Application of deep learning in aquatic bioassessment: Towards automated identification of non-biting midges

[Display omitted] •Identification of non-biting midges is time-consuming and requires expertise.•Convolutional Neural Networks was used to automate the identification process.•An automated identifier was built to identify chironomid species, genera, and subfamilies.•CNN models and humans relied on same morphological characteristics for identification.•The identifier based on deep learning classified taxa with extremely high accuracy. Morphological species identification is often a difficult, expensive, and time-consuming process which hinders the ability for reliable biomonitoring of aquatic ecosystems. An alternative approach is to automate the whole process, accelerating the identification process. Here, we demonstrate an automatic machine-based identification approach for non-biting midges (Diptera: Chironomidae) using Convolutional Neural Networks (CNNs) as a means of increasing taxonomic resolution of biomonitoring data at a minimal cost. Chironomidae were used to build the automatic identifier, as a family of insects that are abundant and ecologically important, yet difficult and time-consuming to accurately identify. The approach was tested with 10 morphologically very similar species from the same genus or subfamilies, comprising 1846 specimens from the South Morava river basin, Serbia. Three CNN models were built utilizing either species, genus, or subfamily data. After training the artificial neural network, images that the network had not seen during the training phase achieved an accuracy of 99.5% for species-level identification, while at the genus and subfamily level all images were correctly assigned (100% accuracy). Gradient-weighted Class Activation Mapping (Grad-CAM) visualized the mentum, ventromental plates, mandibles, submentum, and postoccipital margin to be morphologically important features for CNN classification. Thus, the CNN approach was a highly accurate solution for chironomid identification of aquatic macroinvertebrates opening a new avenue for implementation of artificial intelligence and deep learning methodology in the biomonitoring world. This approach also provides a means to overcome the gap in bioassessment for developing countries where widespread use techniques for routine monitoring are currently limited.