通过空间结构编码和亚像素聚合技术提升高光谱图像分类效果
《Optics & Laser Technology》:Enhanced hyperspectral image classification via Spatial-structure encoding and subpixel aggregation
【字体:
大
中
小
】
时间:2025年12月11日
来源:Optics & Laser Technology 4.6
编辑推荐:
本文提出一种基于双分支网络(SSE和SPA)的高光谱图像分类方法,通过空间结构编码模块捕捉邻近像素的自相似性特征,以及亚像素聚合模块增强边缘纹理细节,实现高精度分类。实验表明,该方法在四个公开数据集上准确率分别达97.21%、99.16%、97.77%和98.45%,优于现有Transformer和CNN模型,同时保持低计算成本。
hyperspectral image classification has become a critical research direction in remote sensing and intelligent recognition. The core challenge lies in effectively integrating spectral characteristics with spatial structural information while preserving fine-grained details. This paper proposes a novel dual-branch framework called SESA (Spatial-structure Encoding and Subpixel Aggregation) that systematically addresses these three key aspects.
First, the spatial-structure encoding (SSE) module introduces a self-similarity computation approach to model contextual relationships. By analyzing the similarity between central pixels and their neighboring regions, this module captures the intrinsic spatial coherence present in hyperspectral data. Unlike traditional global embedding methods, SSE explicitly preserves relative positional information through a multi-scale self-similarity mapping process. This structural awareness enables the model to better distinguish between objects with similar spectral signatures but distinct spatial arrangements.
Second, the subpixel aggregation (SPA) module innovatively adapts super-resolution techniques to hyperspectral classification. Through carefully designed subpixel convolution operations, SPA enhances the resolution of extracted features while maintaining spectral integrity. This allows the model to resolve subtle edge patterns and texture variations that are crucial for differentiating between closely related land cover types.
The integration strategy involves weighted feature fusion before classification. This dual-branch approach not only compensates for the limitations of individual modules but also creates synergistic effects through complementary information. The spatial context from SSE helps establish reliable feature associations, while the high-resolution details from SPA provide precise discriminative signals.
Experimental validation on four benchmark datasets demonstrates SESA's comprehensive advantages. On Indian Pines (IP) dataset with 220 bands, SESA achieves 97.21% accuracy while maintaining computational efficiency. The performance gap widens in complex scenarios like Pavia University (PU) where SESA outperforms HybridSN and GLMGT by 1.4-2.3 percentage points.特别值得注意的是在WHU-Hi-HanChuan (HC)和WHU-Hi-HongHu (HH)等真实复杂地物分布场景中,SESA展现出更强的细节捕捉能力和噪声抑制效果。
Comparative analysis reveals three fundamental improvements:
1. Structural-Spectral Fusion: Unlike conventional methods that process spectral and spatial features sequentially, SESA's parallel architecture enables simultaneous refinement of structural context and spectral discriminants
2. Detail-Preserving Aggregation: The SPA module's subpixel operations effectively maintain edge information loss, critical for distinguishing urban structures or agricultural patterns
3. Adaptive Feature Integration: The weighted fusion mechanism dynamically adjusts the contribution of structural and detail features based on data characteristics, achieving optimal performance across different scenarios
Practical applications show significant advantages in real-world scenarios. For precision agriculture, SESA demonstrates superior capability in detecting crop types with similar spectral profiles but distinct growth patterns. In military surveillance, the model maintains high accuracy for distinguishing伪装植被 (disguised vegetation) and natural cover using fine spatial details. When compared to state-of-the-art methods like HybridSN and GLMGT, SESA consistently achieves higher F1-scores (from 0.96 to 0.98) while requiring 30-50% less computational resources.
The proposed architecture addresses two fundamental limitations of existing methods:
1. Structural Ambiguity Problem: Traditional patch-based approaches often misinterpret structural relationships due to fixed-size windowing. SESA's self-similarity computation inherently captures multi-scale structural dependencies
2. Detail Information Loss: Common spectral-spatial fusion methods discard important subpixel-level features during dimensionality reduction. SPA's subpixel aggregation preserves these details through iterative upsampling
Methodological innovations include:
- A hierarchical self-similarity calculation framework that simultaneously considers local (3×3 neighborhood) and global (256×256 region) structural relationships
- Subpixel convolution with adaptive kernel size selection mechanism, which dynamically adjusts convolution parameters based on spectral band characteristics
- Dual-path feature fusion with context-aware weight adaptation, where the fusion ratio is automatically calibrated during inference
The computational efficiency is achieved through three optimization strategies:
1. Single-pass self-similarity computation with band compression techniques
2. Lightweight subpixel convolution operations using 1×1×1 kernel
3. Parameter-efficient fusion module with fixed-size feature map processing
Real-world validation on WHU-Hi-HongHu dataset shows SESA maintains >95% accuracy for urban land cover classification even in low-light conditions. The model's robustness is further demonstrated by its ability to handle mixed pixels with spectral similarity up to 98.5% while maintaining subpixel accuracy.
Ethical considerations include rigorous data augmentation protocols to prevent overfitting, especially in cases with limited training samples. The open-source codebase provides extensive documentation and pre-trained models for different spectral resolutions, promoting reproducibility in scientific research.
In conclusion, SESA represents a significant advancement in hyperspectral image analysis by systematically addressing structural modeling and detail preservation challenges. The proposed dual-branch architecture not only achieves state-of-the-art classification accuracy but also provides a scalable framework that can be adapted to various spectral configurations and application scenarios. The computational efficiency demonstrated in this work makes it particularly suitable for real-time hyperspectral processing in precision agriculture and environmental monitoring applications.
生物通微信公众号
生物通新浪微博
今日动态 |
人才市场 |
新技术专栏 |
中国科学人 |
云展台 |
BioHot |
云讲堂直播 |
会展中心 |
特价专栏 |
技术快讯 |
免费试用
版权所有 生物通
Copyright© eBiotrade.com, All Rights Reserved
联系信箱:
粤ICP备09063491号