Comparative Study of Non-Iterative Fine-Tuning Strategies in Mosquito Larvae Image Classification

A major challenge in classifying mosquito larvae images using deep learning is the limited data availability, high inter-species similarity, and intra-species variations, which leads to dataset scarcity. Transfer learning can overcome this, but relies on knowledge compatibility between the source and target domains, especially for non-iterative fine-tuning strategies. This study analyzes the performance of various non-iterative fine-tuning strategies to prevent negative transfer on a small annotated mosquito larvae dataset. We evaluate five strategies: Frozen Fine-Tuning, Full Fine-Tuning, First Half Fine-Tuning, Final Half Fine-Tuning, and Adaptive Partial Fine-Tuning on three pre-trained models, namely DenseNet-121, VGG-16, and Inception-V3. Fine-tuning performance is measured using accuracy, precision, sensitivity, specificity, and F1-score with 3-fold cross-validation. Results indicate that Adaptive Partial Fine-Tuning consistently outperforms other methods across all models, achieving maximum values in all metrics. Furthermore, Adaptive Partial Fine-Tuning on DenseNet-121 achieved perfect accuracy with an AUC of 1.0 across all classes, demonstrating its effectiveness in adjusting network weights for small annotated datasets in mosquito larvae classification.