A Topographic Wetness Index for Forest Road Quality Assessment: An Application in the Lakeland Region of Finland

Abstract

Research Highlights: A Topographic Wetness Index calculated using LiDAR-derived elevation models can help in identifying unpaved forest roads that need maintenance. Materials and Methods: Low-pulse LiDAR data were used to calculate a Topographic Wetness Index to predict unpaved forest roads’ quality. Results: The results of this analysis and comparison of road-quality features derived from LiDAR data at resolutions of 1, 10 and 25 m for assessing road quality in the boreal forests of Finnish Lakeland show that the wetness index can predict road quality correctly in up to 70% of cases and up to 86% when combined with other auxiliary GIS-based variables. Conclusions: Road-quality assessments, using airborne LiDAR data, can greatly help forest managers to decide which sections of the ageing road network will benefit the most from maintenance, while reducing the need of field visits.

1. Introduction

Forest roads are essential for silvicultural and harvesting activities, and their quality and regular maintenance are essential for securing access to forest resources [1]. Their importance lies in giving access not only for harvesting purposes but also for recreational use and firefighting. Finland has an extensive forest road network, over 3000 km of which needs maintenance each year [2]. The annual tasks include the construction of new roads and bridges, and the rehabilitation and maintenance of existing roads in the form of grading, drainage improvements, clearing of roadside vegetation and resurfacing with gravel or stone chippings. The Finnish state-owned forestry organization Metsähallitus, which is responsible for over 37,000 km of forest roads [3], spends about 16 million euros a year on ditch maintenance, 46 million euros on the construction and basic improvement of roads, and 17 million euros on the basic maintenance of forest roads [2]. It is thus important to be able to identify the stretches of road that require maintenance, as the whole road network cannot be repaired at once, due to the high costs and labor requirements.

Light detection and ranging (LiDAR), such as Airborne Laser Scanning (ALS), is often used in forestry for assessing forest stocks and the growth and composition of stands, but, during recent decades, its applicability to road and forest road maintenance has also been tested [4,5]. The use of ALS requires more consideration in forested areas than in urban environments, due to differences in canopy conditions, but several studies have focused on extracting forest-road information from low-pulse-density LiDAR data. A 1.12 pulses per m² point cloud density was used to create a 1 m Digital Elevation Model (DEM) and intensity images extracted to estimate road centerlines and curves in a Douglas fir-dominated forest in the McDonald Research Forest in Oregon, USA [6]. Road grades were estimated to an accuracy of 1% of the field value of the slope, and the average error of the curve radii was 3.17 m and the RMSE of the road centerlines 1.21 m in the horizontal plane and 0.28 m in the vertical plane.

Since high pulse data enable more road features to be accessed, 6 pulses per m² LiDAR data were used to extract the position, gradient and total length of forest haul roads in the Santa Cruz Mountains, California, USA, from a 1 m DEM with a positional accuracy of 1.5 m [7]. The road grade was within 0.53% of the true value and road length within 0.2%. LiDAR-derived DEMs from the Lakeland region of Finland were compared by using surface analysis at two pulse densities (11 and 1.1 pulses per m²) [8,9] to identify features of road surface quality such as ruts and holes, and also general road characteristics. In Canada, the vegetation invading forest roads on Vancouver Island was assessed, and the classification method correctly assigned 73% of the 171 km of roads into four vegetation growth categories [10]. In Central Ontario, 1.7-pulse ALS combined with optical imagery with a spatial resolution of 0.4 m (RGB and infrared images) enabled the identification of all the roads [11].

Soils are an essential part of any forest ecosystem, but there has been no extensive research into soil types and their effects on transportation and harvesting. The Finnish Natural Resources Institute has analyzed various aspects of soils with respect to the harvesting and transporting of forest resources, including soil moisture, tyre track-soil interference, bearing capacity and soil deformation, in order to increase forest operations and forest growth and decrease or minimize their environmental impacts and fuel consumption [12]. The Finnish National Forest Inventory (NFI) categorizes forest soil types and particle sizes by means of visual and tactile assessments [13]. The most common Finnish soil types are organic soil, bedrock, stony soil (or boulder fields), glacial till and sorted soil [14].

