Active learning algorithm can establish classifier of blueberry damage with very small training dataset using hyperspectral transmittance data

The aim of this study was to estimate the performance of active learning algorithm for detecting blueberry damage using hyperspectral transmittance data with the very low labeling cost. A hyperspectral transmittance imaging system was first applied to collect the hyperspectral transmittance data of blueberries. Subsequently, the mean hyperspectral transmittance data was extracted. With only 9 labeled berries, the estimated error reduction could achieve the accuracy, precision and recall of 0.87, 0.93 and 0.78 respectively, and it consistently improved or maintained the performance of classifier for the remainder of the queries. In contrast to the SOM and SVM models, the classifier based on estimated error reduction also provided higher accuracy, precision and recall with the much fewer labeled samples. The active learning algorithms can be extended to the large scale applications in which the labeled samples are very limited or expensive and the models are required to be frequently transferred. In our case, due to the significant biological variations existing among blueberry samples, the classifier required frequent updates in practical applications, and the active learning algorithms could remarkably reduce label effort during the model updating processes.