Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Abstract. This paper proposes an online unsupervised deep learning approach for automatically collecting new image data for continuously training more accurate image classifier models of a multi-class insect identification algorithm. Once images are collected by the sticky paper trap image monitoring system, insect images are cropped from the images and screened out based on an unsupervised data collection technique. The proposed unsupervised data collection technique uses reference models trained with supervision for the collection of images based on different classification probability threshold settings. After testing 15 months of data, it was found that reference models trained with 8 months of image data could consistently perform unsupervised data collection with minimal collection error. After 7 months of the unsupervised online training cycle, the technique was able to improve the average F₁-score of the classifier models from 0.88 to 0.926, close to the 0.94 F₁-score achieved with supervision. The proposed technique reduces the effort required to train new image classifier models, as it does not require the traditional manual collection of new images. The presented technique can be applied to not only insect identification, but also other image monitoring applications.