Improving Balanced Accuracy for Minority Plant Species Under Data Imbalance: A Multi-Architectural Ensemble Approach

Imbalanced datasets pose significant challenges for training classifier models. In this paper, we investigate the efficacy of ensemble strategies to enhance classification performance for both majority and minority plant species under data imbalance. Using two annotated datasets, CR Leaves and PlantCLEF 2017, we explore novel ensemble approaches. Our experiments reveal noteworthy improvements over baselines: for the CR Leaves dataset, the ensemble achieved a balanced accuracy for minority species of 0.777 (+16.6%), and for the PlantCLEF 2017 dataset, the ensemble achieved a balanced accuracy for minority species of 0.363 (+5.2%). Our findings highlight the potential of ensem-ble strategies to rectify imbalanced classifications and improve performance for minority species, with implications for various domains beyond plant species identification.