calculated as follows:

• Overall Accuracy = (TP + TN)/(TP + TN + FP + FN). Represents the overall proportion of correctly classified pixels (both strip roads and non-strip roads). It gives a general sense of model performance but can be misleading when classes are imbalanced.
• Precision = TP/(TP + FP). Indicates how many of the pixels predicted as strip roads were actually strip roads. High precision means fewer false positives (non-strip roads areas wrongly classified as strip roads).
• Recall (Sensitivity, TPR) = TP/(TP + FN). Measures the proportion of actual strip roads pixels that were correctly identified. High recall means fewer false negatives (missed strip roads pixels).
• Cohen’s Kappa = (Observed Accuracy − Expected Accuracy)/(1 − Expected Accuracy). A statistic that adjusts accuracy for agreement that occurs by chance. Values range from −1 (complete disagreement) to 1 (perfect agreement). A value of 0 implies random classification.
• Intersection over Union (IoU) = TP/(TP + FP + FN). Indicates the overlap between predicted and actual strip roads areas. A higher IoU means a greater spatial match between method output and control.
• Dice Similarity Coefficient (DSC) = (2 × TP)/(2 × TP + FP + FN). Another measure of spatial overlap. Dice is more sensitive to small overlaps and often used in image segmentation tasks.
• False Positive Rate (FPR) = FP/(FP + TN). Represents the proportion of non-strip roads areas incorrectly predicted as strip roads. Lower FPR values indicate better performance.
• Specificity = TN/(TN + FP). Reflects the ability of a method to correctly identify non-strip roads areas. High specificity complements recall by ensuring that the background (non-strip roads) areas are not falsely marked.

The raster values were extracted and compared using logical operations to derive true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN). Each of these metrics captures a different aspect of classification performance. While accuracy provides a general view, precision and recall offer insight into the balance between omission and commission errors. Cohen’s Kappa ensures that agreement is not due to chance, and spatial metrics like IoU and DSC quantify the geometric match of trail shapes between prediction and reference.

To quantify variability and assess statistical differences among methods, we applied a non-parametric bootstrap procedure based on the balanced subsampling results. Each of the 1000 balanced iterations yielded a value for each accuracy metric, forming a bootstrap distribution per method and per metric. For each distribution, we computed the mean and 95% confidence intervals (CIs) using the 2.5th and 97.5th percentiles. To test for statistically significant differences among methods, a one-way ANOVA was performed separately for each metric, followed by Tukey’s Honest Significant Difference (HSD) post hoc test.

The computations were carried out in R version 4.4.3 using the terra [54], ggplot2 [55], and reshape2 [56] packages.

3. Results

The length of the identified strip road network and the related percentage of soil surface affected by machine traffic per each investigated method is reported in

Table 1. Identified strip road network per each method and control dataset. Percentage of impacted surface calculated considering the actual harvested surface of 7.85 ha and an average strip road width of 3.8 m.

Method	Strip Road Network Length (m)	Impacted Surface (%)
Control	2981	14.43
RGB	1552	7.51
RDM	2621	12.68
LRM	1636	7.91
Hill	1056	5.11

, while the visual representation of the strip roads detected with the various methods is available in Figures S3–S6 in Supplementary Materials. Only the RDM method estimated an impacted surface similar to the actual one, while the other three methods strongly underestimated the actual surface affected by machine traffic.

Below,

Table 2. Accuracy metrics and results of the statistical comparison. Abbreviations: CI: confidence interval; IoU: Intersection over Union; DSC: Dice Score; TPR: True Positive Rate; FPR: False Positive Rate. Different lowercase letters in the column “Tukey group” indicate statistically significant differences at p < 0.05.

