A key element of traffic safety in paved roads is Skid Resistance, a measure of the resistance force of pavement surface to sliding or skidding of the vehicle [1
]. This force is essential for providing tire-pavement grip needed for vehicle control and emergency stopping [2
]. The physical property used to measure skid resistance is termed Dynamic Friction Coefficient (DFC), which quantifies the friction amount between the road and the tire surfaces while the vehicle is moving. The friction coefficient is calculated by dividing the motion frictional resistance by the load acting perpendicular to the interface [3
]. Under wet conditions, newly constructed roads exhibit high friction, whereas older roads may experience aging effects, causing structural damage and compositional alteration of the pavement, resulting with friction loss. These effects are a consequence of mechanical weathering inflicted by passing vehicles, and of environmental effects changing the asphalt pavement’s surface properties (temperatures, dust, rain, snow, and oxidation). Furthermore, friction loss can be caused by tire erosion, leaving skid marks on the road surface, and by oil or gasoline leaks leaving material on the road surface [4
Asphalt pavement is a compound of rocky aggregate and a binder material. The rocky material is usually indigenous in respect to the local mineralogy and is prominently dominated with limestone and basalt minerals in Israel. The binder material is a complex substance mostly comprised by bitumen and polymers. The Bitumen is produced by refining crude oil and has adhesion and water sealing properties. Polymers are added to the bitumen to enhance its qualities and reduce the chance of cracking under load [6
The two major components creating friction in asphaltic roads are adhesion and hysteresis. The adhesion results from the shearing of molecular bonds formed when the rubber tire is pressed into close contact with the pavement surface particles. Changes in asphalt adhesion are caused by the surface’s micro-texture (0.1 to 0.5 mm), which is an aggregate composition related property, as the aggregate ability to retain the roughness against the traffic polishing action controls the rate of decrease in adhesion. Hysteresis is the lagging of peak deformation behind peak load, occurring to the tire after passing across asperities of rough surface, caused by the asphalt’s macro texture (0.5 to 50 mm), created by cracks and fractures in the asphalt [7
Road segments with low DFC contribute to a significantly longer stopping distance for passing vehicles, along with the reduction of the driver’s control over the vehicle. As this could lead to an increasing potential for accidents, DFC is considered the key measured property imperative for defining the road safety [3
]. Usually, state or city authorities are responsible for maintenance operations of the roads within their jurisdiction. In Israel, this is an important endeavor, as insurance companies may claim damages in case of accidents caused by poor road conditions. The maintenance operation itself is a costly exertion with environmental impacts and effects on the transportation grid function. Therefore, an informed decision must take place when considering repairing or replacing a road segment, based on the best available data of the road conditions. To avoid unnecessary operation costs while maintaining transportation safety, an accurate measurement of the road’s conditions must be available for decision makers.
The most common method for mapping the road conditions is by measuring DFC using a test wheel installed on a specialized service car, on which some retardation is applied to cause a skid over the pavement. Using the Dynatest H6875 system, the skid is applied in constant time intervals by an automated breaking system (ABS), and the ratio between the car velocity and the test wheel velocity is correlated to the DFC of the road segment [9
]. To create wet conditions, a water tank is coupled with the service car, spraying water in a constant pressure on the test wheel’s path. As the service car is operating in a constant velocity, the water column stays constant over all measurements [10
This technique directly measures the pavement’s friction and therefore is highly accurate, although the produced data is somewhat limited. The main concern is the minimum operation speed of 60–70 km/h, precluding operation in residential streets in urban areas, as well as increasing sensitivity to ongoing traffic. As this system measures a ~15 cm width section of the road, determined by the test wheel’s dimensions, multiple repetitions of the same road must be performed to achieve a more realistic map—a procedure which is not applied in practice. In addition, operating this system requires highly skilled personnel, as it is comprised of multiple mechanical, electronic, and computational hardware.
Using remote sensing for mapping road conditions have been tested by various academic and industrial groups, by means of different remote sensing technologies. Mainly, these studies were focused on providing a structural measure of the road, locating and mapping holes, gaps and cracks in the pavement. This can be achieved using either a stereoscopic imaging system creating a high resolution 3D map of the road, or by using LIDAR technology, both mounted on a moving vehicle [11
]. Although the available product contains important information for evaluating the road’s conditions, it is not providing the DFC of the road, and requires the use of vehicle to be in-situ. Using hyperspectral remote sensing (HRS) can provide with the chemical and physical information needed to model DFC, and is applicable both in-situ and using airborne platform. As it relies less on the special resolution and more on the spectral resolution, the available product may be superior in terms of coverage and operational productivity.
Works that deal with pavement monitoring using hyperspectral imaging can be divided into two groups—those that uses just the hyperspectral data (unsupervised) and those which uses reference data (supervised). Pascucci et al. (2008) used a thermal airborne multispectral sensor to quantify the road conditions via the limestone absorption feature in 11.2 microns [13
]. This was because the aggregate mixture used in Italian roads is mainly composed of limestone. Mei et al. (2014) applied a combination of digital imaging processing (DIP) and spectral measurements from an ASD spectrometer to quantify the exposed aggregate index (EAI) which is correlated to the amount of bitumen removed from the road surface [14
]. Mohammadi (2012) conducted a research where hyperspectral images were correlated to a road status assessment produced using orthophotos and field visits information. In this work, two types of product were developed—a classification to differentiate between concrete, gravel and asphalt, and a classification to classify three states of the asphalt surface (good, intermediate and bad). A more chemometric approach was applied by Roberts and Harold (2005 and 2008). In this work, spectral characteristics of asphalt under aging effects where discovered in order to qualitatively identify cracks and crevices in the pavement. Additionally, they tried to correlate the spectral information to a pavement condition index (PCI) used by authorities as a standardized method to evaluate the pavement status [4
]. Recently, Carmon and Ben-Dor (2016) showed that a spectral based model to predict the DFC can be extracted using ground spectral measurements, acquired using a car-mounted field spectrometer and geo-referenced friction measurement points [16
]. In this work, the spectral information was modeled to predict the road’s DFC using an artificial neural network (ANN) data-mining technique.
