滨海湿地稀疏采样重构高光谱图像分类精度评价

Evaluation of the Classification Accuracy of Coastal Wetland Hyperspectral Image Reconstructed from Sparse Sampling

高光谱遥感影像维数高、数据量大、波段之间的相关性强,分类时易出现"Hughes"现象,因此在分类过程中如何有效减小数据处理过程中的计算量,又保证原始数据重要的地物信息不丢失具有重要的意义。压缩感知理论可通过远低于耐奎斯特的采样率和少量观测数据实现信号的精确重构,具有对硬件读写要求低、图像恢复效果好等优势。通过利用基于小波变换的压缩感知算法对黄河口地区的高光谱影像进行图像重构,然后分别采用SVM算法、最大似然法以及神经网络分类法对重构后的影像进行分类,并对分类结果的精度分别从空域和小波域、不同的测量值等维度进行了分析和比较。结果表明:(1)压缩感知理论重构后的影像保留了原始影像的基本信息,保证了分类精度;(2)SVM算法的分类精度最好,空域和小波域的分类精度基本一致;(3)分类精度随测量值的增加先逐渐提高,然后趋于稳定。

As is known to all that the hyperspectral remote sensing images have the characteristics of high dimensionality,data volume,as well as strong correlation between bands,so it tends to appear the ＂Hughes＂ phenomenon in the process of classification. Therefore,it is of great significance to effectively reduce the data processing computation in the process of classification while keeping the important feature information of the original data. The compressed sensing theory,well known as its low sampling rate,presents sampling at a rate much lower than the Nyquist sampling rate,and the signals can be accurately reconstructed with a small amount of observation data. It is a kind of image reconstruction algorithm with low requirements for storage and transmission and better effect on image restoration. In this paper,the hyperspectral image of the Yellow River Estuary Area is reconstructed based on the Wavelet transform compression algorithm,and then the reconstructed image is classified by the methods of support vector machine（SVM）,maximum likelihood and neural network.Finally,the classification results are compared and analyzed by different dimensions,including the airspace and wavelet domain,as well as different measured values. Results show that the reconstructed image has kept the basic information of the original image,ensuring the classification accuracy; the SVM algorithm has the best classification accuracy which is essentially the same in the airspace and wavelet domain; the accuracy firstly increases and then remains stable with the increase of measured value.