ResNet-152 model that was trained to classify images of 162 butterfly and moth species that occur in Austria.
Model Details
ResNet-152 pre-trained on ImageNet was used and a full fine-tuning of the pre-trained ResNet-152 model, with all parameters rendered trainable, was conducted. The model was trained for a master thesis in the ULG data science of the University of Innsbruck. A link to the thesis will be added when it is published.
Model Description
- Developed by: Friederike Barkmann, Andreas Lindner
- Funded by:
- Viel-Falter Butterfly Monitoring which is financially supported by the Federal Ministry for Climate Action, Environment, Energy, Mobility, Innovation and Technology (BMK).
- EuroCC Austria which has received funding from the European High Performance Computing Joint Undertaking (JU) and Germany, Bulgaria, Austria, Croatia, Cyprus, Czech Republic, Denmark, Estonia, Finland, Greece, Hungary, Ireland, Italy, Lithuania, Latvia, Poland, Portugal, Romania, Slovenia, Spain, Sweden, France, Netherlands, Belgium, Luxembourg, Slovakia, Norway, Türkiye, Republic of North Macedonia, Iceland, Montenegro, Serbia under grant agreement No 101101903.
- License: cc-by-4.0
- Finetuned from model: ResNet-152 pre-trained on ImageNet
Uses
The model can be used to identify butterfly and moth species that occur in Austria. It can classify 162 species, 131 of which are butterflies.
Bias, Risks, and Limitations
The model does not cover all butterfly and moth species that occur in Austria. Of the about 210 butterflies, 131 were used for model training. Of the about 4000 moth species it were only 31. Not all butterfly and moth species can be determined based on images alone.
Training Details
The model was trained on the EuroHPC supercomputer LEONARDO, hosted by CINECA (Italy) and the LEONARDO consortium. Training was parallelized using the Pytorch DDP framework and the Hugging Face Accelerate library. See https://github.com/FriederikeBarkmann/CNN_butterfly_identification for the scripts that were used for model training.
Two methods to balance the distribution of the imbalanced dataset were applied. In one approach, oversampling of minority classes and undersampling of majority classes during training was used. In another a wighted loss function was applied. In addition, the model was trained without correction of data imbalance based on the original distribution of the data.
Training Data
The model was trained with a dataset of over 500,000 images of butterflies and moths that were recorded in Austria. The images were taken by users of the App "Schmetterlinge Österreichs" of the foundation "Blühendes Österreich" all over Austria. Images that showed more than one species or showed butterfly or moth eggs, larvae and pupae were excluded from training. Species with less than 50 images were excluded from training. The final dataset contains images of the adult life stages of 162 species (31 moth species and 131 butterfly species).
BibTeX:
Will be updated when the thesis is published.
APA:
Will be updated when the thesis is published.