Automatic Culicoides Biting Midges Classification Using Transfer Learning and Shallow Learning Techniques

Accurate identification of Culicoides species is essential for effective entomological monitoring, but remains challenging due to subtle morphological variation between species. This work proposes the development of a Culicoides species classification method based on a pre-trained ResNet-18 backbone for feature extraction and a set of four shallow learning classifiers to maximize predictive performance. The proposed method was trained and validated on a publicly available database, which was increased in the training partition by using five data augmentation operations. The highest mean F1-score of 0.964 ± 0.01 was obtained by the SVM classifier using a twotimes stratified five-fold cross-validation strategy. Also, it reached a successful classification result of F1-score = 0.966 on the test set, suggesting good generalization and ensuring the proposed method's output. The comparison against two previously developed state-of-the-art methods highlighted the proposed method's superior performance with an area under the receiver operating characteristic curve score of 0.99. The proposed method leveraged the transfer learning strategy to minimize the need for extensive image data in the model's training process and increased the deployment opportunities by implementing a lightweight and scalable classification scheme, which can be considered as a potential tool for classifying Culicoides species in the field.