Deep Convolutional Neural Networks for Semantic Segmentation of Multispectral Sentinel-2 Satellite Imagery: An Open Data Approach to Large-Scale Land Use and Land Cover Classification

In this thesis, the applicability of deep convolutional neural network s (CNNs) for large-scale land use and land cover (LULC) classification is evaluated. A state-of-the-art image recognition CNN architecture was adapted and re-trained from scratch using a novel dataset. LULC classification is a common task in remote sensing. Large-scale LULC maps are mainly used for scientific analyses and serve as a basis for decision making by governments and non-governmental organizations (NGOs). Several products with supranational to global coverage have been presented. However, they involve a large amount of manually labeled data, which remarkably limits the ability to frequently update the maps. Approaches for automatic classification of remote sensing imagery exist, but do not scale well to large-scale applications. Recently, novel methods for image recognition and semantic segmentation were developed in the computer vision domain. As these tasks are closely related to the problem of LULC classification, they might help to overcome current limitations. Various approaches were explored and a suitable CNN-based approach was selected for prototypical implementation. In order to train a deep CNN, vast datasets are required. Since there are currently no suitable datasets in the remote sensing domain, the creation of a custom dataset is necessary. Manually labeled reference data is, however, prohibitively expensive. Hence, a new method for deriving a dataset for training from open data is proposed. This includes, in particular, multispectral Sentinel-2 imagery and volunteered geographic information (VGI) extracted from the OpenStreetMap (OSM) database. The produced dataset, called openLL, contains more than 500 000 samples for ten classes. Image patches were extracted from 451 Sentinel-2 images, acquired over Germany in 2016. Labels containing class-wise confidences were derived from five monthly snapshots of the OSM database. A deep image recognition CNN architecture was successfully trained on this custom dataset. Based on a prototypical implementation, the performance of a CNN trained for LULC classification was assessed. Experiments on the error propagation from corrupt training data to classification results were conducted. Analyses revealed that the developed classifier is robust to considerable amounts of defective training examples. This is a notable, because it validates using error-prone VGI as reference data. The achieved average recall of 45% and overall accuracy of 51% for a nine-class classification scheme proves the viability of this approach. Classification results further improved with additional experiments on multiview approaches, displaying potential for future work.     «

In this thesis, the applicability of deep convolutional neural network s (CNNs) for large-scale land use and land cover (LULC) classification is evaluated. A state-of-the-art image recognition CNN architecture was adapted and re-trained from scratch using a novel dataset. LULC classification is a common task in remote sensing. Large-scale LULC maps are mainly used for scientific analyses and serve as a basis for decision making by governments and non-governmental organizations (NGOs). Several pr...     »