Other Finnish research is looking for connections between soil types and road quality by means of a machine learning method [15]. Although the method classified roads into four categories according to the potential risks, it also pointed out that road width plays a significant role in forest road quality in Finland. The bearing capacity of an unpaved road has rarely been investigated or assessed, the main accent within forestry having been on the bearing capacity of harvesting areas or paved roads [16] in order to eliminate access problems for heavy machinery. The bearing capacities of Finnish forest roads were measured [17] by using different tools to compare the variations in bearing capacities during the spring thaw period between 2009 and 2012. Road structure, volumetric water content, the groundwater table and stiffness were assessed, and the research set up regression models between the results obtained with the various devices and concluded that cheaper, portable machines can be used on forest roads in most cases rather than costlier falling weight deflectometers.

The Loadman measuring device is a reasonably reliable tool for measuring the stiffness of a road surface, and thus it can be used to estimate trafficability during the thawing season, while the dynamic penetrometer (DCP) is more suitable for defining thawing depths [17]. The Loadman measuring device is also reliable for assessing trafficability if no information is required concerning the subgrade. In Latvia, two methods have been developed for estimating the bearing capacity of roads [18]. The earlier method used only the minimum and maximum percentages of clay in the top layer of the soil, while these new methods consist of a tool suitable for roads with low traffic volumes but high bearing capacity requirements. The first method is based on the geometrical properties of sand and gravel and the percentage of shells in coarse aggregates, and the second on measurements of bearing capacity and compaction level with a light drop weight.

Various wetness indices derived from digital elevation models [19] are often used to describe the effect of topography on soil moisture, but the first hydrological approach to topographic effects was through hydrological forecasting models. The applications of the index include water flow prediction in small- and medium-sized basins, using topographic and dynamic contributory area parameters in Yorkshire, England [20], while another study compared several Topographic Wetness Index (TWI) calculation methods [21] but did not find any of them to be consistently better than the others. Landsat TM images of the Sangay National Park in Ecuador were used to assess the influence of environmental variables, including climate, substrate, topography and anthropogenic disturbances, on the distribution of Andean forests at the tree line [22]. TWI, temperature and precipitation have also been analyzed to determine spatial autocorrelations between standard growth, topographic and climatic factors [23]. Topographic Wetness Models had not previously been used directly for forest road quality assessment but only for modelling forest growth and other ecologically related parameters.

The aim of this study was to test whether LiDAR data combined with soil information and a Topographic Wetness Index can be of help in identifying roads and stretches of roads that need regular check-ups or maintenance due to unsatisfactory drying conditions and/or water accumulation. As Finland has a dense forest road network, highlighting the roads is a priority, as check-ups could reduce the manual workload significantly. The aim was also to test which road quality features are most closely connected to road conditions in terms of the soil type of the surrounding area and its wetness properties, and to consider how these features can be used to evaluate larger areas and road networks and to locate stretches of road that are more affected by water accumulation and therefore have a higher risk of rapid deterioration of the ditch system or road body.

2. Materials and Methods

The study area is situated in Tuusniemi in the Finnish Lakeland region (Figure 1). The field data were collected in July 2014, in accordance with the generally accepted Finnish road-quality classification [24]. The area consists of boreal forests in which the dominant species are Scots pine (Pinus sylvestris L., 34%), Norway spruce (Picea abies (L.) Karst, 52%) and broad-leaved birch (Betula spp., 11%) [25]. The area has low-to-medium elevation differences, with a maximum road steepness of 20%.

Road centerline data for the Tuusniemi area were available from the Digiroad database of the Finnish Transport Agency [26]. This is a shapefile containing the different categories of roads over the whole country, including unpaved forest roads, 356.8 km of which was within the study area.

Several stretches of road had been maintained recently, and 49 field plots had been inventoried, including the recording of several quality parameters in 10 m stretches of road and 5 m of their surroundings. The plots were selected in order to get sample points from each road-quality class category. Random selections of samples would not ensure this, as the poor-quality roads are underrepresented. The field data were collected during the driest period of the year, the months of June, July and August (as were the LiDAR data). Following the Finnish national road-quality classification [24], the following measurements were recorded for the present purposes: structural condition, road surface quality, drying conditions (including ditch quality on both sides of the road) and bearing capacity. Structural condition, road-surface quality and drying conditions were recorded in the form of three categories based on visual assessments incorporating the dimensions (depth, length and width) of the two biggest ruts or holes and a statement of good, satisfactory or poor quality. The depths of the ditches on both sides of the roads were recorded in each field plot.

