While forest roads are important to forest managers in terms of facilitating the exploitation of wood and timber, their role is far more multifunctional. They permit access to emergency services in the case of forest fires as well as acting as fire breaks, enhance biodiversity, and provide access to the public to enjoy recreational activities. Detailed maps of forest roads are an essential tool for better and more timely forest management and automatic/semi-automatic tools allow not only the creation of forest road databases, but also enable these to be updated. In Spain, LiDAR data for the entire national territory is freely available, and the capture of higher density data is planned in the next few years. As such, the development of a forest road detection methodology based on LiDAR data would allow maps of all forest roads to be developed and regularly updated. The general objective of this work was to establish a low density LiDAR data-based methodology for the semi-automatic detection of the centerline of forest roads on steep terrain with various types of canopy cover. Intensity and slope images were generated using the currently available LiDAR data of the study area (0.5 points m-2). Two image classification approaches were evaluated: pixel-based and object-oriented classification (OBIA). The LiDAR-derived centerlines obtained with the two approaches were compared with the real centerlines which had previously been digitized in the field. The road width, type of surface and type of vegetation cover were also recorded. The effectiveness of the two approaches was evaluated through three quality indicators: correctness, completeness and quality. In addition, the accuracy of the LiDAR-derived centerlines was also evaluated by combining GIS analysis and statistical methods. The pixel-based approach obtained higher values than OBIA for two of the three quality measures (correctness: 93% compared to 90%; and quality: 60% compared to 56%) as well as in terms of positional accuracy (± 5.5 m
The forest road network comprises roadways within the forest which are used for access or for timber extraction, an infrastructure which, among other things, is essential for carrying out the sustainable management of forests (
From the point of view of forestry production, in order to carry out profitable and sustainable forest management it is crucial to create management tools that help to establish the availability of resources and the economic viability of their use, so that logging costs can be minimized. To efficiently calculate transport costs (including hauling and road transport phases), it is essential to have access to information on the extent and conditions of the current road forest network. The role of GIS (Geographical Information Systems) in this respect is increasingly important and such data can provide a solid framework for forestry companies seeking to improve their logistics, as it allows the optimization of transport planning according to various criteria (
In general terms, the task of detecting and extracting the course of roads from digital information has been addressed in a number of ways over the years: photointerpretation and manual digitization, the integration of cartography from different sources, data capture in the field, and data capture from remote sensors (
Within this unfavorable context for the application of remote sensing images in forest areas, LiDAR technology brought about a great advance in the automatic detection and updating of forest road networks, because it provides accurate measurements of ground elevation, which allows high resolution topographic mapping, even under dense canopy cover (
In general, while
In this respect, Spain is one of the few countries in the world to have a National Plan for Territory Observation (LiDAR-PNOA© National Geographic Institute of Spain) which ensures the capture of LiDAR data at various times across the whole of its territory. At the moment, data from only one flight (between 2008 and 2015, depending on the region) is available for civil use, but more flights to update the current data are expected to take place in the next couple of years. While this data has one main drawback - the low cloud point density used for the vast majority of the territory (only 0.5 points m-2) - this is counterbalanced by the fact that data is available for the whole national territory at no cost. In addition, the fact that the data is to be updated periodically with further flights means that there is an ongoing source of date to use to ensure that the mapping of the forest road network is regularly brought up to date.
Despite the easily accessible availability of LiDAR data and the great potential that this technology has shown in various studies, in Spain no great efforts have been made to build a public cartography of forest road networks. This fact is especially worrisome in the northern part of the country (specifically in Asturias) where according to the Spanish National Forest Inventory (
In view of the above, this study aims to test the hypothesis that it possible to use low-density LiDAR data to semi-automatically detect and extract forest roads and that this is a more efficient alternative than manual digitization. The study aims to design a semi-automatic methodology to obtain the centerline of forest roads from low density LiDAR data in a steep forested environment. Two approaches have been evaluated: a pixel-based methodology and one which is object-oriented (OBIA). In addition, the influence of the characteristics of forest roads (width and type of road surface) and the surrounding vegetation (type of canopy cover bordering the road) on the results of these two types of semi-automated detection was also analyzed. Steep environments are acknowledged to be an extremely difficult context for forest management (
The study area is located in Asturias, a mountainous region situated in the north-west of Spain. A state-owned forest of 178.77 ha was chosen as a pilot area because it is a representative sample of the type of forests found in the region, which are generally characterized by steep slopes and a variety of species. Elevation range is between 36 and 335 meters a.s.l. and over half the area (55.7%) has a slope ranging between 31% and 60%, while 20.7% is above 60%. These characteristics provide a challenging though realistic environment where the detection methodologies of this study could be assessed.
