Share Email Print
cover

Proceedings Paper

An iterative approach to optimize change classification in SAR time series data
Author(s): Markus Boldt; Antje Thiele; Karsten Schulz; Stefan Hinz
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

The detection of changes using remote sensing imagery has become a broad field of research with many approaches for many different applications. Besides the simple detection of changes between at least two images acquired at different times, analyses which aim on the change type or category are at least equally important. In this study, an approach for a semi-automatic classification of change segments is presented. A sparse dataset is considered to ensure the fast and simple applicability for practical issues. The dataset is given by 15 high resolution (HR) TerraSAR-X (TSX) amplitude images acquired over a time period of one year (11/2013 to 11/2014). The scenery contains the airport of Stuttgart (GER) and its surroundings, including urban, rural, and suburban areas. Time series imagery offers the advantage of analyzing the change frequency of selected areas. In this study, the focus is set on the analysis of small-sized high frequently changing regions like parking areas, construction sites and collecting points consisting of high activity (HA) change objects. For each HA change object, suitable features are extracted and a k-means clustering is applied as the categorization step. Resulting clusters are finally compared to a previously introduced knowledge-based class catalogue, which is modified until an optimal class description results. In other words, the subjective understanding of the scenery semantics is optimized by the data given reality. Doing so, an even sparsely dataset containing only amplitude imagery can be evaluated without requiring comprehensive training datasets. Falsely defined classes might be rejected. Furthermore, classes which were defined too coarsely might be divided into sub-classes. Consequently, classes which were initially defined too narrowly might be merged. An optimal classification results when the combination of previously defined key indicators (e.g., number of clusters per class) reaches an optimum.

Paper Details

Date Published: 18 October 2016
PDF: 11 pages
Proc. SPIE 10005, Earth Resources and Environmental Remote Sensing/GIS Applications VII, 100050X (18 October 2016); doi: 10.1117/12.2240941
Show Author Affiliations
Markus Boldt, Fraunhofer IOSB (Germany)
Karlsruhe Institute of Technology (Germany)
Antje Thiele, Fraunhofer IOSB (Germany)
Karlsruhe Institute of Technology (Germany)
Karsten Schulz, Fraunhofer IOSB (Germany)
Stefan Hinz, Karlsruhe Institute of Technology (Germany)


Published in SPIE Proceedings Vol. 10005:
Earth Resources and Environmental Remote Sensing/GIS Applications VII
Ulrich Michel; Karsten Schulz; Manfred Ehlers; Konstantinos G. Nikolakopoulos; Daniel Civco, Editor(s)

© SPIE. Terms of Use
Back to Top