This study aimed at elaborating the use of spectral information over asphalt pavement and use airborne HRS technology to generate continues friction map of roads. The advantageous of airborne HRS can consolidate the product available from the standard method by: (1) covering the entire road in one pass, (2) provide a shorter time gap between data acquisition, (3) provide friction data in previously unaccusable roads (i.e., residential streets, bicycle lanes, remote locations), (4) reduce operation costs. Accordingly, we engaged and demonstrate an operational procedure, were the available data are an HRS image and geo-rectified DFC points, used both for developing the prediction model and for mapping the entire road surface in the scene.
A direct use of hyperspectral pixel data for developing a prediction model to be applied back on the same image is innovative to some extent. The common method is to develop a model based on spectral and chemical/physical measurements taken in a laboratory setting. Using surface measurements taken in situ combined with direct HRS pixel data is unique and may have a higher validity when projecting the model for practical applications. In general, hyperspectral airborne data may hold different spectral signatures than data acquired at the laboratory (e.g., dust accumulates on the surface) along with different quality (e.g., less SNR). Moreover, the atmospheric correction routine applied on the airborne image may add noise to the data. Even if the laboratory data is carefully resampled to simulate the spectral configuration of the airborne sensor, major differences still may be found between the datasets. In addition, the airborne system has a much larger surface footprint compared to the laboratory data. Accordingly, upscaling laboratory models to image data may be problematic and hence, to ensure an optimal performance, modeling has to be generated and applied directly with and to the image data itself.
Although there was a time gap of 3–4 months between the friction measurements and the hyperspectral campaign, the spectral assignment data extracted from the developed model suggests constant mechanisms over this time. This was confirmed by the Israeli Road Authorities Engineers, but has to be considered in any future work. The suggested mechanisms are in coherence both with works specific to asphalt erosion, as well as to works in reflectance spectroscopy and asphalt aging effects. Two distinctive processes with spectral registration were found: (1) binder erosion exposing the aggregate and (2) material from the environment adhering to cracks and holes in the pavement.
The advantages of using imaging spectroscopy over using in situ car mounted field spectrometers are two-fold. First, the obtained airborne data holds a larger dynamic range, as the car-mounted spectrometer has a very small surface footprint (160 cm2), and the measurement is taken place during movement. This leads to a limited dynamic range in the data, as the radiation amount coming into the sensor is small. Second, while the car system's geometry creates non-constant illumination conditions on the surface, as well as operational only when driving from south to north (do avoid shade in the northern hemisphere), using an airborne system is much more realistic.
As we demonstrated a controlled modeling process design to reduce bias, the reported prediction capabilities can be considered valid. Providing both a full-cross validation and an internal validation results was done for comparison purposes with future studies, which may use either type of validation. The decision to withdraw from any preprocessing routines came in the light of our ambition to represent the data and the application potential as realistically as possible.
Some limitations of using this technique should be pointed out. First, the standard criteria for friction measurements accuracy is much higher than the accuracy demonstrated in this work. A standardized measuring method should be accurate and constant, whereas the suggested method in this study is still not accurate enough. Nonetheless, a significant increase in accuracy could be obtained via a number of key points. Foremost, reducing the time gap between friction and spectral measurements will ensure the absence of changes to the road surface and the representation of the same reality in both datasets. These changes may be due to sand and dust setting on the pavements’ surface, as well as oil and gasoline leaks from deriving vehicles washed off the road. Moreover, although in this work we successfully modeled the systematic aging effects of the pavement affecting friction loss, the reduction of the pavement’s friction as measured by the DFC service car is related also to cracks and holes in the pavement, which are not detectable in a 1 × 1 m/pixel resolution and requires a higher magnitude in resolution. As a possible solution for this obstacle which may increase the overall accuracy of the prediction may be combining a high resolution RGB camera with the hyperspectral system. An elaborate yet expensive solution is to have a LIDAR sensor and a hyperspectral camera on the same aircraft. The LIDAR is capable of providing a very high resolution digital terrain model (DTM) and identify locations with structural damages to the pavement as needed. Moreover, as the LIDAR and hyperspectral sensors are covering the same area, the latter may benefit with a much better geometric correction and improve the overall data quality. The second limitation of this technique is the alteration of mineral aggregate and binder mixture between different roads in different locations. Although in Israel the aggregate mixture is relatively constant, the binder composition may change between asphalt factories. Because the prediction model is based on spectral features of the observed materials, one model may not be correct for all roads. Although a more robust model may be developed, the tradeoff will be a more generalized determination and a limit to the prediction accuracy. Therefore, for using this suggested method in a large scale mapping survey, multiple calibrations are needed in order to account for the different road types.
In spite of the mentioned limitations and disadvantages, using this approach as is, will provide a major upgrade to the data available today using the traditional method. The produced maps are far superior in terms of resolution and continuity, and despite of the limit in prediction accuracy they provide important information about the road's friction. Additionally, the operational complexity and financial resources needed for using this technique are minor relatively to the traditional technique. Performing an airborne hyperspectral campaign to map an entire metropolitan is a matter of a few weeks, compared to a few months using the friction measurement vehicle. In addition, the airborne system has no accessibility limitations and can operate in urban environments and other inaccessible locations of the DFC measuring vehicle.