Crop classification with polarimetric syntheticaperture radar images: comparative analysis

Abstract

Polarimetric Synthetic Aperture Radar (PolSAR) images could provide beneficial information regarding the complete scattering about the objects or targets and this could be advantageous to derive the physical and geometrical structure. Due to the benefits of the imaging capability day/night and weather-independent, Synthetic Aperture Radar (SAR) sensors are of vital importance for time-critical practices, especially in agricultural applications. In specific to agricultural practices, multi-temporal or time series data is a pre-requisite for timely monitoring or identification of crop pattern. This is because crops have a dynamically changing structure in temporal domain. Each crop has different structural and physical changes in temporal domain and the use of multi-temporal data leads to better separation of crops. The PolSAR data by itself (2×2 complex Sinclair scattering matrix) do not explicitly/directly provide the “ready-to-use” information about the three elementary scattering (surface, double bounce and volume scattering) for natural targets and the data needs to be converted to second order statistical formalism (3×3 complex matrices) for extracting the scattering properties. In such a case, polarimetric decomposition methods can be used to extract the three elementary scattering for the targets precisely. In this thesis, the comparative performance of the original features (linear polarizations and coherency matrix) and polarimetric features (incoherent polarimetric decompositions) from multi-temporal PolSAR data was investigated for crop pattern identification through three different machine learning algorithms (Light Gradient Boosting Machine, Support Vector Machine and Random Forest). In order to create the polarimetric features, three different incoherent polarimetric decompositions were utilized as follows: Cloude-Pottier decomposition (eigenvector-based), Freeman-Durden decomposition (model-based) and Van Zyl (hybrid) decomposition. Among these machine learning algorithms, Light Gradient Boosting Machines was recently introduced to machine learning community and have not been much explored in remote sensing for classification purposes. The experimental results demonstrated that highest classification accuracy (0.96) were received by Van Zyl decomposition as well as Freeman-Durden through LightGBM. The results also addressed that LightGBM is much faster (almost ten times) than RF and SVM for linear polarizations, coherency matrix and Cloude-Pottier decomposition. This thesis also highlights the benefits of model-based and hybrid decompositions about obtaining the higher performance in comparison to original features for crop pattern classification.