样本尺寸对遥感影像FCN训练模型的影响分析

Analysis of sample size influence on FCN training model in remote sensing image

针对如何选择合适尺寸的影像样本来得到较好的网络模型这一问题,该文基于全卷积神经网络(FCN)的遥感影像分类方法,开展了不同样本尺寸下的网络模型训练实验,分析了样本尺寸分别为128、256、512像素大小时对FCN网络模型的影响。结果表明:512像素×512像素大小样本尺寸下像素准确率、平均准确率、平均交叉联合度量和带权交叉联合度量4个评价指标的精度值均高于128像素×128像素和256像素×256像素大小的值,比128像素×128像素样本尺寸平均高出20%以上,比256像素×256像素样本尺寸高出10%以上,因此,在计算机内存允许范围内采用大尺寸样本进行网络模型的训练,有利于提高模型训练精度,可得到更好的分类结果。

Aiming at the problem of how to select the appropriate size image samples to get a better network model,this paper developed a full convolutional network(FCN)to deal with the problem of what sample size could get a better network model.The remote sensing classification experiments using FCN were carried out with the different sample sizes.The influences of 128,256,and 512 sample sizes on the FCN network model were analyzed.The training results showed that four accuracy values(Pixel Accuracy,Mean Accuracy,Mean IOU and FW IOU)of the 512 pixel×512 pixel sample size were higher than the values of 128 pixel×128 pixel and 256 pixel×256 pixel,which were 20% higher than the 128 pixel×128 pixel sample size,and 10% higher than the 256 pixel×256 pixel sample size.Therefore,the use of the large-scale samples for a network training model within the allowable range of computer memory is beneficial to improve the training accuracy of the model and obtain better classification results.