基于小波分析的Hyperion影像地物分类波段宽度

【目的】高光谱遥感为地物的精细识别提供优势的同时,也带来了数据量多、波段间相关性大、处理精度和效率下降等问题,而且在遥感分类中并不是使用的通道越多、波段越窄效果越好。因此笔者从光谱角度出发探讨降低高光谱数据量,以寻求适宜遥感分类波段宽度的方法。【方法】首先对Hyperion影像进行处理,主要包括去除未定标和受水汽影像波段、坏线、条纹和Smile效应,辐射定标和大气校正处理后得到161个波段,对选用的LIR级数据进行几何校正。根据样地调查情况确定试验区待分类别,对提取的14类地物样本平均光谱进行7次Sym3小波分解,由得到的小波细节系数方差和小波细节系数熵分析适宜各类型识别的光谱区间,然后将不同光谱区间内窄波段进行小波融合,最后选取支持向量机方法进行分类识别。【结果】美人松林、落叶松林、樟子松林、针叶混交林、阔叶混交林、火烧迹地、水体、耕地和未利用地9类地物识别的适宜光谱分辨率为40 nm,剩余5种地物识别的适宜光谱分辨率为80 nm,不同光谱区间对应的波段数大大降低,且最终分类精度总体都达到81%以上。【结论】将小波分析与支持向量机方法(SVM)结合,解决了高光谱存在的“维数灾难”问题,提高了高光谱数据的利用率,遥感分类中并不是使用的通道越多、波段越窄效果越好。

Suitability of band width for Hyperion image classification based on wavelet analysis

【Objective】Hyperspectral remote sensing is advantageous for fine identification of features, but has the problem of large data volume, high correlations among bands, and low processing accuracy and efficiency. In remote sensing classification, it is not the case that the use of a greater number of narrow bands ensures better classification perfor-mance. Therefore, the primary objective of the current work was to determine an appropriate bandwidth of Hyperion images by exploring the reduction of hyperspectral data, from the perspective of spectra.【Method】In this study, the preprocessing procedures for Hyperion images principally included the removal of uncalibrated and water vapor affected bands, eliminating the bad lines, stripes and smile effects, radiometric calibration and atmospheric correction. After this processing, 161 bands were retained to implement further geometric correction. Fourteen land use classes or schemes in the experimental area were selected according to field sample plot inventories, and a 7-scale Sym 3 wavelet decomposition was carried out on the average spectra of the 14 land use types. The spectral bands suitable for each type of recognition were mainly concentrated around 40 and 80 nm because of the variance and entropy of the wavelet detail coefficients. The narrow bands identified in the different spectral intervals were fused using the wavelet function to classify, by implementation, the support vector machines(SVM)algorithm.【Result】It was concluded that the appropriate spectral resolution to identify nine types of ground objects, such as Pinus syluestriformis forest, Larix gmelinii forest, Pinus sylvestris forest, mixed conifers forest, mixed broadleaf forest, fire burned area, water, arable land, and unused land, is 40 nm and the suitable spectral resolution for the remaining five kinds of terrain recognitions is 80 nm. The number of bands in different spectral intervals was greatly reduced. Further, the final classification accuracy was over 81%. 【Conclusion】The combination of wavelet analysis and SVM was helpful to solve the “dimensional disaster” issue with hyperspectral data, which indicates that the proposed method is feasible.