A Few-Shot Learning Approach for Red Palm Weevil Recognition in Low-Data Scenarios

The Red Palm Weevil (RPW), a destructive pest significantly harming palm trees making its early detection important. Unfortunately, labelled datasets for RPW recognition are limited. Therefore, this paper presents a transfer learning-based approach using pretrained ResNet50 and EfficientNet-B0 models to recognize RPW in low data scenarios. Experimental results show that fine tuning using few shot learning techniques (MAML and Prototypical Networks) surpass traditional supervised methods delivering high accuracy despite limited data. Specifically, EfficientNet-B0 and MAML model produced the highest accuracy score of 91.3%, followed by EfficientNet-B0 and Prototypical Networks model at 90.8%, ResNet-50 and MAML model at 90.1%, and ResNet-50 with Prototypical Networks model at 89.5%.