Several criteria, such as road surface quality, structural condition and drying conditions (including ditch quality) were assessed at each location (

Table 1. Distribution of field plots in the Tuusniemi area into road-quality classes based on a field inventory following the Finnish forest road-quality standards [24]. The categories are explained in more detail above; (n = 14.7).

Tuusniemi (2014)
Road-Quality Class	Structural Condition	Road-Surface Flatness	Ditches and Drying
Poor	3	3	10
Satisfactory	13	22	14
Good	33	24	25
Total	49	49	49

and Figure 1), to determine the overall quality of the unpaved forest road [24]. This quality-assessment system is based on the work of a Finnish research and development company owned by the Finnish forest sector, Metsäteho Oy, to enable it to carry out regular manual checks on the quality of the forest roads. The structural condition of a road refers to its general structure and how sturdy and well-kept it is. If the subgrade is in good condition, the road will require little or no maintenance, while satisfactory quality refers to the emergence of ruts, so that driving requires greater care, and poor structural quality means a reduction in vehicle speed and highly focused attention while driving. Since the inventoried roads had a speed limit of 80 km/h in most sections, the reduction of speed was always expressed in relation to the permitted maximum. The flatness of a road refers here to the smoothness of its surface layer. A good-quality road will have almost no unevenness, while the poorer classes possess more surface variations. In the case of satisfactory quality, the driving speed has to be reduced due to ruts and depressions in the surface, while, under poor flatness conditions, even side bulges may appear, which will further affect the drying quality of the road, as well, and hinder daily transportation by requiring lower speeds in order to avoid vehicle damage. The Ditches and Drying quality class refers to the danger of water accumulation and is concerned not only with major depressions but also with roadside ditches and other structures (culverts, etc.). A good-quality road will show no sign of problems that would interfere with trafficability or involve an immediate need for repair. Poor drying conditions indicate that there are a few problems related to water drainage that may include blocked or non-existent side ditches or major depressions in the road. It is important to note that road quality is a constantly changing parameter, and therefore road structure and condition reflect the need for maintenance.

Bearing capacity and soil strength were assessed by penetrating the dirt roads, using the same force (3500 kN) on each occasion and measuring the penetration depth. This was done 3 times on each plot. For further assessments, we either used the average of these 3 values or categorized them into 3 classes (poor, satisfactory and good) to explore any connections with other variables. For the classification we used the average of three measurements of the penetration depth: under 2.5 cm (good), between 2.5 and 5 cm (satisfactory) and more than 5 cm (poor). We note that this is not one of the standard methods for measuring bearing capacity, and therefore this aspect is not a core part of the present work.

Soil strength can be expressed in terms of penetration resistance and used instead of bearing capacity for characterizing clay, silt and till soils but cannot be applied to sandy soils. Soil strength is highly dependent on the moisture content, bulk density and clay content of the soil. Since soil samples were not collected during the fieldwork, only a general categorization based on soil types was possible, and the assumption was that all the soils were in their driest condition. The soil samples had not been taken for exact measurements, and thus infiltration rates for the soil types could only be estimated. The infiltration rate is classified as high when organic matter such as peat has good porosity, which usually means a good infiltration capacity, as well. Silt and till were classified as having medium rates of infiltration, while the rate for sand was rapid.

Soils can be categorized based on their trafficability according to the soil type concerned and its wetness under boreal conditions [12]. In general, the trafficability of a soil deteriorates as its moisture content increases. Peats and organic soils show poor trafficability even under fairly dry conditions. The following soil types were recognized here, from the worse to the best trafficability: organic soils, fine-grained mineral soils with a thick organic layer, fine-grained mineral soils with a thin organic layer, sandy soils, mid-grained soils and coarse-grained or till soils. Both the fieldwork and the LiDAR sampling took place in summer (July), when the soils were in a dry condition.

Data from the Geological Survey of Finland [27] on maps of scales 1:20,000/1:50,000 were used as a source of information on superficial deposits. These soil maps had resulted from the collection between 1972 and 2007 of a large body of soil samples obtained by using a probing rod and spade, backed up with aerial images and GIS data processing. Sediments at a depth of one meter were mapped on a scale of 1:10,000 and the soils were classified by following the Finnish system [14]. The soils of the study area can be divided into the following types, based on particle sizes (

Table 2. Distribution of soils and their bearing capacity identified in Tuusniemi, Finland.

