基于三维卷积神经网络模型的吉林一号光谱星影像森林类型分类

Forest Type Classification with Jilin-1GP Spectral Satellite Image Based on Three-dimension Convolution Neural Network

为探究基于三维卷积神经网络模型应用吉林一号光谱卫星数据的森林类型分类效果,以安图县与敦化市交界地带为研究区,采用吉林一号光谱卫星影像为主要数据源,基于三维卷积神经网络深度学习模型对研究区森林类型进行分类,并与传统的随机森林分类方法进行对比分析。结果表明:当三维卷积神经网络的训练样本数量为2400,训练次数为70000时,分类效果最佳。基于三维卷积神经网络方法的总体分类精度为92.9%,Kappa系数为0.92;与随机森林方法分类结果对比,总体分类精度和Kappa系数分别提高了2.8百分点和0.03;三维卷积神经网络能够更加充分地利用遥感影像丰富的光谱信息和空间信息,提高森林类型的分类精度,在斑块构成和景观破碎度方面均得到较大提升,内部完整度较高,破碎化程度较轻微,更贴合实际工作需要。说明国产吉林一号光谱卫星遥感数据可以有效地对森林类型进行识别,在林业的生产经营上具有极大的应用潜力。

In order to explore the effect of forest type classification with Jilin-1 spectral satellite(Jilin-1GP)data based on the three-dimensional convolution neural network model,the junction of Antu County and Dunhua City was taken as the study area,the image of Jilin-1GP was used as the main datasource to classify forest types in the study area based on the deep learning model of three-dimensional convolution neural network.The method was compared with the traditional random forest classification method.The results showed that when the number of training samples of three-dimensional convolution neural network was 2400 and the number of training times was 70000,the classification effect was the best.The overall classification accuracy based on the three-dimensional convolution neural network method was 92.9%,and the Kappa coefficient was 0.92.Compared with the classification results of the random forest method,the overall classification accuracy and Kappa coefficient increased by 2.8 percentage points and 0.03,respectively.Three-dimensional convolution neural network could make full use of the rich spectral and spatial information of remote sensing images to improve the classification accuracy of forest types.The result had been greatly improved in patch composition and landscape fragmentation.The internal integrity was higher,and the degree of fragmentation was slight,which was more in line with the actual work needs.At the same time,it also showed that the domestic Jilin-1GP remote sensing data could be used to effectively identify forest types.That had great application potential in forestry production and management.