부산대학교 도서관

Nowadays it is known how to resolve many questions through satellite imagery such as Landsat 8 and the like, both from the theoretical point of view, i.e. research, as well as from the practical standpoint, e.g. commercial applications. This study evaluated the possibility of generalizing the training for supervised classification of multispectral images with sub-centimeter resolution. Images were taken under uncontrolled conditions of lighting and sun-target-sensor geometry and in the presence of normal interference in the agricultural environment. The images were obtained by the DuncanTech MS3100 camera (Auburn, CA, USA), a multispectral camera (green, red and near infra-red) mounted on a mobile ground platform and transformed into reflectance. For each element present (leaves, stems, spikes, soil, shadows, spectral references and sampling implements), a representative area was delimited in each image. These regions of interest were used, first, to quantify the separability of the classes. The next step was to define groups for cross-validation within these regions of interest; ten-folds were defined randomly with the constraint of a uniform distribution of classes. These folds were used in training and evaluation of the supervised classification using spectral angle mapper, maximum likelihood and decision trees. Spectral angle mapper correctly classified 49.2 % of cases, the maximum likelihood achieved a success rate of 86.8 % and the decision tree correctly classified 99.5 % of the spectral signatures. These results prove that multispectral images taken under uncontrolled conditions can be successfully classified by a generalized model that takes advantage of the higher spatial resolution. This opens a new line in which those pixels that do not correspond to vegetation, which bias the estimates of the crop parameters and complicate the recognition of objects, could be automatically masked. [ABSTRACT FROM AUTHOR]