Object-Based Classification of Abandoned Logging Roads under Heavy Canopy Using LiDAR

2.1. Study Area

The study area is located on the eastern slope of the Bolinas Ridge above Kent Lake Reservoir in Marin County, California (Figure 1). The site is approximately 3 square kilometers of heavily forested steep terrain, with an average slope of 29° and elevations ranging between 109 and 470 meters above mean sea level. The forest canopy in the study area is primarily characterized by second growth coast redwood and Douglas-fir. Canopy cover is dense and consistent throughout the study site and logging roads within the study site are completely occluded by vegetation (Figure 1).

2.2. Logging Road Background

The Marin Municipal Water District purchased the lands and logging rights for the area containing the study site in 1949 [32]. The extensive network of logging roads found within the study site was likely built shortly after. Between 1949 and 1952 an estimated 3 million board feet of timber was processed at the Ruoff mill, the destination of timber from the study site [32].

Since logging ended in 1952, logging roads within the study site have remained unmanaged and undisturbed. Bank erosion is evident throughout the road network and fallen trees and forest litter are present on most roads (Figure 2). However, roads are still clearly distinguishable on the ground and little tree growth has occurred on road surfaces. Roads are mostly located perpendicular to the hillslope and following ridgelines. Measured road surface slopes within the study site were found between 0° and 24.5° degrees with an average of 10°. The average road width within the study site was 4.1 meters.

4.1. Field Survey

Since logging road maps of the study site were unavailable, a line-intercept field sample was used to gather classification validation points and road characteristics. The field survey was completed over five days between 17 November 2012 and 24 August 2013. Road and non-road points were collected along six transects running from the Bolinas Ridgeline on the western edge of the study site to the eastern edge of the study site. Transect starting points were chosen on the Bolinas Ridgeline in order to cover the entire study site. Originally 200 meter intervals between transects were chosen, but in order to adequately cover the entire study site, it was determined that larger intervals were needed. The exact starting point of each transect was chosen randomly within 100 meter zones. Each transect was then run on a 45 degree angle from the Bolinas ridge trail to the eastern edge of the study site. At each logging road found along a transect, the following road characteristics were collected: width, aspect, slope along road, slope across road, and road position, e.g., within gully, along ridgeline, or following contour. A GPS point was also collected at the center of the road. Non-road GPS points were collected on transects at approximately 100-m intervals to provide non-road validation points. A total of 43 non-road points and 57 road points were collected with a handheld Trimble Juno SB GPS receiver. All GPS points were post-processed with differential correction using reference data from the base provider UNAVCO, Inverness, CA. Despite the stated post-processing accuracy of L1 code of 1–3 m [34], large GPS positional errors can occur in steep forested terrain due to high position dilution of precision (PDOP) and multipath LiDAR returns [35]. A high PDOP indicates that the geometry of the available satellites may lead to positional errors in the resulting GPS points. After a visual inspection of the GPS road points, it became clear that many points were located up to several meters away from their true position. When necessary, GPS road points were moved to the intersection of the corresponding transect and the correct road visually identified on the slope model.

4.2. Object Based Classification

Logging road classification was performed in eCognition® Developer 8.7.2, an Integrated Development Environment (IDE) for image segmentation and the creation of object-oriented classification rule-sets [25]. Within eCognition, an Edge Extraction Lee Sigma filter was first applied to the previously created slope model. This edge extraction algorithm uses an edge filter to create bright and dark edge layers from the original image [25]. In this case dark edges were extracted. A sigma value of 5, determining the strength of the edge detection, was chosen resulting in the image shown in Figure 3. The inclusion of an edge-enhancement filter was precipitated by similar applications by Reiger et al. [18] and Harmon [22]. A sigma value of 5 creates a lee sigma image with strong edge detection.

An initial multiresolution segmentation was performed with the slope image and the edge-enhanced filtered slope image as inputs with equal layer weights. As is often the case with image segmentation in the Ecognition environment, a trial-and-error approach to parameter refinement was needed [36]. In this case the segmentation method, input weights, and shape and compactness parameters were empirically chosen with the goal of creating homogeneous road objects. For this project a scale value of 6 was chosen along with values of 0.1 for shape weight and 0.5 for compactness weight. The shape value of 0.1 gives strong favor to spectral signature over shape during segmentation, while the compactness value of 0.5 indicates equal value was given to maintaining heterogeneity of the “smoothness” and “compactness” of image objects [25]. These parameters were empirically chosen to create the largest possible objects that fall entirely within either a road class or a non-road class.

