首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 273 全文下载次数: 369
引用本文:

DOI:

10.11834/jrs.20210313

收稿日期:

2020-07-28

修改日期:

2021-01-11

PDF Free   EndNote   BibTeX
基于分离卷积与密集连接轻量级神经网络的高光谱图像分类
宋廷强, 宗达, 刘童心
青岛科技大学
摘要:

针对高光谱遥感图像空间分辨率低,标注训练样本困难的问题,本文提出一种基于分离卷积(Separable convolution)与密集连接(Dense connection)的轻量级神经网络SDLN模型。该模型基于DenseNet的思想,同时采用计算量更少的分离卷积代替3D卷积,根据算法提取的光谱信息和空间信息,结合目标及周围像素信息推断其中心像素内容,实现对单像素的分类。基于IP、PU和KSC 三个广泛使用的高光谱数据集进行实验,按照分层抽样的方法,每个类别选取少量样本作为训练集,分类精度分别达到了97.4%、97.6%、99.2%,与SSRN、SVM-RBF、MDGCN、DBDA及pResNet多种先进分类算法对比,分类精度提高且时间成本降低。

A Lightweight Neural Network for Hyperspectral Image Classification Based on Separable Convolution and Dense Connection
Abstract:

To address the problems of low spatial resolution of hyperspectral remote sensing images and difficulty of labeling training samples, this paper proposes a lightweight neural network SDLN model based on Separable convolution and Dense connection. The model is based on the idea of DenseNet, and instead of 3D convolutions, it adopts separable convolutions with less computational effort. According to the spectrum and spatial information extracted by the algorithm, and in combination with the target and its surrounding pixel information to infer the content of the center pixel, the model achieves single-pixel classification. Experiments are conducted based on three widely-used hyperspectral datasets of IP, PU, and KSC. According to the stratified sampling method, a small number of samples are selected from each category as the training set, and the classification accuracy reaches 97.4%, 97.6%, and 99.2% respectively. Compared with multiple advanced classification algorithms of SSRN, SVM-RBF, MDGCN, DBDA and pResNet, the classification accuracy is improved and the time cost is reduced.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群 分享按钮