Ensemble Deep Learning Models for Fine-grained Plant Species Identification

Automated plant species identification for the datasets (images) collected from the natural environment is a challenging task. This study investigates the development and application of ensemble deep learning models for fine-grained plant species identification. Two different types of plant species datasets have been used in this study. The first dataset (UBD_45) consists of 45 medicinal plant species from the natural environment with the imbalanced distribution of classes and the second dataset (VP_200) has 200 medicinal plant species with balanced classes from the natural environment. Six popular deep learning models (InceptionResNetV2, ResNet50, Xception, InceptionV3, MobileNetV2, and GoogleNet) were trained on both datasets and heterogeneous ensembles with various ensemble techniques (mean, weighted mean, voting, and stacked generalization) were performed. The validation and testing accuracy results for individual models were compared with the output generated by the ensemble methods. The highest testing accuracies for base models were found 96.7% and 91.2% for UBD_45 and VP_200 datasets, respectively. Mean, weighted mean, and stacking ensembles showed better performance for both datasets. The stacking ensemble improved the classification accuracy by around 1.8% for the UBD_45 dataset while for VP_200 a significant improvement of around 4.23% was noticed using a weighted mean ensemble.