The study area incorporates a number of stand types. There are pure stands of
LiDAR data for the study area was captured under the framework of the PNOA (
In order to test the accuracy of the LiDAR detection of forest roads, the real centerlines of the forest roads network in the study area were collected in the field during the summer of 2013 using a GPS Trimble Explorer XH™ (Trimble, Sunnyvale, CA, USA) with submetric accuracy. This data was collected in shape format, whereby the lines defining the centerlines of forest roads are associated with a database which records their main attributes (width and type of road surface and surrounding vegetation).
A further set of field data was also captured using a GPS model with centimetric accuracy (TOPCON GR-3™, TopCon Positioning Systems Inc., Livermore, CA, USA) which was then used to assess the accuracy of the LiDAR-derived products to be used in the analyses. Firstly, to assess the accuracy of the LiDAR-derived DTMs, 55 ground points were measured in the field which were located in areas of varying degrees of slope and different types of vegetation to ensure variability in these factors was covered. Secondly, to evaluate the planimetric accuracy of the LiDAR-derived centerlines obtained with the two approaches, a subsample was selected consisting of four sections of forest roads (A, B, C and D), each with a different type of road surface, and sometimes differing in road width and surrounding vegetation (
For the detection of the forest road network two LiDAR inputs were used: slope map and intensity image. The intensity image enables covered areas to be distinguished from uncovered areas, and the slope map provides information about terrain morphology (
In order to get an accurate slope map, the first step was to obtain a DTM of the study area. The procedure to obtain a DTM from LiDAR data involves two differentiated steps: the separation of the LiDAR point cloud into those points belonging to the ground and those belonging to tree cover through a filtering process; and the subsequent interpolation of ground points to generate a continuous surface that comprises the DTM. A two-step validation process (see below) was carried out so that the best combination of parameters in each case was established and then used to produce the final DTM.
Firstly, the filtering of ground points was conducted using the “GroundFilter” function (an adaptation of the IRI filter by
In order to assess the accuracy of the LiDAR DTM, a robust statistical error validation process (
The DTM which obtained the best results following the two validation steps (
Due to the fact that intensity values are influenced by terrain, flight and sensor characteristics, as well as by atmospheric conditions (
To extract the forest roads from the LiDAR data, the workflow shown in
In the first approach, the normalized intensity image and the slope map were subjected to a pixel-level classification using the Maximum Likelihood (ML) algorithm. This method considers that digital levels within each class fit a normal distribution such that each classification category can be described by a probability function deduced from its mean vector and its matrix of variance-covariance. The calculation was performed for all classification categories involved, each pixel being assigned to the category that maximized the probability function (
In the second approach, the normalized intensity image and the slope map were subjected to an object-oriented unsupervised classification. The algorithm for the segmentation of both images was Meanshift (MS) segmentation (
The intensity images and slope maps from the two different classification processes used were then reclassified, based on a manual visual analysis, to obtain a binary image - “forest roads” (1) and “not forest roads” (0) - for both slope and intensity.
These binary slope and intensity images were then refined using a majority filter which removes single pixels or noise and replaces the cell(s) depending on the categories of the neighboring cells. After that, the two binary images from each approach were combined by multiplication to obtain one single final binary image, also classified into 0 (not road) or 1 (roads), for the pixel approach and one for the OBIA approach.
This final binary image resulting from each approach was then subjected to a second refining process in order to remove noise and to achieve continuity within category 1 “forest roads” (dilation and closing filter). The final step was automatic vectorization, a technique which converts raster data into vector entities (
The workflow explained above was automatized with the help of GIS software that uses a Model Builder.
The results from the two approaches were assessed using the methodology described by
From the number of TP, FN and FP found, the quality measures were calculated, as described in the work of
Completeness is the percentage of the reference data (in this case the field data) explained by the extracted data, while correctness represents the percentage of correctly extracted road data. Quality is a more general measure of the final result which combines completeness and correctness into a single measurement. The maximum value for each of the quality measures is 1 (100%).
