融合植被指数的3D-2D-CNN高光谱图像植被分类方法

对于基于高光谱图像的植被分类,利用三维卷积神经网络和空谱结合可以取得良好的效果。但存在计算代价大、参数过多容易过拟合等问题。基于此,设计了一种三维卷积与二维卷积相结合的深度网络,通过数据分块的思想减小了计算量;并提出了一种融合植被指数的特征提取方法,改善了现阶段因高光谱图像样本数量少、光谱层间信息相关度高,造成的容易过拟合的问题。在植物园数据集、IP数据集和PU数据集上的实验结果表明该算法以较低的计算复杂度取得了出色的分类效果,具有较好的应用价值。

A Hyperspectral Image Vegetation Classification Method Using 2D-3D CNNs and Vegetation Index

For hyperspectral images-based vegetation classification, the use of spatial-spectrum information and three-dimensional convolutional neural networks usually can produce competitive results. However, there are several limitations for this kind of approach, including high computational/parameter complexity and model overfitting. To address these problems, a deep vegetation classification network that combines the three-dimensional and two-dimensional convolution is proposed in this paper. More specifically, in the proposed vegetation classification method, the computational cost is reduced via data blocking. Moreover, to ease the model overfitting problem owing to very limited hyperspectral images and the high correlation among spectral layers, the vegetation index is incorporated to aid feature extraction. Experimental results on botanical garden data set, IP data set and PU data set show that the algorithm proposed in this paper achieves excellent classification performance with low computational complexity and shows a promising application prospect.