An adaptive hierarchical loss for taxonomic labels and its application for plant identification

Hierarchical labels in classification have commonly appeared on popular datasets, particularly those derived from natural sources such as ImageNet [1], and MS COCO [2] datasets. From a biological perspective, the categorization of a species can be differentiated according to various taxonomic ranks, encompassing but not limited to species, genus, family, and order. However, this kind of information has not been well utilized when it comes to training deep neuron networks to improve the model’s accuracy. In this paper, we propose an adaptive-hierarchical loss that aims to achieve a balance between hierarchical constraints, such as parent-child and sibling relationships, while also considering the lowest leaf classes within the taxonomic ranks. We conducted an extensive evaluation of the proposed loss function using different network architectures including ResNet, DenseNet, SeNet, and Vision Transformer on two datasets: PlantClef2017 [3] and our forestry plant dataset, consisting of 10,000 and 156 classes respectively. The results indicate that the utilization of the adaptive-hierarchical loss during the training of neural networks leads to an improvement in accuracy of up to 7% for the PlantClef2017 dataset and 1% for the forestry dataset. Notably, our method generates dynamic weight factors for each level in the taxonomic tree, enabling us to discern the contribution of each level to the final accuracy.