Attention based multi-level and multi-scale convolutional network for PolSAR image classification

Due to presence of heterogenous regions with materials and objects with different shapes and sizes in natural scenes, there are various contextual information in polarimetric synthetic aperture radar (PolSAR) images, which can be highlighted in the low-, medium-, or high-level features in different scales. To handle these challenges, the multi-scale and multi-level attention learning (MMAL) network is proposed for PolSAR image classification. A convolutional neural network (CNN) with six convolutional layers is introduced for hierarchical extraction of local contextual features in multiple levels. The cross-attention is used to find the relationships among low-level and high-level features and also among medium-level and high-level features. This process is repeated in multiple scales. Finally, the attention based multi-level and multi-scale features are fused to provide the classification map. An ablation study is done in several PolSAR images to show impact of different parts of the proposed network, which shows the superior efficiency of feature fusion in multiple levels and scales with taking to account the cross-attention among low–high and medium–high levels. The proposed MMAL network generally provides improved classification results compared to a CNN with the same structure and settings. For example, for the AIRSAR Flevoland image containing 15 class, with using 100 training samples per class, the overall accuracy of 96.15% with 2.74% increment with respect to the basic CNN is achieved where this improvement is statistically significant in term of the McNemars test. Moreover, the proposed method shows improvement compared to several state-of-the-art PolSAR classification methods.