[Paper Review] Research on Improved U-net Based Remote Sensing Image Segmentation Algorithm

Qiming Yang, Zixin Wang+2 more|arXiv (Cornell University)|Aug 22, 2024

Remote Sensing and Land UseEarth and Planetary Sciences3 citations

TL;DR

This paper proposes an improved U-Net architecture for remote sensing image segmentation by integrating SimAM and CBAM attention mechanisms, significantly boosting performance: a 19.11% improvement in mean Intersection over Union (mIoU), 16.38% in Mean Pixel Accuracy (mPA), and 14.8% in Accuracy, demonstrating strong generalization and robustness on remote sensing data.

ABSTRACT

In recent years, although U-Net network has made significant progress in the field of image segmentation, it still faces performance bottlenecks in remote sensing image segmentation. In this paper, we innovatively propose to introduce SimAM and CBAM attention mechanism in U-Net, and the experimental results show that after adding SimAM and CBAM modules alone, the model improves 17.41% and 12.23% in MIoU, and the Mpa and Accuracy are also significantly improved. And after fusing the two,the model performance jumps up to 19.11% in MIoU, and the Mpa and Accuracy are also improved by 16.38% and 14.8% respectively, showing excellent segmentation accuracy and visual effect with strong generalization ability and robustness. This study opens up a new path for remote sensing image segmentation technology and has important reference value for algorithm selection and improvement.

Motivation & Objective

• Address performance bottlenecks in standard U-Net for remote sensing image segmentation.
• Improve segmentation accuracy and robustness in complex remote sensing scenes.
• Enhance feature representation by integrating attention mechanisms into the U-Net architecture.
• Evaluate the effectiveness of individual and fused attention modules on segmentation metrics.
• Establish a new benchmark for attention-augmented U-Net in remote sensing applications.

Proposed method

• Integrate the learnable, gradient-free SimAM attention module into the U-Net encoder and decoder pathways.
• Apply the CBAM (Convolutional Block Attention Module) to focus on channel and spatial attention in feature maps.
• Fuse SimAM and CBAM modules in parallel branches to enhance multi-level feature representation.
• Train the hybrid model end-to-end on remote sensing image datasets using standard segmentation loss functions.
• Utilize skip connections from U-Net to preserve spatial details while attending to critical features.
• Optimize the model with standard deep learning training protocols, including data augmentation and learning rate scheduling.

Experimental results

Research questions

• RQ1 How does the integration of SimAM improve U-Net performance in remote sensing image segmentation?
• RQ2 To what extent does CBAM attention enhance feature learning in remote sensing segmentation tasks?
• RQ3 What is the combined effect of SimAM and CBAM on mIoU, mPA, and Accuracy in segmentation performance?
• RQ4 Does the fused attention mechanism improve model generalization and robustness on diverse remote sensing scenes?
• RQ5 Can the proposed architecture outperform standard U-Net and existing attention-augmented variants on benchmark datasets?

Key findings

• The model with only SimAM attention achieved a 17.41% improvement in mIoU over the baseline U-Net.
• The model with only CBAM attention improved mIoU by 12.23% compared to the baseline.
• The fused SimAM and CBAM model achieved a 19.11% increase in mIoU, demonstrating superior performance.
• The fused model improved Mean Pixel Accuracy (mPA) by 16.38% and Accuracy by 14.8% over the original U-Net.
• The proposed method showed strong visual segmentation quality and robustness across diverse remote sensing image types.
• The integration of both attention mechanisms significantly enhanced feature representation and model generalization.