基于改进训练策略的高光谱图像分类  

Hyperspectral image classification based on improved training strategy

作　　者：吴少乔 WU Shao-qiao(School of Information and Communication,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,哈尔滨150001

出　　处：《哈尔滨商业大学学报（自然科学版）》2022年第2期154-161,170,共9页Journal of Harbin University of Commerce：Natural Sciences Edition

摘　　要：将卷积神经网络应用在高光谱图像分类中,提出了一种基于训练集损失的训练策略.这种策略选取固定训练周期后半段训练集损失最小时的权重作为最终使用的权重,模型在固定周期下训练完毕,输出的模型为训练集损失最小时的模型.为了评估提出训练策略的有效性,在Indian Pines、Pavia University、Salina Valley数据集上使用了SSRN、FDSSC、DBMA、DBDA模型,对比于广泛使用的早停训练策略进行了实验,识别精度及其稳定性普遍获得了提高,使用的有标记样本仅为早停训练策略的一半.In the application of convolutional neural network in hyperspectral image classification, a training strategy based on training set loss is proposed. This strategy selected the weight with the least loss of training set in the second half of the fixed training cycle as the final weight, and the output model was the model with the least loss of training set after the training of the model in the fixed cycle. In order to evaluate the effectiveness of the proposed training strategy, SSRN, FDSSC, DBMA and DBDA models were used on Indian Pine, Pavia University and Salinas Valley datasets. Compared with the widely used early stopping strategy, the recognition accuracy and stability were generally improved. At the same time, the labeled samples were only half of the early stopping strategy.

关 键 词：高光谱图像分类 卷积神经网络 早停策略 最小损失 识别精度 图像分类 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]