Deep-learning for automated diatom detection and identification for the ecological diagnosis of fresh-water environments

Abstract

Diatoms are a type of unicellular microalgae found in all aquatic environments. Their great diversity and ubiquity make these organisms recognized bio-indicators for monitoring the ecological status of watercourses, notably in the frame of the implementation of the European Water Framework Directive. In this context, we propose a study on diatom detection on microscope images using a deep-learning object detection architecture. To reduce the number of manually labeled images needed for training, we use a synthetic dataset in pair with a real one and gain more than 10% of precision and 5% of recall. This synthetic dataset represents a significant time saving, especially as it is made from publicly available images provided by diatom atlases, avoiding the extensive task of microscopic image acquisition. Diatom detection can be used for many tasks and notably for further classification of the extracted thumbnails by hand or using machine learning. To illustrate this use, we will also propose an update on automatic diatom classification using the latest advances in image classification. Finally, we will also discuss the applications of artificial taxonomy in the case of hierarchical diatom classification.