Comparison of Data Augmentation Methods for Honeycomb Cell Classification using AlexNet

Monitoring colony health and productivity is critical for the sustainability of agricultural ecosystems in modern apiculture. The time-consuming and error-prone nature of traditional observational methods has increasingly necessitated the adoption of autonomous, deep learning-based solutions. In this study, a comprehensive analysis was conducted using the AlexNet architecture to automate the classification of honeycomb cells. To mitigate class imbalance within the original dataset, four distinct data augmentation techniques-Flip, Rotate, Gamma correction, and Gaussian Noise injection-were implemented, and their respective impacts on model performance were systematically evaluated. Utilizing a transfer learning strategy, pre-trained weights from ImageNet were employed, and the model was trained for 20 epochs on each dataset using an NVIDIA Tesla T4 GPU. Across the four augmented training sets, test accuracy improved from 0.90 on the baseline to 0.93 with Flip, Rotate and Gamma, and to 0.94 with Gaussian Noise; macro F1-score followed a similar pattern, rising from 0.83 on the baseline to 0.89 with Gaussian Noise. The gains were largest on the underrepresented classes (Class 2, Class 5, Class 6) where per-class F1 increased by more than 10 points. Within this dataset the augmented models clearly outperform the baseline, but confirming whether these gains transfer to images from different hives, cameras and lighting conditions would require external evaluation that we have not yet performed.