Fine-grained image classification of microscopic insect pest species: Western Flower thrips and Plague thrips

Accurate identification of insect pests is essential in crop management as they are one of the primary causes of yield losses. However, differences between insect species demand different pest control strategies. Hence, research on new technology for fine-grained classification of insect pests is potentially important. Morphologically similar microscopic pest species classification has received little attention in the literature, and is targeted by this study as a means to address the need for agricultural pest management. We propose a novel computational method for deep learning-based, fine-grained classification of microscopic insects using the Vision Transform (ViT) architecture. This architecture employs an attention mechanism motivated by domain knowledge. The proposed approach consists of two main modules, a Data Preprocessing Module to segment relevant insect features and split the insect into body segments to inform identification, and a Domain Knowledge-Driven Stacked Model based on ViT to generate the prediction from each body segment and to fuse predictions for each segment into an accurate species-level classification. We tested the approach using an image dataset of two economically devastating thrip species - Western Flower thrips (Frankliniella occidentalis) and Plague thrips (Thrips imaginis). These insects are small (-1 mm), exhibit minute inter-species differences, and require different pest control strategies. We compared our model with the original ViT model, RestNet101, and RestNet50. Experimental results achieve an F1-score of 0.978, a 3.27% improvement over the baselines. This is important in the horticultural context given the yield losses that these pest insects are known to cause if their populations remain incorrectly quantified.