Estimation of Road Pavement Surface Conditions via Time Series of Satellite Synthetic Aperture Radar Images

Backscattering intensity images derived from synthetic aperture radar (SAR) observations have been used to estimate road pavement surface conditions. Nevertheless, the potential of freely available SAR images with low spatial resolution has not been fully explored. In this study, international roughness index (IRI)-based condition labels (poor or good) are estimated via Sentinel-1 data while considering the influence of vehicles on roads. For intensity, a minimum-intensity image was obtained via an existing image stacking method. For coherence, we propose a method to select data that are relatively unaffected by vehicles within the time window for calculation. When the angle between the SAR range and road longitudinal direction (incident azimuth angle) was small, the difference in coherence calculated with the proposed method between good and poor condition roads was statistically significant. Random forest models were established to estimate condition labels using the IRI values from previous inspections and SAR metrics as inputs. When the incident azimuth angle ranged from 0°–30°, the overall accuracy was 0.800, and the producer's accuracy for the poor class was 0.731 for the model using the coherence calculated with the proposed method as the input. These values were 0.600 and 0.519, respectively, when the coherence without considering vehicle effects was used as the input. Although the accuracy is not as high as that reported for SAR images with high spatial resolution, the present study demonstrates the possibility of a cost-efficient way using freely available SAR images to select inspection targets.