Soil Type	Additional Properties	Good Quality	Satisfactory Quality	Poor Quality	% Not Showing Good Bearing Capacity	Number of Plots
glacial till	coarse particles	1	0	0	0%	1
glacial till	medium-grained particles	9	8	2	53%	19
glacial till	fine-grained particles	2	0	0	0%	2
sorted soil	fine-grained particles	1	0	0	0%	1
organic soil	sedge and sphagnum peat	4	0	1	20%	5
bedrock	-	9	6	0	40%	15
mixed areas	bedrock with glacial till	6	0	0	0%	6
Total		32	14	3		49

The soil information was obtained from the Geological Survey of Finland [27].

): glacial tills, sorted soils, organic soils and bedrock. Bedrock areas and glacial till soils with medium-grained particle sizes predominate. The organic soils consist of Sphagnum peat or sedge peat. Sorted soils and sediments were also present in the area. The poorest bearing capacities were found in the case of medium-grained glacial tills and on bedrock.

The LiDAR data were recorded at Tuusniemi, Finland, in July 2014, with a pulse density of 1 per m², and the area was mapped between 23 and 30 July 2014, from about 2000 m above ground level, using a Leica ALS50-II laser scanning system at a 20° angle to the field view and with a pulse repetition frequency of 114 kHz, as well as a 20% side overlap. The average sampling density was 1.1 pulses per m².

The DEMs, at two resolutions (25 and 10 m), were created by using the Topographic Database of the Finnish National Land Survey [28], in addition to which a 1 m resolution DEM was generated by using ground points from the LiDAR point clouds by Inverse Distance Weighted (IDW) interpolation [29] in a cell size of 1 m. The resulting low- and high-resolution raster datasets were used later in the hydrological analysis.

The method for assessing forest road quality based on a wetness index and additional variables was implemented in several steps, which are described in more detail below. Firstly, we looked for problematic areas predictable from the soil types, and, secondly, we tested the use of the Topographic Wetness Index as a predictor of road quality. This was divided into several steps and involved assessing TWI itself and in combination with other variables such as soil type. Thirdly, we compared the division into 3 classes with a 2-class classification, with the latter being aimed at identifying only poor-quality roads. In the present work, we also assessed a combination of TWI, soil type and surface quality indices with regard to the determination of forest road quality, in order to see whether problematic locations could be identified with greater accuracy.

The Topographic Wetness Index (TWI) may be viewed as a means of quantifying topographic effects on hydrological processes. There are several methods for computing this index, which differ primarily in the way the upslope contributing area is calculated. As there is no proven best method for performing TWI calculations [21], we followed the workflow of TWI in ArcGIS [30]. Further variables were analyzed regarding soil types, road quality, ditch quality and depth, in order to identify which features, road conditions and classes can be explained by means of TWI. The methods and workflow pattern are explained below. The index can be calculated by using DEMs generated from LiDAR data to identify roads and stretches of road that require frequent maintenance.

The general equation for the Topographic Wetness Index is as follows:

TWI = ln(a/tan B)

where a = upstream contributing area in m² and B = slope.

The index is a number expressed without units, in which higher values mean drainage or depression, while lower values represent crests and ridges. Regarding road quality, we were looking for depressions that were located close to forest roads, as an accumulation of water near a road may damage the road structure.

The DEMs were calculated from the LiDAR ground point cloud by means of IDW (Inverse Distance Weighted) interpolation [29] at a resolution of 1 m, using ArcMap tools. DEMs with other levels of resolution, 10 and 25 m, were available from the Topographic Database of the National Land Survey of Finland for use in making the slope and flow calculations in the later steps. In the account that follows here, the TWI index is calculated for all three spatial resolutions.

Of the various approaches to calculating the Topographic Wetness Index, we opted for the dispersive application of the D infinite method [31] for obtaining the flow direction. This assigns a flow direction based on the steepest slope of a triangular facet, i.e., the steepest downward slope on a 3 × 3 grid for the center cell. The flow direction is marked with angles in radians from 0 to 2 π, starting from the east and running counterclockwise. To avoid calculation errors when having a zero in our raster in the later steps, we added 0.001 to each cell value and then calculated the cumulative flow, using the number of upslope cells flowing towards the given location.

