基于多维卷积神经网络的多源高分辨率卫星影像茶园分类

Classification of tea garden based on multi-source high-resolution satellite images using multi-dimensional convolutional neural network

武夷山市地形条件、茶园种植结构复杂,云雨天气多、卫星影像难获取。针对单一影像源茶园难提取的问题,以武夷山市新田镇为研究区,综合Sentinel-2影像的光谱信息和Google影像的纹理特征,提出一种基于多源高分辨率卫星影像和多维卷积神经网络(multidimensional multi-source convolutional neural networks,MM-CNN)的茶园分类方法。该方法以一维和二维卷积神经网络为基础,根据不同分辨率的影像,通过建立2种模型,分别提取茶园及疑似区域,并融合2个模型结果,最终得到茶园分布,以相对经济、高效的方式完成研究区茶园分布的高精度提取。结果表明,MM-CNN融合多源高分辨率影像进行茶园提取的空间分布精度优于单一影像源方法,MM-CNN方法具有一定的普适性和鲁棒性,为南方丘陵山区大范围高效监测茶园分布情况提供了方法参考。

The terrain conditions and tea plantation structure of Wuyishan City are complex,with cloudy and rainy weather,so it is difficult to obtain satellite images here.To address the problem of difficult extraction of tea gardens from a single image source,we investigated the spectral information of Sentinel-2 images and the texture features of Google images in Xintian Town,Wuyishan City,coupled with which a tea garden classification method based on multi-source high-resolution satellite images and multidimensional convolutional neural networks(MM-CNN)was established.In this method,tea gardens and suspected tea gardens were extracted,respectively,with two models developed with images with different spatial resolutions,based on one-and two-dimensional CNN.Results obtained with the two CNN models were combined,and the high-accuracy distribution of tea gardens in the study area was generated in a relatively economical and efficient way.The results showed that the spatial distribution accuracy of the tea gardens identified by MM-CNN is better than that of the single image source method.The MM-CNN method is highly universal and robust and provides a reference method for efficiently monitoring the distribution of tea gardens in large-scale hilly areas of South China.