Evaluation of PolInSAR Observables for Crop-Type Mapping Using Bistatic TanDEM-X Data

Abstract

The contribution of Polarimetric SAR Interferometry (PolInSAR) observables to crop-type classification is investigated in this letter. The focus is set on characteristic parameters of the Coherence Region (CoRe), i.e. the representation in the polar plot of the PolInSAR data. For this purpose, time series of dual-pol HH-VV single-pass TanDEM-X bistatic data acquired over an agricultural area in Spain are exploited. In the experiment, up to 13 different crop types are evaluated. Crop classification is performed by means of the well-known Random Forest algorithm. The retrieved accuracy metrics highlight the potential of the evaluated PolInSAR descriptors for this application. Some PolInSAR features have proven to be enough representative of the scene, such as the Trace Coherence, which yields a classification accuracy of 75% and 87% at pixel and field level, respectively, on its own. Using all the PolInSAR parameters jointly as input features, classification reaches around 90% and 94% accuracy at pixel and field level, respectively. However, there are some PolInSAR feature subsets, e.g. the coherence measured at the Pauli channels or the foci of the ellipse which represents the CoRe, which yield accuracy levels very close to these maxima. These results demonstrate the suitability of the PolInSAR parameters for crop-type classification. Results are further improved when both polarimetric and PolInSAR features are combined, reaching 94% and 96% accuracy at pixel and field level, respectively.