2.3.1. Image Pre-Processing

To allow for the comparison of the classification results via spatially consistent image data, the recorded RGB and multispectral images were processed by manually registering 12 reference points (overlapping geometries and artificial black-and-white targets) for thin plate spline [54] transformation (TPS) with nearest neighbour resampling, carried out with the Georeferencer plugin in QGIS (v2.18). This resulted in overlying images, which are necessary to assess the quality of the final results. All non-sediment objects captured in the images were masked manually. Aside from external disturbances like grass patches and other vegetation, this included the prehistoric stone layers that are separating the sediment layers from one another (Figure 2). This step was conducted using the raster package for R [55]. Minimising the impact of image noise and smaller disturbances within the profile, the images were filtered spatially using the raster package for R [55]. To keep the geometric alteration as low as possible and, likewise, to minimise the influence of outliers, we chose a median filter with a window size of 21 by 21 pixels (multispectral data) and 101 by 101 pixels (RGB data) for spatial filtering. The window size was scaled according to the respective spatial resolution and set to quite high values, since our main interest was in the main strata. The sediment layers examined did not include flakes of, e.g., brick, charcoal, or mortar, which renders the loss of texture detail in the image data acceptable [12].
