Automatic snow layer detection in drone-borne radar data using edge detection and morphology

Abstract

The thesis aims to detect the primary interfaces in ground-penetrating radar (GPR) data collected from a snow-pack. An airborne drone was used to collect the data, where a 2D image of the substructures was gattered, including GPS and laser altimeter data. Al these were used under the thesis to develop the method or presentation of the results. The method focused on simpler image processing techniques where more complicated methods would be explored if needed. Ground truth was drawn manually with guidance from a GPR expert.

The primary method used in this thesis was Canny edge detection and morphological operators. Two different techniques were used to detect the two different layers because they showed significantly different characteristics. The technique for the top layer resulted in a root mean square error (RMSE) accuracy of 5 cm, which was within the range resolution of the radar system was achieved. A quality estimate was also given to the top layer, indicating the top estimate's quality found through our method. The bottom estimate showed an accuracy of 20 cm because of the complexity of the bottom layer. On the other hand, the method did have a cross-correlation of 0.9, meaning it could follow the bottom layer in most datasets, but it could struggle to have the exact location correct.

In short, the method presented could be applied routinely to estimate the primary interfaces in other GPR data, where no method previously existed.