Automatic non-biting midge (Chironomidae) identification through the application of object detection and deep learning techniques

This research study introduces a possible new method for the identification of chironomid larvae mounted on microscope slides in the form of an automatic computer-based
identification tool using deep learning techniques. Deep learning is becoming an important tool for ecologists where there are advantages and limitations for its use as a rapid biomonitoring tool. Chironomids collected from the River Stour in Kent had their head capsules mounted on
microscope slides and images of these were then captured using a Raspberry PI. Using these images, a series of object detection models were created to classify several different
chironomid genera. These models were then used to show how different deep learning approaches, focusing on pre-training preparation, could improve the performance of image
classification. The model comparisons included two object detection frameworks (Faster-RCNN and SDD frameworks), three balanced image sets (with and without augmentation) and
variations of two hyperparameter values (Learning Rate and Intersection Over Union). All models were reported using the standard computer science object detection evaluation
protocol, the mean average precision metric. Each model configuration was run three times,to allow for statistical significance evaluation. Additionally, a series of novel post training performance metrics were created examining a model’s prediction accuracy and its givenconfidence value in its prediction choice. The highest mean average precision value achieved was 0.751 by Faster-RCNN. The models highlighted significance between the two object detection frameworks, where the Faster-RCNN framework performed better than SDD
framework; however, there was non-significance between the image sets and the hyperparameters values. All models produced similar accuracy results regardless of
framework used (between 95.5%-97.7%), however, there were large differences between the confidence examinations, wherein Faster-RCNN produced more confident predictions than
SSD. In conclusion, this investigation successfully developed object detection models using SSD and Faster-RCNN to classify between three chironomid genera. As a proof of concept, this study highlighted that automatic and rapid classification models using deep learning techniques can be applied for the correct taxonomic identification of difficult organisms, like
chironomid larvae, further advancing the prospect of using this relatively new field of computer science for ecological research.