dc.contributor | Universidade Estadual Paulista (UNESP) | |
dc.contributor | Universidade Federal de Uberlândia (UFU) | |
dc.date.accessioned | 2022-04-28T19:52:52Z | |
dc.date.accessioned | 2022-12-20T01:41:10Z | |
dc.date.available | 2022-04-28T19:52:52Z | |
dc.date.available | 2022-12-20T01:41:10Z | |
dc.date.created | 2022-04-28T19:52:52Z | |
dc.date.issued | 2022-05-01 | |
dc.identifier | Computers and Electronics in Agriculture, v. 196. | |
dc.identifier | 0168-1699 | |
dc.identifier | http://hdl.handle.net/11449/223751 | |
dc.identifier | 10.1016/j.compag.2022.106901 | |
dc.identifier | 2-s2.0-85127470718 | |
dc.identifier.uri | https://repositorioslatinoamericanos.uchile.cl/handle/2250/5403880 | |
dc.description.abstract | Supervised learning-based methods can identify crop threats in the visual data collected by an Unmanned Aerial Vehicle (UAV). However, as these methods induce classification models from a finite set of a priori known classes, they cannot recognize new patterns emerging in visual data to be classified. In agricultural environments, these patterns may appear over time, so that those related to diseases/pests should be addressed by the classifier timely. This study investigates an extension of a semi-supervised classification algorithm to identify new classes of threats appearing in UAV visual data. To do so, the algorithm aggregates information from clusters with Support Vector Machine (SVM) outcomes operating on the unlabeled (target) data. From an iterative active learning procedure, the classification model is then fed back to learn a new class. Experimental results showed that our algorithm can discover a new threat, named Ceratocystis wilt, in Eucalyptus plantations even with labeled data scarcity and class imbalance. Also, even this new class being the minority one, its error rate was reduced to almost zero in few iterations on a tested dataset. This is due to the adopted Entropy and Density-based Selection approach, which explored the new class better than an SVM Margin Sampling baseline. When operating on VGGNet-16 deep features, our algorithm achieved accuracies between 92% and 97% being slightly better than those results based on hand-crafted features. | |
dc.language | eng | |
dc.relation | Computers and Electronics in Agriculture | |
dc.source | Scopus | |
dc.subject | Active learning | |
dc.subject | New threats classification | |
dc.subject | Novelty detection | |
dc.subject | Semi-supervised learning | |
dc.title | Novelty detection in UAV images to identify emerging threats in eucalyptus crops | |
dc.type | Artículos de revistas | |