NEW STATISTICAL METHODS FOR SATELLITE IMAGE CLASSIFICATION

Abstract

Nowadays remote sensing of the Earth is one of the most efficient methods for acquiring information about the state, properties and spectral characteristics of the Earth’s surface objects. Classification of satellite images is a main task in remote sensing, that contains the assignment of each image pixel to a specific thematic class based on its spectral characteristics. The result of classification procedure is a thematic map, that shows the distribution of various surface types. Statistical classification methods are based on analysis of spectral characteristics of image pixels and their assignment to land cover classes. The maximum likelihood method is one of the most dependable and precise supervised classification methods. The main assumption of this method is that spectral characteristics of pixels within each class follow a normal (Gaussian) distribution. But it was noted, that this assumption is correct only for homogeneous natural surfaces such as water bodies or agricultural fields. This assumption is not valid for mixed cover types and heterogeneous urban territories. It was conducted a comparative analysis of maximum likelihood method applying three statistical distributions for representing spectral characteristics of land cover classes: normal distribution, exponential distribution, and Weibull distribution. The maximum likelihood criterion was applied for classification procedure and formulas of the log-likelihood functions were considered for these three distribution types. It was noted, that normal distribution is effective for classes with homogeneous structure such as, water surfaces, dense vegetation and paved roads. The exponential distribution is optimal for radar imagery classification. It describes intensities of radar signals from rough surfaces. The Weibull distribution is effective for heterogeneous classes with variable texture such as urban territories. This distribution is commonly used to model the amplitude characteristics of radar signals. The choice of statistical distribution should be justified for each specific land cover class. It was shown that exponential distribution is a special case of Weibull distribution. Future research directions include applying ensemble approaches combining various distribution types for different classes and integration with machine learning methods