Metric	Method	Mean	CI_Low	CI_High	Tukey Group
Accuracy	RDM	0.746	0.744	0.747	a
	RGB	0.636	0.634	0.636	b
	LRM	0.635	0.633	0.636	c
	Hill	0.609	0.608	0.609	d
Kappa	RDM	0.491	0.488	0.493	a
	RGB	0.271	0.267	0.273	b
	LRM	0.270	0.266	0.271	c
	Hill	0.217	0.215	0.218	d
Precision	RDM	0.996	0.989	1.000	b
	RGB	0.994	0.981	1.000	c
	LRM	0.994	0.981	1.000	c
	Hill	0.997	0.988	1.000	a
Recall	RDM	0.493	0.493	0.493	a
	RGB	0.273	0.273	0.273	b
	LRM	0.271	0.271	0.271	c
	Hill	0.218	0.218	0.218	d
IoU	RDM	0.492	0.491	0.493	a
	RGB	0.272	0.271	0.273	b
	LRM	0.271	0.270	0.271	c
	Hill	0.218	0.218	0.218	d
DSC	RDM	0.660	0.658	0.661	b
	RGB	0.428	0.427	0.428	c
	LRM	0.426	0.425	0.427	a
	Hill	0.358	0.357	0.358	d
TPR	RDM	0.493	0.493	0.493	a
	RGB	0.273	0.273	0.273	b
	LRM	0.271	0.271	0.271	c
	Hill	0.218	0.218	0.218	d
FPR	RDM	0.002	0.000	0.005	a
	RGB	0.002	0.000	0.005	a
	LRM	0.002	0.000	0.005	a
	Hill	0.001	0.000	0.003	b
Specificity	RDM	0.998	0.995	1.000	c
	RGB	0.998	0.995	1.000	a
	LRM	0.998	0.995	1.000	a
	Hill	0.999	0.997	1.000	a

presents the considered accuracy metrics, including overall accuracy, precision, recall, and Cohen’s Kappa, Intersection over Union (IoU), Dice Score, True Positive Rate (TPR), False Positive Rate (FPR), and Specificity.

The RDM method demonstrated the highest overall accuracy (0.746) and Kappa coefficient (0.491), indicating a strong agreement with the reference strip roads. The RGB method (0.636 accuracy, 0.271 Kappa) followed as the best-performing approach, showing lower agreement with the control data. The LRM method (0.635 accuracy, 0.270 Kappa) performed slightly lower, while the Hill method (0.609 accuracy, 0.217 Kappa) exhibited the weakest agreement with the reference strip roads. The results show that RGB outperformed both LRM (slightly) and Hill in all classical accuracy metrics, demonstrating that high-resolution drone imagery can be a viable alternative to LiDAR for detecting strip roads. However, RGB was still less accurate than RDM, which had the highest overall agreement with the control data.

Precision, which measures how many of the identified strip roads are actually correct, was relatively high for RGB (0.994), indicating that it avoided excessive false detections. However, its recall (0.273) was significantly lower than RDM (0.493), suggesting that it failed to detect a considerable portion of the actual strip roads. Compared to the LiDAR methods, RGB performed better than LRM and Hill in recall, indicating a greater ability to identify strip roads. However, RDM had the best balance of high recall and high precision, making it the most robust method overall.

The Intersection over Union (IoU) and Dice Score (DSC) provide insight into the spatial agreement between the detected and actual strip roads. Again, RDM had the highest IoU (0.492) and DSC (0.660), while RGB showed moderate performance with an IoU of 0.272 and DSC of 0.428. Notably, RGB outperformed LRM (IoU = 0.271, DSC = 0.426) and Hill (IoU = 0.218, DSC = 0.358), reinforcing its effectiveness compared to these LiDAR-derived methods.

The False Positive Rate (FPR) measures the percentage of areas mistakenly classified as strip roads. The RGB method had an FPR of 0.002, meaning it produced a very low level of false detections, similar to LRM and RDM, and slightly higher than Hill (0.001). Specificity, which measures the ability to correctly identify non-strip roads areas, was highest for Hill (0.999), followed by the other three methods with a value of 0.998. This suggests that RGB effectively avoided misclassifying non-strip roads areas while still achieving a reasonable level of recall.

4. Discussion

From an ecological perspective, both the underestimation and overestimation of machinery-induced disturbance can have significant implications for forest management and impact assessment [57]. Underestimating disturbance may lead to an incomplete understanding of the spatial extent and severity of soil compaction and rutting [58]. This, in turn, could result in an underappreciation of long-term effects on soil structure, hydrological processes, and nutrient cycling, ultimately compromising the sustainability of repeated mechanized operations in the same stands [59]. Conversely, overestimating disturbance may artificially inflate the perceived ecological footprint of forest operations, leading to overly restrictive management responses or misinformed policy decisions regarding allowable harvesting intensity or machine types. Both biases therefore affect the accuracy of post-harvest monitoring and the ability to evaluate compliance with sustainable forest management standards. By improving the precision of disturbance mapping, forest managers can more reliably assess the real extent of soil and understory disturbance, enabling adaptive practices that balance operational efficiency with ecological protection. Our study was conducted to specifically address this important topic in the framework of sustainable forest operations.

