基于多核极限学习机的遥感影像林地信息提取

Hyperspectral remote sensing images classification based on multi-kernel extreme learning machine

高光谱遥感影像包含了大量的波段信息,能够很好地应用于地物的识别。基于单核的高光谱遥感影像极限学习机分类模型,因其实施简单、分类精度高、训练时间短,已被广泛地应用于高光谱遥感影像识别。但是核特征的选取,以及单核特征表达的单一性,限制了模型分类精度的进一步提高。为了解决此问题,受多核极限学习机(MK-ELM)思想的启发,首先使用核方法,提取了遥感影像的多核特征表达;然后利用多核极限学习机理论,同时优化极限学习机结构参数以及多核特征融合系数,获得最优的分类模型。为了说明MK-ELM的有效性,在Indian pines数据集上做了对比实验,该实验证明基于多核极限学习机遥感影像分类模型的分类精度较单核极限学习机有明显地提高,MK-ELM的分类整体精度为80.2%,Kappa系数高达78%;同时将多核极限学习机应用到芷江林场的林地信息提取,其分类精度高达89.1%,Kappa系数达86%。

The hyperspectral remote sensing image contains a lot of band information and can be applied to ground object recognition. The multi-spectral remote sensing image classification based on single-kernel extreme learning machine model has been widely used in multi-spectral remote sensing image recognition due to its simple implementation, high classification accuracy and short training time. However, the selection of nuclear features and the uniqueness of expression of monocytes limit the improvement of model classification accuracy and further application. In order to solve this problem, inspired by the idea of multi-kernel extreme learning machine（mkelm）, we first extracted the expression of multi-kernel features of remote sensing images using multiple kernel methods. Then the theory is used to optimize the structural parameters of the extreme learning machine and the fusion coefficient of multi-kernel features and obtain optimal classification model. In order to illustrate the effectiveness of the MK-ELM, we do the contrast experiment on the Indian pines data sets, the experiment proved that the classification accuracy of remote sensing image based on multi-kernel extreme learning machine model is higher than single-kernel extreme learning machine model.And the experiment indicate that MK-ELM can collect forest information better. The overall accuracy of mk-elm classification is 80%, and the Kappa coefficient is up to 78%. At the same time, the multi-kernel extreme learning machine is applied to extract the woodland information of zhijiang forest farm. The classification accuracy is up to 89.1% and Kappa coefficient is up to 86%.