PolicyMamba: Localized Policy Attention With State Space Model for Land Cover Classification.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Multihead attention mechanisms in hyperspectral image (HSI) classification face computational inefficiencies and scalability issues.
• Capturing long-range dependencies in HSI data is challenging due to the quadratic complexity of self-attention.

Purpose of the Study:

• To introduce PolicyMamba, an efficient spectral-spatial mamba model for improved HSI classification.
• To address the limitations of traditional attention mechanisms in HSI analysis.

Main Methods:

• PolicyMamba utilizes a localized policy attention mechanism to reduce computational overhead by focusing on non-overlapping regions and enforcing sparsity.
• A hierarchical aggregation strategy integrates patch-wise attention outputs to preserve spectral-spatial correlations across scales.
• A sliding window patch process is employed to enhance local feature continuity and minimize information loss.

Main Results:

• PolicyMamba demonstrates superior classification accuracy in land cover classification (LCC) compared to conventional and state-of-the-art methods.
• The model efficiently models intricate spectral-spatial dependencies within HSI data.
• Experimental results validate the effectiveness of the proposed localized policy attention and hierarchical aggregation.

Conclusions:

• PolicyMamba offers a computationally efficient and scalable solution for HSI classification.
• The proposed model significantly enhances feature representation and classification performance.
• This work provides a new direction for developing advanced deep learning models for HSI analysis.