The results of this study confirmed the initial hypothesis that high-resolution RGB imagery acquired via UAV is effective in detecting strip roads established by boom-corridor thinning in Mediterranean pine forests. Among the methods tested, the RGB-based approach demonstrated good performance across a variety of accuracy metrics, particularly considering the simplicity and accessibility of the method (Table 2). These findings are particularly relevant in forest contexts where the canopy opening is moderate, as in boom-corridor thinning, and where rapid, cost-effective monitoring tools are needed.

Nonetheless, the Relative Density Model (RDM) emerged as the most accurate method overall, showing superior values for nearly all statistical metrics, including precision, recall, and intersection-over-union (IoU) (Table 2). Our findings reinforce the robustness of the RDM method and mirror results obtained in tropical forest studies, where the approach was initially applied to identify skid trails after selective logging. In those cases, reported percentage of detected trails generally fell within the 65%–70% range [34,51,60,61]. Furthermore, the RDM method was also individuated as the most performing in previous research in the Mediterranean context, where it achieved Cohen’s Kappa up to 0.50 in various forest types [36]. By properly adjusting the parameters, especially the height interval used to capture the reduction in vegetation within strip roads, it effectively identified disturbed areas in even-aged Mediterranean pine plantations. This adaptability reinforces the potential for broader application of RDM across forest types when parameterized appropriately.

However, it is important to highlight the practical limitations of the RDM method. Unlike the other techniques, RDM is computed directly from the LiDAR point cloud, rather than from a Digital Elevation Model (DEM), which makes it considerably more demanding in terms of computational resources. Effective use of RDM, particularly when relying on graphical software such as CloudCompare 2.14 beta, requires high-performance hardware with significant RAM, processing power, and graphic card capacity. This makes it less suitable for practitioners working with standard computing resources. To provide a practical reference, the analysis was performed on a workstation equipped with 64 GB of RAM and an Intel Core i9 processor; under these conditions, the computation required approximately five minutes and utilized up to 80% of the available memory. When the same procedure was attempted on a standard user-grade laptop (8 GB RAM, Intel Core i5 processor), the system failed to complete the operation due to insufficient computational capacity.

In comparison, the hillshading (Hill) method—despite its ease of use and wide availability in standard GIS software—showed the lowest performance. This is consistent with the findings of Affek et al. [49], who also noted the limitations of hillshade visualization for subtle terrain features in forested environments with less than 40% of actual machinery trails detected. Its reliance on visual contrast derived from artificial lighting angles renders it less reliable for consistent identification of strip roads. However, it is worth highlighting that this was the method with the highest specificity, highlighting that when the Hill method detects a strip road, we can be almost sure that the strip roads actually exist.

The Local Relief Model (LRM), while showing intermediate results, offered a good compromise between simplicity and accuracy. As demonstrated in previous studies [38,62], local relief filtering can successfully isolate microtopographic features like strip roads, particularly when parameters are carefully tailored to the size and depth of the disturbances. However, it requires careful adjustment of Gaussian smoothing filters to effectively eliminate broader terrain forms while preserving the smaller depressions associated with trail formation.

However, while RDM and RGB showed high accuracy, the future of such methodologies likely lies in their integration with automated classification techniques. Machine learning offers the potential to greatly enhance efficiency by removing the need for manual digitization. That said, such approaches require large, annotated datasets for training, resources that are often unavailable in small-scale, fragmented forest ownerships like those typical of Central Italy and much of the Mediterranean region. Moreover, the recent literature shows that the accuracy levels achieved through machine learning methods for trail detection are comparable to those obtained in this study using RDM and even RGB, with higher overall accuracy (0.846) achieved only under optimal conditions (even-aged young stands of coniferous species) [63]. Thus, while automation is a clear benefit, it does not yet translate into superior precision.

