In the thesis, we analyzed the impact of noise data on the learning of classification models. We focused on the classification of crops from multispectral satellite images. The existing neural network architecture for crop classification is adapted so that it can be learned with uncertain class designations derived from common surface boundary cells containing two or more crop classes. The method was evaluated on a dataset that includes the entire surface of Slovenia, the evaluation of uncertainty in labels improves the classification accuracy by 5\%, which opens new possibilities for more robust learning of predictive models on similar problems.
|
