Résumé
Multispectral autofluorescence images can be acquired using automated microscopes with a spatial resolution of 1µm and a large field of view (1 to 2 cm 2 ). Such acquisition systems are well suited to observe whole organ sections, while enabling tissue identification. They allowed the acquisition of an increased number of images, although the sample preparation is still the bottelneck. These large image series pave the way for statistical analysis, and could be used to explore the diversity of plant structure and cell wall spectral properties as a function of different factors (genetic, environmental and agricultural conditions), but also on technological fractions obtained from these plant ressources. To quantify the differences between images we proposed to take advantage of the pixel score distributions after Principal Component Analysis (PCA). Series of 30-60 large images may contain millions to billions of pixels. Principal Component Analysis can be applied to image series by iteratively computing the variance-covariance matrix and the score images [1]. It is called large PCA. For each principal component, the pixel score distribution across the entire image series and its percentiles are determined. Local PCA score distributions are then calculated for each image using these percentiles. These local score distributions are regarded as quantitative characteristics enabling multispectral images to be compared. Two examples based on maize stem section and ground maize fractions are presented. Autofluorescence multispectral images were acquired and specific attention was paid to UV and visible fluorescence variability. First of all, 4 maize inbred lines were compared based on 40 stem imaged sections. The overall comparison was carried out by looking at the principal component analysis of the score distributions. Similar approach was used to compare 6 grinding fractions obtained from 6 maize inbred lines.
Pixel score distributions are found to be a promising tool for statistical comparison of multispectral images. The method is easy to implement, can be used as a first descriptive analysis of the image series, and can be easily extended to other spectral imaging techniques.