Taleghanidoozdoozan, Saeid2023-08-142023-08-142023-08-142023-07-06http://hdl.handle.net/10012/19681The RADARSAT Constellation Mission (RCM) utilizes compact polarimetric (CP) mode to provide data with varying resolutions, supporting a wide range of applications including oil spill detection, sea ice mapping, and land cover analysis. However, the complexity and variability of CP data, influenced by factors such as weather conditions and satellite infrastructure, introduce signature ambiguity. This ambiguity poses challenges in accurate object classification, reducing discriminability and increasing uncertainty. To address these challenges, this thesis introduces tailored spatial models in CP SAR imagery through the utilization of machine learning techniques. Firstly, to enhance oil spill monitoring, a novel conditional random field (CRF) is introduced. The CRF model leverages the statistical properties of CP SAR data and exploits similarities in labels and features among neighboring pixels to effectively model spatial interactions. By mitigating the impact of speckle noise and accurately distinguishing oil spill candidates from oil-free water, the CRF model achieves successful results even in scenarios where the availability of labeled samples is limited. This highlights the capability of CRF in handling situations with a scarcity of training data. Secondly, to improve the accuracy of sea ice mapping, a region-based automated classification methodology is developed. This methodology incorporates learned features, spatial context, and statistical properties from various SAR modes, resulting in enhanced classification accuracy and improved algorithmic efficiency. Thirdly, the presence of a high degree of heterogeneity in target distribution presents an additional challenge in land cover mapping tasks, further compounded by signature ambiguity. To address this, a novel transformer model is proposed. The transformer model incorporates both fine- and coarse-grained spatial dependencies between pixels and leverages different levels of features to enhance the accuracy of land cover type detection. The proposed approaches have undergone extensive experimentation in various remote sensing tasks, validating their effectiveness. By introducing tailored spatial models and innovative algorithms, this thesis successfully addresses the inherent complexity and variability of CP data, thereby ensuring the accuracy and reliability of diverse applications in the field of remote sensing.enradarsat constellation missionspatial modelcompact polarimetrymachine learningdeep learningconditional random fieldsignature ambiguityoil spillsea iceland coverDoctoral Thesis