“…For each object, a small region of 75 × 75 pixels is extracted around the central object coordinate, such that it covers the object extent including wakes. The same region is now extracted for the 4 high-resolution bands m = 1, 2, 3, 4 (blue, green, red, and near-infrared) with spatial resolution 10 m. The other 6 bands with 20 m and the 3 bands with 60 m pixel resolution are not used in this analysis because they have less spatial information (see however [25]). For convenience, we combine the red and near-infrared bands m = 3, 4, to have the 3 color images as is commonly used for image recognition in neural nets.…”