ConvAttentionNet: a high-performance model for efficient and accurate PolSAR data classification

This paper presents ConvAttentionNet, a lightweight and high performing deep learning model developed for accurate and efficient classification of Polarimetric Synthetic Aperture Radar (PolSAR) imagery. The proposed architecture combines multiscale convolutional mixer blocks with a directional convolution based attention mechanism to effectively capture spatial features and suppress background noise. Designed to address the challenges of limited labeled data and computational constraints, ConvAttentionNet achieves superior performance while maintaining a compact model size. Experimental results on three benchmark datasets (Flevoland, San Francisco, and Oberpfaffenhofen) demonstrate that ConvAttentionNet consistently outperforms state of the art CNN based, transformer based, and wavelet based models. It achieves an overall accuracy (OA) of 97.24% and a Kappa coefficient of 96.98 on the Flevoland dataset using only 1% of the training data. These results confirm the model’s robustness, label efficiency, and generalization capabilities, making it a practical solution for operational remote sensing scenarios with limited computational resources. The source code for this work will be publicly available at: https://github.com/aj1365/ConvAttentionNet .