叶片含水率推扫式高光谱成像去条纹标定法优化

由推扫式高光谱成像系统所采集的图像中会出现特有的条纹噪声，这些噪声会穿过化学计量学模型，最终出现在反映被测指标空间分布情况的可视化预测图中，干扰其空间特征的呈现及解读。以银杏叶含水率为例，基于偏最小二乘回归（PLSR）预测模型，将经去条纹标定法处理后的图像分别与原始图像及经传统均值滤波增强后的图像进行比较，研究去条纹标定法对化学计量学指标空间分布预测的改进作用。去条纹标定法和传统均值滤波增强不会对感兴趣区域均值PLSR预测模型决定系数R_p²产生明显影响，其随主成分数增加，呈先增后减趋势，当主成分数为10时R_p²均达到最大，且预测准度相当。将化学计量学模型应用到像素光谱，进行指标空间分布预测时，随主成分数由6增至10,模型的波段增益系数逐渐增大，导致化学计量学可视化图像中条纹噪声逐渐增加：在由原始图像或经传统均值滤波增强图像得到的含水率可视化图像中，条纹噪声逐渐增加，甚至完全湮没叶面内部含水率空间分布信息；而去条纹标定法能够明显抑制本征条纹噪声，即使当主成分数增加到8时（R_p

Method of De-stripe Calibration Applied in Water Content Spatial Visualization in Ginkgo Leaf on Spectral Imagery

A distinctive spatial noise pattern in the form of parallel stripes exists commonly in the images that are acquired using pushbroom hyperspectral imaging systems.Passing through chemometric systems,it often resurfaces in resultant images of the spatial distributions of various chemical or quality indices,blocking or breaking the spatial details therein,and undermining consequent interpretation.In regard of this,an image de-striping calibration was investigated for its contribution to improving spatial chemometric predictions.The de-stripe calibration was first applied to the hyperspectral images of 155 ginkgo leaves before mapping the spatial distribution of water content(WC)using partial least squares regression-chemometric models.In comparison,the process was repeated twice,respectively,from either raw hyperspectral image without de-stripe calibration or those through a conventional image-enhancement of spatial smoothing-filtering.Results showed that neither the de-stripe calibration nor the conventional image enhancement would affect the accuracy of chemometric models,and that the coefficient of determination of prediction,or R²_P,irrespective of different preprocessing in all three cases,were risen up with the increase of number of principal components(PCs),until peaking at the number of 10 PCs(R²_P=0.93~0.94).However,difference emerged when applying chemometric models to the spectra at individual pixels to map the spatial distribution of WC over leaf-surface.As the number of PCs was increased from 6 to 10,so did the spectral gains of chemometric models causing strengthening stripy noise in the WC maps from either the un-treated or conventionally smooth-filtered images,with noise-stripes being the most prominent spatial feature at 8 PCs,and even deteriorating to the point,at 9 or 10 PCs,that any possible WC variation over a leaf would be totally blocked up.To the contrary,the de-stripe calibration successfully suppressed the distinctive noise patterns inherent from the pushbroom hyperspectral imaging system,so that no discernible stripes appeared in the WC maps from the de-striped hyperspectral images and stunning spatial details were preserved in the maps derived from the relatively high accuracy chemometric model of 8 PCs(R²_P=0.88).It may be safely concluded from this comparative study that de-stripe calibration of pushbroom hyperspectral images did contribute rich spatial details and high accuracy to spatial chemometric predictions through keeping spatial details intact while enabling the application of models with high spectral gains.