BSDSNet: Dual-Stream Feature Extraction Network Based on Segment Anything Model for Synthetic Aperture Radar Land Cover Classification

• DOI: 10.3390/rs16071150
• Publication Date: 2024-03-26T15:05:33Z
• Authors: Yangyang Wang; Wengang Zhang; Weidong Chen; Chang Chen
• Keywords: Land Cover; Feature (linguistics)

Land cover classification using high-resolution Polarimetric Synthetic Aperture Radar (PolSAR) images obtained from satellites is a challenging task. While deep learning algorithms have been extensively studied for PolSAR image land classification, the performance severely constrained due to scarcity of labeled samples and limited domain acceptance models. Recently, emergence Segment Anything Model (SAM) based on vision transformer (VIT) model has brought about revolution in study specific downstream tasks computer vision. Benefiting its millions parameters extensive training datasets, SAM demonstrates powerful capabilities extracting semantic information generalization. To this end, we propose dual-stream feature extraction network SAM, i.e., BSDSNet. We change encoder part dual stream, where ConvNext utilized extract local VIT used global information. BSDSNet achieves an in-depth exploration spatial images. Additionally, facilitate fine-grained amalgamation information, SA-Gate module employed integrate local–global Compared previous models, BSDSNet’s impressive ability represent features akin versatile receptive field, making it well suited classifying across various resolutions. Comprehensive evaluations indicate that excellent results qualitative quantitative evaluation when performing AIR-PolSAR-Seg dataset WHU-OPT-SAR dataset. suboptimal results, our method improves Kappa metric by 3.68% 0.44% dataset, respectively.