Creating High Resolution Normalized Digital Surface Models through Supervised Deep Learning using Sentinel-2 Data

Kurzfassung

Satellite driven geographic elevation models increasingly have gained importance in terms of city cluster detection and various other human challenges like, for example, natural catastrophes. Those digital models can in particular be used for flood simulations in risk areas. The underlying thesis is dedicated to not measure but generate a high resolution normalized digital surface model - a high informative elevation map - trough deep learning. Although this elevation model is normalized, it does not contain any terrain information and only provides absolutes heights. In order to create one, a neural network has been built, which takes in low resolution satellite data and then converts it into a high resolution surface map. Furthermore, a lot of pre-processing steps are required to be able to feed the satellite data into the network. Thereby, atmospheric corrections and resampling is applied. Additionally, the data is processed and inserted into the network by following a sliding window strategy, that crops smaller frames out of the data, since the model needs to fit the available computational hardware limitations. In terms of the network architecture, Unet will be used. To improve the models performance, many aspects from other network architectures and papers presented in the related work, are implemented. Further, smaller tuneable values are inspired by these. Moreover, the deep learning model will be proposed in different variants, as it will be also investigated, if rather a newly developed classification approach or a commonly regression strategy provides better visual and numerical results. Therefore, height mapping is performed in two of three variants to digitize the height values of surface objects to discrete classes. Doing this, a rough height-mapping model and a finer height-mapping model is developed. Furthermore, a regression model is trained. Those propose the three different variants. Thereby, training the regression approach serves as a comparison between classification and regression. As an evaluation, both approaches deliver proper numerical results, but also both lack in different aspects when it comes to visual height maps. They either lose to much detail in shape of the structures or blur out. However, the degree of detail is not always perfect, both the classification and the regression strategy produce proper results on bigger areas when it comes to creating maps for a rough overview of heights, since arbitrarily large areas can be processed through the sliding window approach. Furthermore, hard borders are avoided using the overlap tile strategy. Also, the absolute height predicted is mostly not far from the truth in the visual tests, verifying the proper numerical results.