By contrast with earlier studies that have assessed the use of surface-quality indices to determine road quality ([8,9]), the current research made use of the Topographic Position Index (TPI), at a resolution of 1 m, as an additional variable. This TPI was calculated for each cell, by subtracting the average elevation of the cell’s neighbors relative to the cell itself. The neighborhood was set at 3 × 3 m.

After calculating the TWI index, road quality could be predicted by evaluating which features and levels of resolution affected the road quality the most using Linear Discriminant Analysis (LDA) [32,33]. The same method was used to assess the variables and create models for predicting road structure conditions, drying conditions and road surface flatness. The features tested in order to define road quality were the following: TWI index, soil types, trafficability, bearing capacity, ditch depth and surface-quality indices. An LDA with cross-validation [34] was carried out, using the predictor variables and the road-quality factors measured in the field, at varying levels of resolution.

The performance of the TWI indices at three resolution levels regarding road-structure quality was assessed by calculating Cohen’s kappa and the percentage of agreement for each (

Table 3. Correctness of classification, using three road quality classes (poor, satisfactory and good) and various combinations of predictors, such as the wetness index, soil data and surface-quality index. Cohen’s kappa, the percentage of agreement and the agreement after k-fold (k = 10) cross-validation were calculated.

	Predictors	Only TWI Index		TWI + Soil Data			TWI + Surface-Quality Index			TWI + Soil + Surface-Quality Index
Resolution	Predicted Variable	Cohen’s Kappa	% of Agreement	Cohen’s Kappa	% of Agreement	Accuracy (k-Fold CV)	Cohen’s Kappa	% of Agreement	Accuracy (k-Fold CV)	Cohen’s Kappa	% of Agreement	Accuracy (k-Fold CV)
1 m	Flatness	0.54	57%	0.56	59%	0.42	0.29	51%	0.4	0.61	67%	0.4
10 m	Flatness	−1.38	22%	0.56	59%	0.76	−0.26	31%	0.52	0.51	61%	0.78
25 m	Flatness	0.46	49%	0.58	61%	0.74	0.41	45%	0.52	0.62	65%	0.6
1 m	Drying	0.24	29%	−0.73	47%	0.48	0.26	49%	0.8	0.54	67%	0.78
10 m	Drying	−0.48	37%	−0.25	49%	0.43	0.07	43%	0.47	−0.19	49%	0.4
25 m	Drying	0.38	47%	−0.25	49%	0.48	0.21	47%	0.63	−0.13	47%	0.65
1 m	Structure	0.68	71%	0.44	53%	0.75	0.67	69%	0.98	0.69	73%	0.78
10 m	Structure	0.13	27%	0.46	55%	0.72	0.45	55%	0.75	0.49	57%	0.4
25 m	Structure	0.54	61%	0.48	57%	0.75	0.3	43%	0.75	0.48	57%	0.65

). The kappa values represent the following degrees of agreement: under 0.00 poor agreement, 0.00–0.20 slight agreement, 0.21–0.40 fair agreement, 0.41–0.60 moderate agreement, 0.61–0.80 substantial agreement and 0.81–1.00 almost perfect agreement.

3. Results

Unpaved forest roads in the Lakeland region of Finland were primarily assessed by using the Topographic Wetness Index. There is no notable difference between using digital elevation models with 10 and 25 m resolution (Table 3 and

Table 4. Correctness of classification, using two road-quality classes (poor and non-poor) and various combinations of predictors, such as the wetness index, soil data and surface-quality index.

	Predictors	Only TWI Index		TWI + Soil Data		TWI + Surface-Quality Index		TWI + Soil + Surface-Quality Index
Resolution	Predicted Variable	Cohen’s Kappa	% of Agreement	Cohen’s Kappa	% of Agreement	Cohen’s Kappa	% of Agreement	Cohen’s Kappa	% of Agreement
1 m	flatness	0.167	75.5	0.232	89.8	0.219	71.4	0.266	83.7
10 m	flatness	−0.008	30.6	0.232	89.8	−0.031	53.1	0.301	85.7
25 m	flatness	0.028	71.4	0.232	89.8	−0.039	49%	0.232	89.8
1 m	drying	0.378	77.6	0.227	57.1	0.551	81.6	0.589	83.7
10 m	drying	0.086	44.9	0.227	57.1	0.041	39	0.247	59.2
25 m	drying	0.004	38.8	0.227	57.1	−0.009	36.7	0.227	57.1
1 m	structure	0.301	85.7	0.301	85.7	0.052	75.5	0.266	83.7
10 m	structure	0.082	79.6	0.301	85.7	0.082	79.6	0.266	83.7
25 m	structure	0.153	85.7	0.266	83.7	0.266	83.7	0.236	81.6

