Patch-based Image Classification for Sentinel-1 and Sentinel-2 Earth Observation Image Data Products

Kurzfassung

Due the continuous growth of Earth Observation (EO) image data collections acquired from a great variety of sensors, we can observe an increasing need for methods and techniques of querying remote sensing images, not only by their annotations but also by their semantic content. Various methods of content based image retrieval (CBIR) have been proposed in the remote sensing domain, but no general approaches are available. Regardless of the used method, the CBIR systems have the same function, to identify the most similar images with the query image. Some authors developed powerful CBIR tools such as GeoIRIS system of Shyu C. et al. which is a content-based multimodal Geospatial Information Retrieval and Indexing System and includes automatic feature extraction, visual content mining from large-scale image databases, and high-dimensional database indexing for fast retrieval, KIM - knowledge-driven information mining proposed by Datcu M. et al. which is based on human-centered concepts and implements new features and functions allowing improved feature extraction, search on a semantic level, the availability of collected knowledge and interactive knowledge discovery, SemQuery,developed by Sheikholeslami G. et al., which is a semantics-based clustering and indexing approach, used to support visual queries on heterogeneous features of images. Regarding the idea of finding a common ground between synthetic aperture radar (SAR), optical data and even data fusion products, the goal is to develop an application capable to join feature extraction algorithms and classification algorithms. Therefore, this paper is presenting a framework of feature extraction methods for SAR, Multispectral and Data fusion image products that can be used in automatic or semi-automatic classification of urban areas. Our results demonstrate the usability of patch based image classification techniques that can be applied on Sentinel-1 and Sentinel-2 public data sets. Also, another goal is to demonstrate how data fusion products perform in the context of patch based image classification and automatic annotation of urban areas. In order to do so, the selected scene is grouped in a few generic classes like buildings, vegetation, forest, water, streets etc. Then we use feature extraction methods such as Gabor filter banks and Weber Local Descriptors in combination with Support Vector Machine (SVM) and k-Nearest Neighbours (k-NN) to define an application to be tested on SAR, optical data and data fusion products. The result of the study is intended to establish the optimum number of classes that can be found in the Sentinel-1 and Sentinel-2 images when using patch based image classification techniques. Also another important objective of this paper is to determine the best patch sizes suitable for this type of classification and that can be used to obtain the best results for Sentinel-1 and Sentinel-2 EO images.