A Target-oriented Feature Selection Framework for Polarimetric SAR Image Classification Based on Multi-scale Two-dimensional Structural Similarity

Concise and informative target feature descriptions are curial for accurate land cover classification of polarimetric synthetic aperture radar (PolSAR) images. Effective feature selection strategy significantly impacts both classification model design and final accuracy.

Ideally, target features should capture diverse polarization scattering characteristics and physical properties, the foundation for PolSAR image interpretation, and also require all these features satisfy the independent and identically distributed hypothesis, as directly using all these features can lead to sparse sample data in the multi-dimensional space, especially with limited samples, hindering model training.

To this end, existing research attempt to utilize multiple features manually or analyze specific scattering characteristics for classification scenarios. However, these studies mainly focus on manual feature selection or using traditional random forest importance-based feature selection strategy, adaptive feature selection tailored to individual situations remains less explored.

In order to address this gap, we propose a novel target-oriented feature selection framework leveraging multi-scale two-dimensional structural similarity measure (MTSSIM). This framework adaptively selects informative features from an initial PolSAR image feature set, encompassing commonly used polarization scattering features, spatial neighbor context features, and morphological features. The core principle lies in designing an efficient algorithm that selects features maximizing intra-class and minimizing inter-class structural similarity.

For enhanced robustness and practicality, the proposed framework incorporates two key modules: 1) Two-dimensional structural similarity representation: This module quantifies the structural similarity between two samples, and 2) Multi-scale feature structural similarity measurement: This module utilizes local feature images at multiple spatial neighborhood scales to assess the intra-class and inter-class structural similarity of each feature relative to the target category.

To validate the effectiveness of the proposed framework, we conducted classification experiments on two real PolSAR image datasets using identical classification methods and parameters for three feature sets: the manually chosen features that commonly used in PolSAR image classification task (Manual feature set), the random forest importance-based features (RF feature set), and MTSSIM-recommended features (MTSSIM feature set).

Experimental results demonstrate that the proposed MTSSIM feature set consistently outperforms traditional approaches, demonstrating significant improvements in classification accuracy. These benefits include: 1) Reduced misclassification rates: MTSSIM significantly decreases misclassified pixels, leading to more accurate and reliable land cover maps; 2) Enhanced homogeneity: MTSSIM-derived feature sets yield spatially consistent and less noisy classification results, facilitating easier interpretation and analysis. 3) Improved performance in small-sample scenarios: MTSSIM effectively utilizes limited data, enabling accurate classification even with limited training samples.

In conclusion, the MTSSIM framework offers a powerful and practical solution for optimizing feature selection in PolSAR image classification. By addressing feature redundancy and leveraging structural information, MTSSIM improves classification accuracy, making it a valuable tool for enhancing remote sensing applications in land cover mapping, environmental monitoring, and various other domains.