Multivariate Classifiers Using Image Texture Features for Nitrogen Doses Discrimination in Wheat

摘要

Sidedress nitrogen fertilization is currently discussed all over the world due to its economical and environmental implications and the lack of current methods to determine N availability on the soil. The aim of this work was to study the nitrogen doses discrimination from the spectral response of the wheat dossel on visible and near infrared images, associated or not with texture features. The classification results were compared to that obtained with a portable chlorophyll meter and leaf nitrogen concentration. Data were collected in plots with three levels ofN in three dates (8, 14 and 20 days after sidedress fertilization). The images were processed using nine spectral indices and elaborated multivariate classifiers with the different features.From each index image were extracted the pixel's mean value and five texture measurements (2nd, 3rd, 4th, 5th and 6th moments of the histogram), so that a vector of features was created from each original image. The chlorophyll data and leaf nitrogen concentration were used in univariate classifiers. Statistical classifiers were developed to discriminate the classes defined by the N doses using combinations of the nine indices. The classifiers were evaluated using the leaving one out cross validation technique, creating the error matrix and the overall accuracy and the kappa index calculation. Using the images, it was possible to discriminate the N doses eight days after sidedress fertilization with no benefit of the use of texture features.