Abstract

Convolutional Neural Networks (CNNs), a prominent deep learning architecture, are progressively establishing themselves as the gold standard for the detection and quantification of objects within digital images. However, a preponderance of research in the field tends to train and evaluate these neural networks using data derived from a single image source, thereby impeding the ability to generalize model performance across more heterogeneous contexts. The primary aim of this investigation was to examine the robustness of models when trained on data from a variable number of sources. To this end, images of yellow sticky traps containing psyllids and a wide variety of other objects were procured using nine disparate devices. Models were subsequently trained and tested employing diverse combinations of this data. The findings from these experiments enabled the drawing of several conclusions regarding optimal training procedures and the influence of data quantity and variety on the robustness of the trained models.