M.C. Codrea, T. Aittokallio, M. Keränen, E. Tyystjärvi, and O.S. Nevalainen (Finland)
machine learning, time-series, feature selection, plantspecies identification, classification.
In the present work, we study the effect of tempo ral changes on the identification accuracy in fluorescence fingerprinting of plants. Here, the time series comprises consecutive datasets (chlorophyll fluorescence transients), measured seven times, from plants of four distinct species, during an interval of six weeks. The raw data is subject to parameterization using the automatic feature learning sys tem, where the parameters of the feature extraction proce dure are tuned towards the maximization of the recognition accuracy. The optimization of the features is performed using time windows of successive datasets and the actual classification tests are done on the following time-points. An initial study on temporal changes in fluorescence finger printing [1] (Keranen et al. / Precision Agriculture, 2003) revealed, surprisingly, a non-monotonous trend in the iden tification accuracy which could not be explained by possi ble correlations with the physiological state or the growing conditions of the plant material. Therefore, we further in vestigate whether the oscillatory behavior of the classifica tion accuracy can be attenuated by using short-range and long-range time windows for optimizing features that have prediction capability.
Important Links:
Go Back
