Research on Complex Spectral Recognition Methods and Performance Evaluation Based on Deep Learning
基于深度学习的复杂光谱识别方法及性能评价研究

摘要

复杂光谱数据具有波段数量多、相邻波段相关性强、类别边界模糊和空间结构差异明显等特点，传统 分类方法难以充分提取深层判别信息。深度学习能够通过分层网络自动学习光谱、空间及空谱联合特征，已 成为高光谱识别的重要方法。文章分析一维卷积、二维卷积、三维卷积、混合卷积和注意力机制的基本特点， 并结合Indian Pines、Pavia University 等公开数据集的实验结果评价不同方法的分类性能。结果表明，空 谱联合网络整体精度较高，但小样本、类别不平衡和跨场景泛化仍是制约实际应用的重要因素。

Abstract

Complex spectral data are characterized by a large number of bands, strong correlations between adjacent bands, ambiguous class boundaries, and significant differences in spatial structures, making it difficult for traditional classification methods to fully extract deep discriminative information. Deep learning can automatically learn spectral, spatial, and joint spectral-spatial features through hierarchical networks, and has become an important method for hyperspectral recognition. This paper analyzes the basic characteristics of one-dimensional convolution, two dimensional convolution, three-dimensional convolution, hybrid convolution, and attention mechanisms, and evaluates the classification performance of different methods in combination with experimental results on public datasets such as Indian Pines and Pavia University. The results show that spectral-spatial joint networks achieve relatively high overall accuracy, but small samples, class imbalance, and cross-scene generalization remain important factors restricting practical applications.