A rule-based classification approach, whereby classification was performed on the image objects based on thresholds for two object parameters, was used to create a road seed class and a road candidate class. The two object parameters selected were the “mean slope value” of an image object and the “mean difference of an object’s slope to its brighter neighbors”. The “mean slope value” parameter is the average slope of each object within the image. The “mean difference of an object’s slope to its brighter neighbors” refers to the mean difference between the mean slope value of each image object and the mean slope value of each adjacent object with a higher mean slope value [25]. More conservative thresholds, meaning a lower “mean value” threshold and higher “mean diffierence to brigher neighbor” threshold, were chosen for the road seed class ensuring a high likelihood that road seed objects were correctly classified (Figure 4a). Remaining unclassified objects were classified into a road candidate class using less conservative parameter thresholds, meaning a higher “mean value” threshold and lower “mean diffierence to brigher neighbor” threshold, in order to ensure most remaining road objects were captured within this class (Figure 4c). Classification parameter thresholds are shown in Table 1.

Road seed objects were merged and small objects below 10 pixels in size were removed from the road seed class. A straightening process was then applied to the road seed class to create straightened elongated road objects more suitable for the road growing. Object straightening was achieved by running multiresolution segmentation iteratively over a reducing scale parameter with a shape criterion of 0.9 and a compactness criterion of 0.9. A comparison of objects before and after straightening is shown in Figure 4a–b.

An iterative road growing and road straightening process was used to grow road seed objects into adjacent road candidate objects. Road seed objects were merged with adjacent road candidate objects based on a decreasing density requirement. Density is a measurement of the distribution of pixels within an object. Denser objects have a structure approaching a square while less dense objects have a shape more similar to a filament [25]. It is expected that as a road increases in length, its density will decrease. Therefore road seed objects were only merged with adjacent road candidate objects if the density of the combined object was less than the density of the road seed object. This process was run iteratively to grow the road into adjacent candidate objects (Figure 4d). Finally, non-road objects fully enclosed by road class objects were classified as roads. The resulting road classification was exported as a shapefile into ArcGIS.

4.3. Road Post Processing

Two levels of post-processing were performed on the results of the object based road classification within ArcGIS 10.1. First, misclassified gullies and streams were removed. A streams layer was created in ArcGIS using the following workflow: LiDAR derived DEM > Flow direction raster > Flow accumulation raster > Streams layer. The resulting stream layer was buffered to 4 meters and used to erase roads falling within natural drainages.

The second post processing level involved hand digitizing ridge roads that were missed by the original object based classification. A plan curvature raster was created from the LiDAR derived DEM using the curvature tool in the Spatial Analyst toolbox [33]. Plan curvature is the rate of change in the flow direction of the surface, following the contour direction. The plan curvature image is helpful in distinguishing ridgelines [37] and ridge roads (Figure 5). Ridge roads visible on the plan curvature image were hand digitized in ArcGIS and joined to the existing classified road layer. Six ridge roads with a total length of 3.9 km were added.

4.4. Unsupervised Pixel-Based Classification

An unsupervised k-means classification of the LiDAR derived slope image was performed to compare with the object-based classification results. Following this approach pixels are separated into a user defined number of clusters based on a minimum distance method [38]. The classification was performed with ERDAS Imagine 2010 with two classes: a road and non-road class.

4.5. LiDAR Point Density Reduction

In order to examine the sensitivity of the road classification to LiDAR point spacing, the original LiDAR point cloud filtered for ground points was thinned at several levels and the previously described slope model creation and object-oriented classification procedures were re-applied. The original LiDAR ground point cloud with a point spacing of 0.91 points per m² was randomly thinned by a fraction of 0.5. Iterative point thinning resulted in 6 levels of LiDAR ground point spacing (Table 2).

4.6. Classification Accuracy Assessment

Logging road classification accuracy was assessed using GPS ground reference data acquired during the field investigation. Accuracy statistics are summarized in the form of a confusion matrix for each classification result. Classification accuracy assessment was performed for the initial object oriented classification as well as for each post-processing step. Results for the unsupervised classification and for each point spacing level were also calculated.

5.1. Object-Based and Pixel-Based Classification Results

The results for the object-based classification of logging roads summarized in Table 3 show a total accuracy of 86% was obtained from the initial object oriented classification of the LiDAR derived slope image. This accuracy increased to 88% and 90% respectively after extracting misclassified gullies and adding manually digitized ridge roads (Table 3). For the object-oriented classification with gully roads extracted and with ridge roads added, the user’s accuracy was 98% while the producer’s accuracy was only 84%, indicating that a majority of classification errors were errors of omission or roads missed by the classification. Full error matrices are shown in Table 4. A total of 0.275 km² of area were classified as roads within the study site (Figure 6).

