一种遥感影像自动识别耕地类型的机器学习算法

A machine learning algorithm for automatic identification of cultivated land in remote sensing images

耕地作为重要的土地资源,关系着国家的粮食安全问题,因此迫切需求快速准确获取耕地信息的方法。传统的遥感影像监督分类方法以训练样本和待分类像元/图斑的光谱特征或纹理特征的一致性作为分类依据,这对训练样本的依赖性较强。对此提出了一种基于影像窗口子区的耕地类型自动识别算法,通过提取一定大小影像窗口子区的多光谱和多层次特征,利用机器学习算法,实现影像窗口子区耕地和非耕地类型的自动判别。依据该算法,可以通过建立某个区域内遥感影像耕地类型的特征库,实现对影像窗口子区类别的非监督自动判别,提高目前分类算法的自动化程度。以东北地区高空间分辨率遥感影像为例进行实验,精度达到了90. 8%。该算法为耕地信息自动化快速获取提供了技术支持,也可用于遥感影像中某一种纯净地物类型的快速提取。

As an important kind of land resources,cultivated land is related to the country9s food security.So it is very significant to have a fast and accurate method for obtaining information of cultivated land.The traditional supervised classification methods of remote sensing image are based on the consistency of the spectral features or texture features between the training samples and the pixels/patches to be classified.These methods have strong dependence on training samples.This paper proposes an automatic classification algorithm of cultivated land based on the image window subarea.By using the machine learning algorithm,the automatic classification of cultivated land or non-cultivated land in the sub region of the image window can be realized by extracting the multi-spectral and multi-level features.Using this method,the unsupervised automatic classification of the type of the image window subarea is realized by establishing the feature database of the remote sensing image of the cultivated land in a certain area.With the high spatial resolution remote sensing image of Northeast China as an example,the experimental results show that the accuracy of the automatic classification algorithm is90.8%.Being able to automatically acquire the cultivated land information,this method can also be used to extract any pure ground object from remote sensing images.