SPECIES IDENTIFICATION FOR AQUATIC BIOMONITORING USING DEEP RESIDUAL CNN AND TRANSFER LEARNING

Aquatic insects and other benthic macroinvertebrates are mostly used as bioindicators of the ecological status freshwaters. However, an expensive time-consuming process species identification represents one key obstacles for reliable biomonitoring aquatic ecosystems. In this paper, we proposed a deep learning (DL) based method that evaluated on several available public datasets (FIN-Benthic, STONEFLY9, EPT29) along with our Chironomidae dataset (CHIRO10). The relies three DL techniques to improve robustness when training is done relatively small dataset: transfer learning, data augmentation, feature dropout. We applied by employing ResNet-50 convolutional neural network (CNN) pretrained ImageNet 2012 dataset. results show significant improvement compared original contributions confirms there considerable gain multiple images per specimen.