The pixel-based classification results show that the unsupervised classification of logging roads had a total accuracy of 78% (Table 3). A lower user’s accuracy was found in the pixel-based classification than in the object-based classification. Unlike the object-based classification, errors of commission were of significant concern. Errors of commission occurred largely on ridgelines where large areas of low slope were present (Figure 7). The object-based region growing approach was better able to constrain road classification correctly to roads in these areas. Within road areas, the pixel-based classification was less successful in classifying continuous road surfaces. Large and frequent classification gaps are found on the roads in the pixel-based classification results (Figure 7.)

5.2. Point Reduction Classification Results

Thinning the LiDAR ground point cloud had a strong impact on the resulting classification accuracy (Table 2). Total accuracy ranged from a high of 86% with a ground point spacing of 0.91 points per m² to 46% with a ground point spacing of 5.66 (Table 2) when the object-based classification was applied without post-processing. The decline in producer’s accuracy indicates that errors of omission are the cause of reducing classification accuracy. This effect is evident in the classification results shown in Figure 8.

6.1. Object-Based Classification

The fully automated object-based classification performed well with a total accuracy of 86%. This fully automated object-based approach was an improvement over pixel-based unsupervised classification which resulted in a road classification accuracy of 77%. The object-based approach also outperformed a supervised classification approach to forest road detection attempted by Harmon [22], which yielded a classification accuracy of 73% before post processing. Applying shape conditions to road classification allowed the object based classification to minimize errors of commission in low slope areas. The object-based classification was comparable to total classification accuracies of 85%, 76%, and 74% found by Espinoza et al. [19] for a manual road classification based on visualizing the LiDAR point cloud. Distinguishing natural terraces and drainages from roads was a difficulty shared by several authors [16,19]. This problem derives from the fact that some natural features, like ridgelines, can have similar slope characteristics to manmade road objects. While, in most cases, ridge roads were over-classified [16,19], in our case, a majority of the total error was caused by an under-classification of roads on ridgelines. The road detection parameter “mean difference to brighter neighbors” was largely responsible for reducing the effectiveness of road classification on larger ridgelines. Since slope is fairly uniform surrounding a ridge road, the “mean difference to brighter neighbor” threshold eliminates some ridge roads from the classification. This effect was visible when the “mean difference to brighter neighbor” parameter value was refined using a trial-and-error approach. While adding hand-digitized ridge roads from the plan curvature model requires manual interpretation, it does help to reduce missed ridgeline roads in the classification. A two percent increase in total accuracy after adding hand-digitized roads to the initial object-based classification demonstrates that the plan curvature model is a useful tool in distinguishing ridgeline. However, several ridge roads were still missed, because of difficulties in distinguishing ridge roads from the background terrain. Higher LiDAR point density, allowing for increased slope resolution may help in road detection on ridgelines. Several authors found that incorporating an intensity image in forest road classification methods was helpful [16,23]. However, the high level of debris found on abandoned logging roads would reduce its usefulness for abandoned logging road classification.

Misidentification of natural drainage features as roads was another source of classification error. Because of the lack of field points directly within gullies or streams, the extent of this error was not captured in the error results. However, the simple post-processing step of removing drainages from the road classification seems to largely take care of this problem. Espinoza et al. [19] suggest examining the height profile across a potential road object to determine whether it is a road or a natural drainage. Natural drainages tend to have height profiles with a “V” shape while roads will have a partially flat profile.

6.2. LiDAR Point Reduction

Several authors note that LiDAR ground point densities were greatly reduced under canopy [17,23]. Over 94% of initial LiDAR returns in a study by White et al. [17] were classified as vegetation or other non-ground point returns. By examining the accuracies of manual road classification across several study sites, Espinoza et al. [19] found that point spacing was a large factor in determining classification quality. Our examination of the effects of LiDAR point thinning on the road classification results confirm that finding and reflect a first attempt at isolating the effects of LiDAR point spacing in order to quantify LiDAR point spacing requirements for road detection under canopy. As shown in Figure 8, the usefulness of LiDAR for road detection is drastically reduced as point spacing increases. Between an average ground point spacing of 1 point per m² and 2 points per m², producer’s accuracy falls from close to 70 percent to below 50 percent (Figure 9). Within this range, errors of omission reach a level where the usefulness of the resulting classification would be questionable. A similar LiDAR point decimation strategy was performed by [39] to examine the impact of LiDAR post spacing on flood zone delineation. Raber et al. [39] suggest a 4 meter post spacing requirement for flood plain delineation. Our result suggests a much higher LiDAR post spacing is needed for forest road detection.