), although the higher resolution provided significantly better results in several cases. The best predictions were achieved by using 1 m spatial resolution regarding almost all the variables tested. The predictions can be divided into two categories: those in which only the Topographic Wetness Index was used as a predictor and those where other predictors such as soil data and/or surface quality indices were added. Ditch quality and drying of roads cannot be directly explained by the Topographic Wetness Index for the surroundings or by the soil types. The poorest results were recorded when predicting drying conditions in the ditches and on the roads, in which case the use of additional predictors improved the classification.

When only the Topographic Wetness Index was used as a predictor, the general quality of the road structure was predicted best at a resolution of 1 m, with 71% correct classifications, followed by a resolution of 25 m, with 61% correct classifications (Table 3). The addition of soil information led to a better classification only in a few cases, but the use of TWI, soil information and surface indices together improved the classifications, especially when the aim was to determine only the poor-road-quality classes rather than all three classes (poor, satisfactory and good).

We found that surface quality indices can identify road problems with high precision when poor versus non-poor classification was carried out. When operating with three categories, 57% of the roads can be classified correctly, while up to 85% of the poor-quality roads were identified (Table 4) when using the Topographical Position Index in combination with this method. The linear coefficients (TWI, TPI and soil types) for a three-class classification using all the variables to predict the road quality parameters are shown in

Table 5. Combinations of coefficients for each linear discriminant based on the Linear Discriminant Analysis for a three-class classification. Three road-quality criteria (flatness, structure and drying) were predicted, using the wetness index, the surface quality index and different soil types. Each predictor was calculated at three resolutions: 1, 10 and 25 m. LD1: first linear discriminant. LD2: second linear discriminant.

Predictive Variable	Flatness						Structure						Drying
Resolution (m)	1		10		25		1		10		25		1		10		25
Linear Discriminant Coefficients	LD1	LD2	LD1	LD2	LD1	LD2	LD1	LD2	LD1	LD2	LD1	LD2	LD1	LD2	LD1	LD2	LD1	LD2
Wetness Index	−0.006	−0.009	0.05	0.029	0.036	0.103	−0.096	0.058	−0.135	0.022	−0.125	0.067	0.048	0.063	0.154	−0.032	0.153	−0.096
Fine-grained glacial till	3.125	−1.861	2.493	−1.429	2.189	−2.891	2.233	−2.237	2.827	−1.469	2.566	−1.958	1.469	−0.478	−0.827	0.176	−0.877	0.517
Medium-grained glacial till	3.783	−2.218	3.468	−1.273	3.008	−3.052	3.605	−0.564	2.522	1.404	2.8	0.86	3.778	0.942	2.384	−0.278	1.745	−1.211
Fine-grained sorted soil	0.623	−4.515	2.771	−3.338	1.689	−5.135	1.512	−1.809	2.368	1.393	2.577	0.846	0.564	1.158	2.589	−0.318	2.005	−1.226
Bedrock	2.474	−1.767	2.375	−0.941	2.226	−2.359	2.348	−2.695	3.298	−0.509	3.448	−1.109	0.998	−0.142	−0.359	0.063	−0.643	0.117
Bedrock with glacial till	2.961	0.568	4.326	0.472	4.526	−0.688	1.961	−1.487	2.556	1.302	3.011	1.145	0.678	−0.777	1.017	−1.681	−0.146	−2.13
Coarse glacial till	3.154	0.548	4.267	0.889	4.254	−0.758	2.174	−1.404	2.368	1.476	2.278	0.881	−0.406	−4.648	−1.727	−4.514	−2.646	−3.162
Organic soils	3.987	0.438	4.689	−0.928	4.471	−0.793	2.887	−0.904	2.815	0.745	3.078	1.061	0.818	−4.687	−1.652	−4.474	−3.569	−3.257
Surface-quality index	6.894	−0.912	0.061	−0.185	−0.02	−0.01	5.851	4.17	0.005	−0.071	−0.041	−0.021	10.158	−0.283	0.057	−0.006	0.047	0.018

, and those for a 2-class classification in

Table 6. Combination of the coefficients for each linear discriminant based on a Linear Discriminant Analysis for a two-class classification. Three road-quality criteria (flatness, structure and drying) were predicted by using the wetness index, the surface-quality index and different soil types. Each predictor was calculated at three resolutions: 1, 10 and 25 m. LD1: first linear discriminant.