An analysis of variance (ANOVA) was performed to evaluate the influence of various factors on the detection of forest roads, particularly on the quality measures: completeness, correctness and quality. The factors analyzed were: surrounding vegetation type, road surface and road width. All possible combinations of the three factors were considered resulting in a forest road classification of 25 classes. In addition, the influence of the two different approaches used in the detection was also evaluated.
Besides estimating the quality of the two methods tested in terms of completeness correctness and quality, the positional accuracy of the centerlines of the LiDAR-derived centerlines was compared to the field-survey centerline of sections A, B, C and D using a simple method proposed by
The LiDAR-derived centerlines obtained with each of the two approaches are shown graphically in
In terms of completeness, the value obtained in the two approaches was similar, meaning that total percentage of the forest roads network detected with both approaches is around 60-65%. The correctness value was also very similar for both, and indicates that around 90% of the LiDAR-derived centerlines detected automatically represented real roads. Commission error in both cases was therefore low.
Quality provides a more global measure of reliability, since it takes into account both the completeness and the correctness of the extracted data (
Based on the real centerlines collected in the field, the total length of the forest roads network is 24 km, 70% of which were identified in the visual inspection as being in good condition. In relation to road width, the vast majority of forest roads (75%) were wide enough to allow for the circulation of forestry machinery (2.5 m or more).
In this scenario, and despite the heterogeneity of the study area which presented great differences in terms of orography, types of surrounding vegetation and road surface, both approaches were able to detect and draw the centerline of the principal forest roads. Both approaches were particularly effective in areas where road width was > 2.5 m and when roads were not occluded by vegetation, conditions particularly prevalent in the southern part of the study area (
The strongest influence on detection was that of the surrounding vegetation, as can be observed in
Comparing these results with those of other authors, a study carried out by
In the future, in line with the PNOA, when LiDAR data of higher density will be captured for the whole of Spain, the results obtained with the methods detailed in this work will be improved in terms of both quality and positional accuracy. Until such data will be available, the approaches described here can serve to provide the first step in developing a large-scale national database of forest roads networks.
Two limitations of the approaches used here must, however, be recognized. These pertain to forest roads width and the type of vegetation bordering forest roads. The majority of the forest roads that were not detected by either approach are narrow and, on the whole, surrounded by dense broadleaved vegetation. The influence of the type of vegetation is made especially evident by the fact that the centerline of forest roads running through broadleaved stands is at times completely occluded, thus making road detection exceedingly difficult. It should be noted that the LiDAR data used in this study was captured during summer, when broadleaved canopies were in full leaf. Thus, the laser beam cannot penetrate through the dense canopy to the ground, resulting in a lower quality DTM not representing the shape of the land surface, but rather of the vegetation. In the case of intensity image, the leafy areas appear very dark because they represent the highest points of the vegetation, especially in those areas where the high canopy density hinders the centerline of a forest road to be detected from the air, or is completely masked. In areas where there are two types of vegetation stand adjacent to each other, such as pine and broadleaved stands, it can be seen that the LiDAR-derived centerline was detected without problem in the pine forest area but disappears or is interrupted in the part bordered by broadleaved trees.
In reforested stands however, the percentage of LiDAR-derived centerlines detected was high, since the relatively recently planted vegetation was not very dense and did not occlude the centerlines of the roads. However, the number of false positives in these areas was also high, because the presence of bare soil among the trees gives rise to high intensity values, and the classification algorithm has problems distinguishing soil between lines of smaller young trees from the surface of the road. According to the results of
In this study, forest roads that did not have pronounced gradient were difficult to identify, as can be observed in the bottom left corners of
The results of the assessment of the positional accuracy of the LiDAR-derived centerlines obtained with the method of
The average positional error was ± 5.50 m for the pixel-based approach and ± 6.88 for OBIA. In the case of LiDAR-derived centerline accuracy, the fact that the pixel-based classification was more accurate may be related to the resolution of the images used in the classification process. According to
Comparing the results obtained in this study with other similar works, accuracy values are lower than those obtained by
One final point to note is that forestry road extraction is typically a manual process where positional error depends on a number of factors, but the human factor (
In this study a methodology for the detection and extraction of forest roads from freely available LiDAR data of low density was designed and applied. The results obtained confirm the initial hypothesis that it is possible to semi-automatically reconstruct the forest roads network even in steep forested environments. As the analysis procedure has been implemented in a GIS Model Builder, it can be applied quickly and easily in other forest areas with similar or less complex and challenging characteristics.
