Statistical analysis of airborne imagery combined with GIS information for training data generation

In recent years, the task of land cover classification from airborne image and elevation data advanced considerably due to enhanced applicability of CNNs (Convolutional Neural Networks). Nevertheless, CNNs require a huge amount of training data. Traditionally, few essential feature values, such as elevation or vegetation index, had been chosen to provide a coarse distinction of classes, but very often these values have to be adapted depending on the imagery. To improve this process, freely available GIS data are combined with spectral and spatial features (and their variations) following the K-Means and Mean-Shift algorithm. Based on cluster assignments to pixels, statistical analysis for extracting plausible values for distinguishing between land cover classes is applied. The resulting labeled databases are evaluated using ground truth data, and will form the basis for the training data required for CNNs.