利用U-Net神经网络的多光谱图像草地特征提取

Grass Feature Extraction for Multi-spectral Images Based on U-Net Neural Network

针对多光谱图像目标物特征提取中波段信息融合、计算数据量大、上下文语义信息易混淆等问题,提出了一种融合U-Net神经网络的特征分析方法进行多光谱图像草地识别。首先对图像进行预处理,通过几何校正、辐射校正和图像配准等消除各种不良干扰;然后利用主成分分析法和相关性分析法实现特征波段选择;最后搭建U-Net神经网络进行草地特征提取,辨识出样本图像中的草地范围。仿真实验表明,该方法具有较高的特征辨识和提取精度,具有一定的可行性。

In this paper,we proposed a feature analysis method incorporating U-Net neural network for grass recognition in multi-spectral images,which addressed the problems of fusion of waveband information,large amount of computational data and easy confusion of contextual semantic information in feature extraction of UAV multi-spectral images.Firstly,we preprocessed the images,and eliminated undesirable interferences through HO correction,radiation correction and image alignment.Then,we used principal component analysis and correlation analysis to select the feature bands.Finally,we built a U-Net neural network to extract grass features and identify the extent of grass in the sample images.The simulation experiments show that the method has high feature recognition and extraction accuracy and is feasible.