高光谱成像
卷积(计算机科学)
计算机科学
人工智能
对偶(语法数字)
图像(数学)
计算机视觉
模式识别(心理学)
遥感
人工神经网络
地理
文学类
艺术
作者
Xiang Fan,Zhaohui Wang
出处
期刊:International Conference on Digital Image Processing
日期:2025-07-22
卷期号:: 116-116
摘要
In the field of hyperspectral image (HSI) classification, the accurate categorization of each pixel into specific land cover types remains a critical challenge. The development of efficient spectral-spatial joint feature extraction models is essential for improving classification accuracy. In recent years, Convolutional Neural Network (CNN)-based methods have made significant strides in HSI classification due to their powerful feature extraction capabilities. However, these methods face two major limitations: the computational complexity escalates as network depth increases, and they struggle to effectively integrate global and local spectral-spatial features. On the other hand, Transformer-based methods have shown unique advantages in global spectral modeling, as their self-attention mechanisms can effectively capture long-range dependencies. Nevertheless, these methods are often inefficient in extracting local spatial features and suffer from high computational and memory costs. To address these issues, we propose a novel HSI classification model based on the Mamba model, designed for HSI classification. It is capable of modeling long-range interactions across the entire image and capturing spatial and spectral information in an adaptive manner, while maintaining high computational efficiency. Specifically, efficient Mamba and CNN modules are employed to extract global and local spectral-spatial features. Initially, we utilize adaptive Mamba modules and spatial-spectral convolutions with residual connections to extract fused spatial-spectral features. Then a spectral-spatial information fusion module is used to combine global Mamba features with local convolutional features, achieving deep integration of spatial-spectral features. Experimental analysis on three standard datasets indicates that the proposed method Dual-Branch Spectral – Spatial Mamba and Convolution Network (DBMC) surpasses existing classification techniques in classification performance, achieving leading classification results.
科研通智能强力驱动
Strongly Powered by AbleSci AI