Özet:
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.
