基于改进的三维卷积神经网络的高光谱遥感影像分类技术研究

Research on hyperspectral remote sensing image classification based on 3D convolutional neural network

高光谱遥感影像数据具有多样化的光谱信息和空间信息,然而传统的高光谱影像分类只是针对目标的光谱特征进行处理。基于三维空间滤波操作可以作为一种简单高效的提取高光谱影像光谱和空间特征的方式,基于此提出一种改进的三维卷积神经网络框架以实现更加准确的高光谱遥感影像分类。利用高光谱遥感影像数据立方体有效地提取光谱-空间组合特征,而不依赖于任何预处理或后期处理。另外,与其他传统的基于深度学习的方法相比,该方法去除了池化层,从而达到所需参数更少,模型规模更小,更容易训练的效果。将该方法与其他基于深度学习的高光谱遥感影像分类方法进行了比较,并使用两个真实场景的高光谱遥感影像数据集作为测试。实验结果表明,该方法在地物分类准确度方面较传统的基于深度学习的高光谱遥感影像分类方法取得了更好的分类效果。

Hyperspectral remote sensing image contains both rich spectral and spatial information.However,traditional hyperspectral image classification is usually based on spectral features.Based on three-dimensional spatial filtering,it can be used as a simple and effective method to extract spectral spatial features of hyperspectral images.An improved three- dimensional convolutional neural network framework is proposed for accurate hyperspectral remote sensing image classification.The data cube effectively extracts spectral-spatial combination features without relying on any pre-processing or post-processing.In addition,compared to other traditional deep learning-based methods,a smaller model size which requires fewer parameters,and less likelihood of overfitting can be much easier to train.This method is compared with other hyperspectral remote sensing image classification methods based on deep learning,and is tested on two hyperspectral remote sensing image datasets.Compared to traditional deep learning methods,3D-CNN gets better accuracy among the results of hyperspectral images.