Woźniak E., Kofman W., Lewiński S., Wajer P., Rybicki M., Aleksandrowicz S., Włodarkiewicz A. (2018), Multi-temporal polarimetry in land-cover classification. International Journal of Remote Sensing, doi:10.1080/01431161.2018.1483084

Research 2018 | 12/09/2018

This study uses time-series Sentinel-1(S-1) synthetic aperture radar images to evaluate the impact of multi-temporal polarimetric processing on land-cover classification. Various polarimetric processing methods are applied to multi-temporal S-1 data set in order to obtain several inputs parameters for land-cover classification: e.g. time-series coherence matrices from dual-polarization data (shows coherence among polarizations in matrix for separated time points t1, t2, to tn); scatter zone time series; multitemporal single and dual-polarization coherence matrices (reveal coherences among time points for one or two polarizations); and parameters from the H/α decomposition. Then, the classification potential of each polarimetric data set is compared to a reference classification, which was derived from time series of dual-polarization backscatter (σ0) images. We evaluate if polarimetric processing of dual-polarization images brings better classification results than alone classification of backscatter image. Finally, we evaluate the impact of segment size and the classifier on classification accuracy.
The classification based on polarimetric data sets is consistently better than that of backscatter time series. A maximum overall accuracy of 93.2% was achieved for the classification of four basic land-cover classes (urban, agriculture, forest, and water) using a composite data set made up time series of scatter zone derived from the H/α plane and scatter zone temporal stability maps. This accuracy was 4.5% better compared to our reference classification based on σ0 time series. Similar trends were observed for more detailed land-cover classes. Classification accuracy is heavily influenced by segment size and can drop by about 15% for very small segments. The most suitable classifier proved to be the Support Vector Machine, which performed up to 12% better than the worst one. This study demonstrates the suitability of multi-temporal polarimetry processing for land–cover classification.