Classification evaluation and improvement of airborne PolSAR images for land use mapping using deep learning

Abstract

Polarimetric synthetic aperture radar (PolSAR) images have been widely used in many fields due to its advantage in obtaining full polarization information, especially in land use classification. To evaluate the performance of airborne PolSAR images in land use classification, this paper systematically evaluated the potential value of PolSAR images in land use classification by using machine learning and deep learning algorithms, and improved the classification performance of airborne PolSAR images by constructing a multi-structural feature aware attention network (MSFA-Net). The classification performance of different algorithms was evaluated by the overall accuracy. The results showed that the machine learning algorithm Random Forest (RF) produced the lowest overall accuracy of 71.93%. Using RF as the baseline accuracy, the deep learning-based classification algorithm significantly improved the classification accuracy. Specifically, Wide Contextual Residual Network achieved an overall accuracy improvement of 8.87% to 80.80%, and Two-branch CNN generated an overall accuracy of 84.49%. Multi-source CNN produced the second highest overall accuracy of 87.25%, an improvement of 15.32%. The proposed MSFA-Net algorithm achieved the highest classification accuracy of 89.29% with an improvement of 17.36%, which further improves the land use mapping capability of airborne PolSAR images. A moderate reduction in resolution helps to improve classification accuracy, with 84.36% accuracy after 4-fold downsampling compared with the original resolution accuracy of 69.30%. On the other hand, noise, polarization error and motion error had a negative impact on the classification performance. This paper highlights the application of airborne SAR images in land use classification, and strengthens the understanding of deep learning and the influence of different factors on PolSAR images in land use classification.