极化SAR具有全极化观测和不受天气条件影响的优点，是一种重要的对地遥感手段，其在地物分类应用领域受到广泛关注。本文以极化SAR图像的地物分类作为主要应用背景，对极化SAR图像的斑点滤波、图像配准、极化分解、地物分类方法等关键问题展开研究。本文的主要工作和创新点总结如下： 1) 围绕极化SAR图像的斑点滤波、图像配准的图像预处理问题，本文首先在极化差异特征表达方面提出了Wishart梯度，所提梯度可良好的适应极化SAR图像斑点噪声并表达局部差异特征的幅度和方向。然后，围绕斑点滤波问题，结合偏微分方程方法，提出了基于Wishart梯度与各向异性扩散的滤波方法，可兼顾纹理细节保持与斑点噪声滤除，提高滤波效果。围绕图像配准问题，结合SIFT类方法，提出了Wishart-SIFT配准方法，在多视角、多波段配准情况下提高了配准稳定性与配准精度。实验验证了所提滤波与配准方法的有效性，显著优于对比方法。 2) 围绕极化SAR的极化分解问题，本文针对体散射模型建模问题，结合散射体的不同形状与取向分布，提出了一种新体散射模型。新体散射模型包含的四种情况互为补充，可更全面的描述不同场景下的体散射模型，尤其是可在短波长下较好的建模来自于植被冠层表面的散射。进一步，本文结合基于优化方法的极化分解策略，利用所提体散射模型作为候选散射模型，提出了基于散射体形状与取向分布的$L_1$约束优化极化分解方法。通过仿真与实测数据上的实验，验证了所提分解方法的有效性，所提方法可在散射机制混合区域以及短波长体散射区域较好的自适应选择合适的散射模型，取得较好的分解效果。 3) 围绕极化SAR图像地物分类问题，本文分为单波段地物分类与多波段融合分类两种情况展开研究。在单波段地物分类问题中，针对模型感受野问题，本文引入Transformer模型及自注意机制，提出了基于Vision Transformer的地物分类方法。针对有标签样本量少的问题，引入遮罩自编码器方法，实现无标签数据的模型预训练。在多波段融合分类问题中，本文在自注意模块中引入了多波段的信息交互，提出了多波段交叉注意机制，并进一步提出了基于交叉注意力的多波段融合分类方法。通过实验验证，表明了所提分类方法在分类精度方面的优势，并讨论了模型超参数以及预训练对分类性能的影响。

Polarimetric SAR (PolSAR) has the advantages of full polarization observation and is not affected by weather conditions. It is an important earth remote sensing system and has received widespread attention in the field of land cover classification applications. This dissertation takes the ground object classification of PolSAR images as the main application background and focuses on key issues including despeckling, image registration, polarization decomposition, and land cover classification methods of PolSAR images. The main work and innovations are summarized as follows. 1) Aiming at the despeckling and image registration of PolSAR images, a new polarization difference feature expression method called Wishart gradient is proposed, which can not only well adapt to the speckle noise, but also express both magnitude and direction of the local differences in PolSAR images. Then, focusing on the despeckling problem, combined with the partial differential equation method, a despeckling method based on Wishart gradient and anisotropic diffusion is proposed, which can maintain texture details and filter out speckle noise. Focusing on the problem of image registration, combined with SIFT methods, the Wishart-SIFT registration method is proposed, which improves registration stability and accuracy in the case of multi-view and multi-band registration. Experimental results verified the effectiveness and superiority of the proposed despeckling and registration method. 2) Focusing on the problem of polarization decomposition, a new volume scattering model combining the different shapes and orientation distributions of scatterers is proposed. The four situations included in the new model complement each other and can more comprehensively describe the volume scattering model in different situations especially the scattering from the vegetation canopy surface at short wavelengths. Furthermore, an $L_1$ constrained optimization polarization decomposition method based on the scatterer shape and orientation distribution is proposed. The proposed decomposition method introduces the polarization decomposition strategy based on the optimization method, and utilizes the the proposed volume scattering model as candidate scattering models. Experimental results on simulation and measured data demonstrate the effectiveness of the proposed decomposition method. The proposed method can adaptively select the appropriate scattering model in the scattering mechanism mixing region and the short-wavelength volume scattering region, and achieve better decomposition effect. 3) Focusing on the problem of PolSAR image land cover classification, this dissertation conducts research in two situations: single-band land cover classification and multi-band fusion classification. For the single-band land cover classification problem, to address the issue of model receptive field, the Transformer model and self-attention mechanism are introduced, and a Transformer-based land cover classification method is proposed. Aiming at the issue of the small number of labeled samples, the masked autoencoder method is introduced to achieve model pre-training on unlabeled data. For the problem of multi-band fusion classification, multi-band cross-attention mechanism is proposed to introduce multi-band information interaction in the self-attention module. Then, a multi-band fusion classification method based on cross-attention is proposed. Experimental results demonstrate the advantages of the proposed classification method in terms of classification accuracy. The impact of model hyperparameters and pre-training on classification performance is also verified and discussed.