SAM2-ARAFNet: adapting SAM2 with an attention-enhanced residual ASPP fusion network for high-resolution remote...

Area of Science:

• Computer Vision
• Remote Sensing
• Geospatial Analysis

Background:

• High-resolution remote sensing image segmentation is vital for environmental monitoring and resource management.
• Challenges include intra-class variability, complex scenes, and high computational costs of deep learning models.
• Existing methods often struggle with practical deployment on resource-constrained systems.

Purpose of the Study:

• To develop an efficient and accurate segmentation framework for high-resolution remote sensing images.
• To address the computational burden and parameter inefficiency of current deep learning models.
• To enable practical deployment of advanced segmentation models in edge-focused remote sensing scenarios.

Main Methods:

• Introduced SAM2-ARAFNet, integrating Segment Anything Model 2 (SAM2) with lightweight adapter modules.
• Incorporated an Attention-Enhanced Residual Atrous Spatial Pyramid Pooling (ResASPP) for enhanced multi-scale feature representation.
• Employed a distillation strategy to compress the SAM2 model into a compact EfficientNet_b0-based student network.

Main Results:

• SAM2-ARAFNet achieved mIoU scores of 85.43% (Vaihingen) and 87.44% (Potsdam), outperforming PSPNet.
• The distilled student model reduced parameters by 97% (222.98M to 6.68M).
• The student model retained over 99% of the teacher network's accuracy, demonstrating high efficiency.

Conclusions:

• SAM2-ARAFNet offers significant performance gains in remote sensing image segmentation.
• The distilled model provides a highly efficient solution suitable for edge devices.
• The framework effectively balances accuracy and computational efficiency for practical applications.