Predicted Variable	Flatness			Structure			Drying
Resolution (m)	1	10	25	1	10	25	1	10	25
Linear Discriminant Coefficients	LD1	LD1	LD1	LD1	LD1	LD1	LD1	LD1	LD1
Wetness Index	−0.002	0.036	−0.022	−0.108	−0.123	−0.112	0.023	0.131	0.143
Fine-grained glacial till	3.335	2.87	3.329	2.814	2.274	2.179	1.429	−0.704	−0.789
Medium-grained glacial till	4.027	3.729	4.105	3.448	2.85	2.943	3.012	1.867	1.741
Fine-grained moraine	1.948	3.834	4.053	2.011	2.699	2.719	0.139	2.04	1.846
Bedrock	2.731	2.58	3.089	3.082	3.018	3.214	0.915	−0.296	−0.332
Bedrock with glacial till	2.438	3.894	4.182	2.3	2.851	3.206	0.829	1.902	1.947
Coarse glacial till	2.615	3.685	3.987	2.462	2.724	2.428	1.063	1.983	2.048
Organic soils	3.385	4.75	4.189	2.924	2.932	3.256	2.144	2.007	1.817
Surface-quality index	6.369	0.126	−0.012	3.768	−0.017	−0.045	8.959	0.044	−0.001

. To test the reliability of the LDA results, Cohen’s kappa was calculated for each LDA analysis. The best models showed substantial or moderate agreement with respect to Cohen’s kappa, but addition of the soil data did not significantly improve the road quality predictions in all cases. Besides Cohen’s kappa, k-fold cross validation (k = 10) was also calculated for the combined data (Table 3).

Considering this two-class classification (Table 4), the methods using the wetness index, surface quality index and soil types as variables provided 81.6–89.8% correct classifications regarding the various road-quality parameters. The surface-quality index, due to its nature as an assessment of the road surface, did not yield reliable results at lower resolution levels, as the roads were an average of 3–4 m in width, so that the surface-quality indices with resolutions of 10 and 25 m did not limit their assessment to the road surface and its immediate surroundings. However, 69% and 73% correct predictions of road quality were achieved with a resolution of 1 m.

The most precise prediction of poor classes was achieved by using all three predictors (TWI index, soil type and surface-quality index) at once. In this case, the road structure was predicted at least 75.5% correctly, but when using the wetness index and soil data, the prediction was over 83% correct, and while using all the variables, it was over 81% correct (Table 4). If we look at the road-structure reference classes, the wetness index performed equally well with or without the surface-quality index. When the 356 km of roads in the area were assessed we found that only 7% of them, about 25 km, were exposed to extremely wet conditions as defined by the Topographic Wetness Index.

Comparing the soil types of the field plots and measurements, we found that bad or only satisfactory bearing capacity was usually connected with medium-grained glacial tills and bedrock areas; however, this did not directly mean that such areas had the worst road quality the whole year round. The roads’ crossing areas with other till soils, on the other hand, were in good shape. Only one out of the five peat areas had poor quality in the road inventories. Conversely, the roads in sandy or rocky areas were in the worst condition, while the road that was surrounded by silt areas was found to be in good condition in summer.

4. Discussion

Unpaved-forest-road-quality information generated from remote sensing data, especially from LiDAR, can help forest owners and managers to allocate resources better for road maintenance. We present here an assessment of forest roads of varying quality in the Lakeland region of Finland for maintenance purposes, in which the Topographic Wetness Index helped greatly in classifying the roads relative to the use of LiDAR data alone. The index can indicate areas where roads are more likely to be damaged and/or to need more frequent maintenance in the years to come. This is important, as Finland’s road network is ageing and deteriorating quickly, and it will be important to identify and renovate the stretches of road for which renovation is most urgent.

