Palarimetric Synthetic Aperture Radar Image Classification using Bag of Visual Words Algorithm

Land cover is defined as the physical material of the surface of the earth, including different vegetation covers, bare soil, water surface, various urban areas, etc. Land cover and its changes are very important and influential on the Earth and life of living organisms, especially human beings. Land cover change monitoring is important for protecting the ecosystem, forests, farmland, open spaces, water quality, wildlife habitat, and human health. Uncontrolled changes in land cover (deforestation, extensive agricultural activities, vast urban and exurban development, etc.) will not only destroy the natural environment and the ecosystem but also will affect different aspects of human life including the economy. As a result, various effective methods for land cover monitoring are always deliberated by researchers. Remote sensing based techniques, due to their unique capabilities and extensive data availability, have attracted great interest for decades. In addition, fully Polarimetric Synthetic Aperture Radar (PolSAR) images, due to the rich data and the ability of day-night and all weather condition data acquisition, has got high capability to study and monitor land cover changes. Recent developments of SAR sensors in acquiring fully polarimetric, continuous, and very high resolution data, besides enabling precise land cover monitoring, has necessitates the usage of more powerful and robust algorithms in different image processing steps. Classification algorithms are among the most fundamental tools of remote sensing image processing; and an increasing number of researches have been focused on developing novel and more robust classification methodologies for remote sensing data in recent years. Mid-level representations have also been introduced to the remote sensing data classification to enhance the reasonability and semantic rationality of the land cover classifications through remote sensing techniques. Bag of Visual Words (BOVW) model, originally inspired by the Bag of Words (BOW) model from text mining, is one of the most authoritative mid-level representation models which has achieved state-of-the-art results in different image processing tasks. BOVW representation model has recently been introduced to remote sensing data processing and proved to be very beneficial for high resolution remote sensing data representation. In the present paper, a novel segment-based BOVW framework has been developed for PolSAR data representation and classification. The obtained classified maps are quantitatively and qualitatively compared with the result of classification with well-known classifiers including Support Vector Machine (SVM), Artificial Neural Network (ANN), and Wishart Classification Algorithms. The main contributions of the current study are (i) covering the semantic gap between low-level features extracted from the PolSAR image and high-level concept of land cover in remote sensing data classification (ii) introducing novel segment-based BOVW framework for PolSAR image classification, and (iii) extensive evaluation of the performance of the BOVW model for PolSAR data classification in terms of the model parameters and low-level features selection. The experimental dataset is acquired by RADARSAR-2 from San Francisco Bay, California, in C band and at fine quad-pol mode, in 2008. The achieved overall accuracy, 90.1%, illustrates the capability of the proposed model for SAR image classification purposes. Moreover, the proposed segment-based framework for the BOVW model has decreased the undesired speckle effect of the SAR data, without speckle filtering.