计算机科学
人工智能
高光谱成像
上下文图像分类
模式识别(心理学)
图像(数学)
统计分类
计算机视觉
特征提取
网(多面体)
图像处理
图像分割
支持向量机
数据分类
人工神经网络
自适应滤波器
像素
遥感
作者
Arafat Al-Jawari,Li Fang,Radhwan A. A. Saleh,Huifang Shen
标识
DOI:10.1109/jstars.2026.3694796
摘要
Hyperspectral image classification (HSIC) remains a challenging task due to the complex relationship between space and spectrum in hyperspectral images, in order to improve the classification accuracy. This paper presents the Adaptive Multi-Scale Attention Net (AMSA-Net), a novel deep learning architecture that integrates multi-scale feature extraction with dynamic spectral-spatial attention for robust HSIC. AMSA-Net combines hierarchical encoding-decoding with a Dynamic Spectrum-Aware Attention (DSAA) block at its bottleneck, which processes features through three complementary pathways: global self-attention for long-range dependencies, local depthwise convolution for channel-specific patterns, and patch-based attention for intermediate-scale relationships. A learned gating mechanism dynamically fuses these branches based on local spectral-spatial characteristics. Comprehensive experiments on four benchmark datasets (Pavia University, WHU-Hi-HanChuan, WHU-Hi-HongHu, and QUH-Tangdaowan) demonstrate that AMSA-Net outperforms state-of-the-art methods. Ablation studies validate the contribution of each model's component, showing that AMSA-Net is particularly crucial for handling spectral variability in agricultural scenes. Visual analyses confirm AMSA-Net's superior spatial coherence and ability to preserve fine boundaries compared to CNN and transformer-based approaches. The proposed architecture provides an effective solution for hyperspectral analysis, particularly in applications that require precise discrimination among spectrally similar land-cover classes. The source code is available herehttps://github.com/Arafat-Aljawari/AMSA-Net.
科研通智能强力驱动
Strongly Powered by AbleSci AI