A Novel Decision Tree Algorithm for Land Cover Classification Using Hybrid Polarimetric Sar Data

Abstract

The polarimetric information contained in polarimetric synthetic aperture radar (SAR) images represents great potential for characterization of natural and urban surfaces. However, it is still challenging to identify different land cover classes with polarimetric data. Hybrid polarimetric SAR data (RH, RV) from RISAT - 1 is found to be suitable for land cover classification of significant features that are well distinguished. The availability of high resolution hybrid polarimetric data from RISAT - 1 SAR systems supporting FRS -1 mode, made it possible to analyze the scattering mechanism for different land use and land cover features using the Raney decomposition (m-alpha, m-chi, and m-delta) techniques. Further to perform both supervised classification(parallelepiped, minimum distance, maximum likelihood and isodata classifiers) and machine learning (artificial neural net) classification also performed Decision tree classification.The proposed statistical Gumbel distribution model has been implemented and retrieves the threshold intensity values. In this proposedwork classification approach has been evaluated for RISAT-1 SAR hybrid polarimetric data of 21^st October 2014 over an urban city, Visakhapatnam, in the state of Andhra Pradesh, India. Since the hybrid polarimetric radar data contains all the scattering information for any arbitrary polarization state, data of any combination of transmitting and receive polarizations can be synthesized, mathematically from hybrid polarimetric data. The RISAT-1 SAR hybrid polarimetric data were decomposed to retrieve the surface and volume scattering information. Both supervised classification and machine learning classification methods were appliedto land cover and few other land use classes based on ground truth measurements using maximum-likelihood (ML) distance measures that are derived from the complex distribution of SAR data at various polarization combinations. The results show that Decision tree classification accuracies for m-alpha, m-chi and m-delta methods were 99.743, 96.873 and 99.857 respectively.  RISAT-1 hybrid polarimetric SAR data helps to classify land cover features efficiently.