Contribution of VHR Coherence, Complex Intensity, and Sigma0 to the Identification of Urban Structures in Hue, Central Vietnam

Kurzfassung

This study analyzes the information content of various derivatives of VHR radar image products over the city of Hue in Central Vietnam. 24 PAZ images, acquired in Spotlight mode in both ascending and descending orbits were used to compute calibrated radar cross section (Sigma0, n=24), interferometric coherence of different temporal and perpendicular baselines (n=30) and complex intensity which is computed from the combination of real and imaginary parts of the reference and support image of an interferometric pair and contains less speckle and also information from two acquisitions (n=30). These grouped derivatives were used as features to test their contribution for the detection of dwelling density, the allocation of built-up areas and the classification of land use and landcover. Reference data were provided by Vietnamese project partners and values of 5000 reference points were extracted based on a stratified random sampling to include representative training data for all land cover types. Building density (ranging between 0 and 2.6 buildings within a radius of 10 meters): A feature ranking based on the ReliefF measure confirmed that coherence layers have a distinctively higher predictive value for the building density than complex intensity and Sigma0. However, no correlation between perpendicular or temporal baseline and the prediction accuracy has been found. None of the feature groups was able to reliably predict building density based on the 24 images based on linear regression (R² between 0.0 and 0.16). Built-up area (binary: built-up and open): The feature ranking again confirmed the superiority of coherence and complex intensity over Sigma0 data, but indicates no clear trend between baseline, orbit direction and ReliefF. A decision tree based on all three feature groups resulted in a classification accuracy 0.84 and a F1 score of 0.82. The most important feature to split the training data was a coherence product with a temporal baseline of 11 days and a perpendicular baseline of 22 m. While complex intensities were also included in the tree, calibrated Sigma0 was only rarely used to split the samples. Land cover (9 classes: crops, forest, open, paved, public, residential, vegetation, water, other): Coherence-based features resulted in highest ReliefF values compared to complex intensity (slightly lower) and Sigma0 (clearly lower). A random forest classification based resulted in an overall classification accuracy of 32.5 % with strong variation between the classes. The classes with the highest accuracies were water (UA: 83 %) crops (PA: 81.7 %), and residential (PA: 66.8 %), while classes with less distinct physical characteristics (vegetation, public or paved) performed comparably poor. Concludingly, orbit direction and perpendicular baseline did not have significant impacts on the classification results, but coherence (especially combinations of several temporal baselines) were found most suitable to characterize urban features in the case city.