The invention discloses a polarimetric SAR ground object recognition method based on learning superpixel and QCNN. The method comprises: acquiring polarimetric radar data; performing polarization decomposition on the polarized radar data to obtain a polarized radar pseudo-color image; classifying the polarized radar pseudo-color images based on a quaternion convolutional neural network of pixels to obtain a classification result; generating a plurality of super-pixel blocks by using a PAN network based on the polarized radar pseudo-color image; and generating a classification result of the polarized radar data based on the classification result of the polarized radar pseudo-color image and the superpixel block. According to the method, high-level semantic features in a color space are utilized, and local region segmentation better than that of a traditional superpixel generation method can be obtained. In combination with quaternion convolutional neural network pixel-level classification and PAN superpixel segmentation regions, a high-precision polarimetric SAR image terrain classification effect based on color features is obtained.