engleski

Development of a Novel Methodology for Vegetation Mapping using Radar Satellite lmagery

Synthetic Aperture Radar (SAR) is an active microwave instrument that has many favorable characteristics. It provides high-resolution, day- and-night, and weather-independent imagery for a multitude of applications, which make it an indispensable tool for earth observation from space. This research evaluates different preprocessing steps of single-date and multitemporal SAR imagery using machine learning methods in order to develop a novel methodology for vegetation mapping. C-band vertical-vertical (VV) and vertical- horizontal (VH) SAR imagery acquired from the Sentinel satellite mission (i.e., Sentinel-1 ; S1) were used for vegetation mapping. In the preprocessing phase, the inevitable drawback of SAR imagery is the presence of noise-like phenomena called speckle. Speckle is a signal dependent noise that strongly degrades the image quality, which leads to image interpretation difficulties. Therefore, this research evaluated different adaptive filters for speckle reduction and estimated impact of filtering techniques on classification accuracy. Since S1 imagery provides data from a dual- polarization (i.e., VV and VH) C-band SAR sensor, including texture information from the grey-level co-occurrence matrix (GLCM), can produce new images by making use of additional spatial information and different land-cover classes, which reflects in improving the classification accuracy. GLCM reveals how often different combinations of pixel brightness with specific values and specified spatial relationships occur. Thus, various texture bands derived from GLCM were tested for vegetation mapping. Furthermore, multitemporal (MT) S1 imagery was used to characterize phenological changes in vegetation land-cover classes since single-date imagery often does not produce satisfactory results. Using MT imagery vegetation mapping, the number of input features is rapidly increasing. Therefore, parameter optimization of machine learning methods will increase vegetation mapping accuracy, as well as reduce the time needed for building a machine learning model. In order to develop the above-mentioned methodology for vegetation mapping, various machine learning methods were used, e.g., Random Forest, Support Vector Machine, boosting classifiers (XGBoost, AdaBoost), artificial neural network (Multi-Layer Perceptron).

Sentinel-1, synthetic aperture radar, vegetation mapping, machine learning, grey-level co-occurrence matrix, speckle filtering, multitemporal, feature selection

nije evidentirano