Of the two approaches evaluated, the pixel-based classification method yielded slightly better results than the OBIA one with regard to the quality measures and positional accuracy. The completeness, correctness, and quality values were 65%, 90% and 60%, respectively, compared to 59%, 93% and 56% obtained with OBIA. Despite this, the results of the ANOVA demonstrate that neither the methodology used nor the type of road surface had any significant influence on the detection and digitization of the road centerline, although road width and the type of surrounding vegetation do. In fact, the results indicate that low density LiDAR data is suitable for the detection and digitization of forest roads over large areas, especially those where forest roads are wider (over 4 m) and are not surrounded by broadleaved stands. With respect to positional accuracy, the values obtained by pixel-based classification are on average, 1.38 m more accurate than those from the OBIA for each of the sections examined (± 5.50
Regarding the methodology limitations, future research should focus on both improving the effectiveness of image classification and achieving more defined and continuous lines (road centerlines) in adverse vegetation and slope conditions.
Finally, efforts at a national level by governments to capture higher cloud point density data on a country-wide level open the door to large-scale detection trials. In Spain, the data that will be captured in the future, within the framework of PNOA, will allow the methodology presented in this study to be used to develop a cartography of forest roads, including mountainous areas, which can be updated every time new data is released. However, in the meantime, the approaches described in this work offer a valuable first step towards such a complete large-scale database of forest roads networks.
This study was funded by the SCALyFOR project (R&D Projects “Research Challenges”, Spanish Ministry of Economy and Competitiveness). The authors would like to thank the Forestry Service of the Principality of Asturias (Spain) for providing the information used. Thanks also to Ronnie Lendrum for revising the English.
(A) Distribution of different stand types across the study area; (B) DTM resulting from the validation step; (C) Slope map resulting from the DTM; (D) Intensity image of the study area.
The workflow followed to obtain LiDAR-derived centerlines.
Measures used to calculate the quality measures in forest road detection for the two assessed methodologies. (A) “True Positive” (TP): LiDAR-derived centerlines that are real forest roads. (B) “False Positive” (FP): LiDAR-derived centerlines that do not follow the real road. (C) “False Negative” (FN): forest roads not identified.
Forest road network obtained with the two detection approaches: pixel-based (A) and OBIA (B).
Relationship between the values of the quality measures evaluated in the pixel-based classification and the surrounding vegetation.
Percentage of LiDAR-derived centerline lying within the buffer as a function of buffer width using pixel-based approach (A) and OBIA approach (B).
Characteristics of the subsample sections.
Section | Length(m) | Road surface | Road width (m) | Surrounding vegetation |
---|---|---|---|---|
A | 994.3 | Aggregate | 2-4 | Pine |
B | 809.2 | Dirt | 2-4 | Pine |
C | 266.1 | Rock | 2-4 | Reforested |
D | 359.8 | Aggregate | >4 | Pine |
Quality measures obtained for the two detection approaches: pixel-based (Pixels) and OBIA.
Approach | Completeness | Correctness | Quality |
---|---|---|---|
Pixels | 0.65 | 0.90 | 0.60 |
OBIA | 0.59 | 0.93 | 0.56 |
Results of Analysis of Variance (ANOVA) to quantify the influence of the factors on the quality measures.
Metric | Vegetation | Road surface | Methodology | Road width | ||||
---|---|---|---|---|---|---|---|---|
F | prob | F | prob | F | prob | F | prob | |
Completeness | 7.92 | 0.0002 | 1.27 | 0.2967 | 0.25 | 0.6205 | 3.97 | 0.0258 |
Correctness | 6.64 | 0.0008 | 0.98 | 0.4123 | 0.54 | 0.4667 | 2.86 | 0.0677 |
Quality | 7.49 | 0.0004 | 1.14 | 0.3453 | 0.44 | 0.5120 | 4.8 | 0.0129 |
Positional accuracy values (in meters) for the two approaches for each section.
Section | Length(m) | Roadsurface | Roadwidth (m) | Vegetation | Pixel | OBIA |
---|---|---|---|---|---|---|
A | 994.3 | Aggregate | 2-4 | Pine | 7.00 | 10.00 |
B | 809.2 | Dirt | 2-4 | Pine | 5.50 | 5.50 |