biunit / dragonfly

Species identification model for Japanese dragonflies and damselflies

Dragonfly Classification

Usage

Preparation

The pre-trained weights for species or genus classification models can be downloaded with the following scripts. File names ending with _resnet152.pth and _vgg19.pth are the weights of PyTorch models of ResNet152 and VGG19, respectively. File names starting with meshmatrix_ are the summary of ecological survey data of dragonflies and damselflies.

mkdir weights
wget -P ./weights https://bitdessin.dev/storage/dragonfly/species_resnet152.pth
wget -P ./weights https://bitdessin.dev/storage/dragonfly/species_vgg19.pth
wget -P ./weights https://bitdessin.dev/storage/dragonfly/genus_resnet152.pth
wget -P ./weights https://bitdessin.dev/storage/dragonfly/genus_vgg19.pth
wget -P ./weights https://bitdessin.dev/storage/dragonfly/meshmatrix_species.tsv.gz
wget -P ./weights https://bitdessin.dev/storage/dragonfly/meshmatrix_genus.tsv.gz

Species Identification

To predict species name of dragonflies and damselflies with ResNet152 model, run the following scripts. Note that it can use VGG19 model for prediction by changing resnet152 to vgg19 in the scripts. The prediction result will be saved in inf_probs.txt specified with -o option.

python predict.py --class-label  classes_species.txt             \
                  --model-arch   resnet152                       \
                  --model-weight ./weights/species_resnet152.pth \
                  -i data/dataset_T/example_01.jpg               \
                  -o inf_probs.txt

To perform the prediction with combined model (i.e., image model and additional ecological filtering), add --mesh option and run the following scripts.

python predict.py --class-label  classes_species.txt                 \
                  --model-arch   resnet152                           \
                  --model-weight ./weights/species_resnet152.pth     \
                  --mesh         ./weights/meshmatrix_species.tsv.gz \
                  -i data/dataset_T/example_01.jpg                   \
                  -o inf_probs.txt

Genus Identification

To predict genus of dragonflies and damselflies with image models, run the following scripts with the model weight for the genus level (e.g., genus_resnet152.pth).

python predict.py --class-label  classes_genus.txt             \
                  --model-arch   resnet152                     \
                  --model-weight ./weights/genus_resnet152.pth \
                  -i data/dataset_T/example_01.jpg             \
                  -o inf_probs.txt

To use the combined model, add --mesh option and run the following scripts.

python predict.py --class-label  classes_genus.txt                 \
                  --model-arch   resnet152                         \
                  --model-weight ./weights/genus_resnet152.pth     \
                  --mesh         ./weights/meshmatrix_genus.tsv.gz \
                  -i data/dataset_T/example_01.jpg                 \
                  -o inf_probs.txt

Training

python train.py --class-label   classes_species.txt                 \
                --model-arch    resnet152                           \
                --model-outpath ./weights/example_model.pth         \
                --traindata     ./data/dataset_W1/augmentated_image \
                --validdata     ./data/dataset_F/raw                \
                --epochs 5 --batch-size 32 --lr 0.001

Citation

@article{Sun_2021,
    doi = {10.3389/fevo.2021.762173},
    url = {https://doi.org/10.3389/fevo.2021.762173},
    year = 2021,
    volume = {9},
    author = {Sun, Jianqiang and Futahashi, Ryo and Yamanaka, Takehiko},
    title = {Improving the Accuracy of Species Identification by Combining Deep Learning With Field Occurrence Records},
    journal = {Frontiers in Ecology and Evolution}
}