There is an increased need for assessing the quality of unpaved forest roads, as their drivability changes over the year much more frequently than does that of paved forest roads. Although ensuring that the necessary machinery can reach harvesting sites via good quality paved roads [16] is an important part of forestry logistics, this paper has discussed the quality, and therefore trafficability, of unpaved roads using LiDAR data analysis and improving the current Finnish forest road-quality assessment standards, which are highly empirical. The predictors constructed here, using a Topographic Wetness Index, classified roads in terms of their quality, using low-resolution data with greater accuracy than those used in earlier studies, where only surface-quality indices were employed [8,9,35].

The method presented here was based on using IDW interpolation. As data processing is a crucial aspect of remote sensing, it should be noted that other interpolation methods may slightly alter the results in both directions. One thing to note is that the sink detection was removed during the DEM preparation step. This may affect the result if a pixel is classified as a sink and removed, which could mean that some potential road beds may not be detected. The classification accuracy achieved was high, however, which means that the results are not notably affected by pit removals. The Topographic Wetness Index is similarly a good means of predicting surface quality indices [8], [9] even when using lower resolution (10–25 m DEMs), while the latter performed best using high- density and -resolution data sources, being able to predict road structure quality at different DEM resolutions with up to 80–89% accuracy. Peuhkurinen and Puumalainen [15] maintain that a machine learning approach could achieve an accuracy of 64–78% in predicting road quality in Southern Finland.

We would like to note some crucial parts of the method. One of these is the timing of reference data and LiDAR data collection. Both took place in summer (2013 and 2014) in our experiment, which is the driest period of the year, so that the roads were in their best possible condition. Most of the road-quality problems are not crucial at that time but rather during the spring, when the snow melts, and in autumn, when there is more precipitation. One of the weaknesses in the method was that the data collection was carried out only once during the year. Although the data collection was compatible with the LiDAR data collection, as both were carried out in summer, when the roads were driest, multiple field data collections could further improve the results, as they would give a more precise picture of road quality throughout the year and the effect of the thaw period.

One clear strength of the method was its applicability to sparse LiDAR datasets. The method was successfully applied to low pulse density LiDAR data to assign unpaved forest roads to quality categories without the need for high pulse density datasets. The assessment of bearing capacity and trafficability must be evaluated carefully due to their dependence upon weather conditions and the significant changes occurring throughout the year. Sandy and rocky soils show good trafficability, and therefore road conditions in such areas do not change significantly during the year, i.e., the majority of such roads remain trafficable in the worst weather, while peat areas are highly affected by the accumulation of water during rainy seasons. The good results can be explained by the fact that extra attention had to be paid to road construction in such cases. Conversely, the roads in sandy or rocky areas were in the worst condition, which may indicate that there is no need to pay attention to the sturdiness of the road structures in such areas, as the surroundings are not greatly affected by water accumulation in the rainy season. Fine-grained soils such as silt have a bearing capacity that is highly dependent on the level of wetness and on particle sizes [36]. The bearing capacity measurements were obtained with only one set of equipment as compared with the wide range of equipment and different timings employed for the comparable measurements conducted by Kaakkurivaara [17].

While the identifying of road features was not within the scope of this project, as the road centerlines were provided by the Finnish Transport Agency [26], roads can be mapped with high precision using similar LiDAR datasets [6,7]. Future research could therefore combine these two crucial aspects: road location and road quality. Identifying roads from remote sensing data reduces the time and costs expended on laborious fieldwork, as the method is based on LiDAR data, which are available from forest inventories (or as lower resolution DEMs from the National Land Survey of Finland) and therefore do not burden forest road owners with the substantial extra costs of data collection. Furthermore, the price of acquiring high quality LiDAR data is likely to drop in the coming decades, making this source accessible to more forest managers and owners, which would in turn allow more detailed road analyses to be carried out.

5. Conclusions

TWI itself is a good variable for describing road structure quality, while a combination of variables can be used for determining the drying conditions on roads or their flatness. This research introduced new methods, using variables derived from LiDAR data, to improve the Finnish forest road-quality assessments and make them less subjective. The results of comparing three spatial resolutions (1, 10 and 25 m) show that the wetness index can help in categorizing unpaved forest roads and can predict poor road quality up to 70% correctly and up to 86% when it is combined with other GIS-based auxiliary variables, such as a surface quality index and soil types. Road-quality assessments using airborne LiDAR data can greatly help forest managers to decide which stretches of road in the network could benefit most from maintenance, while at the same time reducing assessment costs and the need for field visits.