基于混合卷积网络的高分辨率高光谱图像分类研究  

High-Resolution Hyperspectral Image Classification Based on Hybrid Convolutional Network

作　　者：沈柄志 聂若梅 蒋海鹏 杨智帅 宋洺睿 陈思琪 李鑫伟 Shen Bingzhi;Nie Ruomei;Jiang Haipeng;Yang Zhishuai;Song Mingrui;Chen Siqi;Li Xinwei(College of Science,Beijing Forestry University,Beijing 100083,China)

机构地区：[1]北京林业大学理学院,北京100083

出　　处：《激光与光电子学进展》2022年第24期278-286,共9页Laser & Optoelectronics Progress

基　　金：中央高等学校基本科研业务费项目(BLX201610);北京林业大学“北京市大学生科学研究与创业行动计划”(S202110022147)。

摘　　要：传统卷积神经网络模型未能充分利用高分辨率高光谱图像中丰富的空间-光谱信息,存在计算成本大、小样本数据分类精度低的问题。提出一种轻量化多尺度金字塔混合池化混合卷积模型。以混合卷积网络为基础,所提模型采用改进的金字塔池化模块增强对空间-光谱特征的提取能力,使用较少的卷积层和深度可分离卷积,并用全局平均池化层代替部分全连接层以实现卷积层到全连接层的过渡,显著降低了参数量。采用三个高分辨率高光谱数据集对所提方法进行测试,同时与经典高光谱图像分类方法进行对比实验,结果表明所提方法在分辨率高、地物种类多、边界复杂的情况下仍然能取得最佳的分类结果。在WHU-Hi-LongKou、WHU-Hi-HanChuan、WHU-Hi-HongHu数据集上仅使用1%、2%、2%训练样本的情况下,所提方法的总体精度分别达99.12%、98.43%、98.84%,优于传统卷积网络,证明了所提模型计算成本小,在小样本问题上准确率高,能很好地适用于高分辨率高光谱数据集。Traditional convolutional neural network models fail to fully utilize the rich spatial-spectral information in highresolution hyperspectral images,and have the problems of high computational cost and low classification accuracy for small sample data.This study proposes a lightweight multiscale pyramid hybrid pooling hybrid convolution model.Based on the hybrid convolution network,the proposed model uses an improved pyramid pooling module to enhance the ability to extract spatial-spectral features,uses fewer convolution layers and depth separable convolution,and uses the global average pooling layer to replace a part of the full connection layer to achieve the transition from the convolution layer to the full connection layer,significantly reducing number of parameters.In this study,the proposed method is tested on three high-resolution hyperspectral datasets and compared with classical hyperspectral image classification methods.The results show that the proposed method can achieve the best classification results under high-resolution conditions,multiple ground object types,and complex boundaries.The overall accuracy of the proposed method on WHU-Hi-LongKou,WHU-HiHanChuan,and WHU-Hi-HongHu datasets are 99.12%,98.43%,and 98.84%,respectively,when only 1%,2%,and 2%training samples are used,which is superior to that of the traditional convolutional networks.It is proved that the model proposed in this study has a low computational cost,high accuracy for small sample problems,and can be well applied to high-resolution hyperspectral datasets.

关 键 词：遥感 高光谱图像分类 混合卷积网络 混合池化 特征融合 高分辨率 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]