It is also necessary to acknowledge some limitations of the present study. The first one consists of using manual digitalization, which implies a certain level of subjectivity depending on the operator. The application of machine learning (ML), as previously stated, could be a possible solution to make strip road identification more objective and faster when working on large areas. However, although machine learning (ML) methods represent an increasingly powerful tool for automated feature extraction in remote sensing, their application was not pursued in this study for several reasons. First, the extent of the study area (approximately 8 ha, corresponding to about 3 km of strip roads) results in a dataset that is too limited to train data-intensive models such as convolutional neural networks, which typically require large numbers of labeled examples to achieve reliable generalization [64,65,66]. In addition, the operational scale considered here, namely the forest management unit, does not justify the use of complex ML workflows, as manual digitization proved to be highly efficient, requiring only 10–15 min per method to delineate all strip roads. At this scale, ML would not meaningfully reduce processing time or improve practical usability. Furthermore, the purpose of the study is to offer forest managers solutions that are directly applicable within routine monitoring practices; introducing more elaborate ML procedures would increase technical complexity without delivering proportional benefits. While ML techniques could play a valuable role in large-area applications, such as regional- or management-plan–level disturbance mapping where airborne laser scanning would serve as the primary data source, these scenarios fall outside the scope of the present work.

A second limitation arises from the fact that the analysis was carried out in only one study area, which, while highly representative of Mediterranean even-aged pine forests undergoing mechanized thinning, does not capture the full variability of forest structures or terrain conditions that may affect method performance. Nevertheless, the chosen site provides a suitable and relevant context for testing our hypothesis and drawing meaningful conclusions. A further potential limitation of this study concerns the transferability of the results to other forest types and terrain conditions. The analyses were conducted in a relatively homogeneous black pine stand located on moderate to steep slopes typical of Mediterranean mountain forests. While the selected site is representative of many artificial coniferous stands in Central and Southern Italy, the structural characteristics and understory development may differ substantially in other forest ecosystems. Similarly, steeper or more heterogeneous terrain could affect both UAV data quality and the performance of terrain-based LiDAR derivatives such as LRM and RDM [36]. Consequently, the accuracy levels and method rankings reported here should be viewed as context-dependent, and further validation across different species compositions, canopy densities, and slope gradients is needed to confirm the robustness of these findings under broader environmental conditions.

5. Conclusions

This study addressed the challenge of detecting machinery-induced strip roads in Mediterranean pine forests following boom-corridor thinning operations. Given the need for sustainable forest management of these widely distributed stands, the ability to monitor soil disturbance accurately and efficiently is essential. We compared four different methods for identifying strip roads—one based on UAV-derived RGB imagery and three based on LiDAR data, including Hillshading, the Local Relief Model (LRM), and the Relative Density Model (RDM).

Our results confirm that RGB method, despite its reliance on passive sensors, can effectively detect the strip road network under the moderate canopy openings characteristic of boom-corridor thinning. This finding supports our initial hypothesis and highlights the operational value of low-cost UAV-RGB surveys in such forestry contexts. Among the LiDAR-based approaches, the RDM method proved to be the most accurate and reliable, even when applied in a temperate Mediterranean conifer forest—significantly different from the tropical environments where it was originally developed. However, the RDM method is also the most computationally intensive, requiring high-performing hardware and careful parameter tuning.

From a practical standpoint, these findings provide clear guidance for forest managers and practitioners. Specifically, the RDM approach delivered the highest performance, but UAV-based RGB surveys, thanks to their lower cost, satisfactory accuracy, and user-friendly implementation, represent a highly suitable option for identifying the strip road network after boom-corridor thinning in coniferous stands.

Nevertheless, the generalizability of the results remains limited by the characteristics of the study area, a single, relatively homogeneous black pine stand on moderately steep terrain. Further research should validate the findings across a wider range of forest types, canopy structures, and slope gradients to confirm their applicability in more diverse contexts. Looking forward, the integration of automation and machine learning offers promising directions for future work. Approaches based on convolutional neural networks (CNNs), deep learning segmentation of point clouds, or hybrid optical–LiDAR feature extraction could greatly reduce operator dependency and increase processing scalability. Such advances would allow continuous, near-real-time monitoring of forest disturbance over larger areas while maintaining the precision required for sustainable forest management.