Machine learning for improved size estimation of complex marine particles from noisy holographic images

Zonghua Liu^1,2*Marika Takeuchi²Yéssica Vanessa Contreras Pacheco³Thangavel Thevar⁴Alex Nimmo-Smith⁵John Watson⁴Sarah Lou Carolin Giering^2*

• ¹School of Computing, Engineering and Technology, Robert Gordon University, Aberdeen, United Kingdom
• ²Ocean BioGeosciences, National Oceanography Centre, Southampton, United Kingdom
• ³Division of Oceanography, Center for Scientific Research and Higher Education of Ensenada, Ensenada, Mexico
• ⁴School of Engineering, University of Aberdeen, Aberdeen, United Kingdom
• ⁵School of Biological and Marine Sciences, University of Plymouth, Plymouth, United Kingdom

Size estimation of particles and plankton is key to understanding energy flows in the marine ecosystem. A useful tool to determine particle and plankton size -besides abundance and taxonomy -is in situ imaging, with digital holography being particularly useful for micro-scale (e.g., 25 -2,500 µm) marine particles. However, most standard algorithms fail to accurately size objects in reconstructed holograms owing to the high background noise. Here we develop a machine-learning-based method for determining the size of natural objects recorded in digital holograms. A structured-forests-based edge detector is trained and refined for detecting the particle (soft) edges. A set of pixel-wise morphology operators are then used to extract particle regions (masks) from their edge images. Lastly, the size information of particles is calculated based on these extract masks. Our results show that the proposed strategy of training the model on synthetic and real holographic data improves the model's performance on edge detection in holographic images. Compared with another ten methods, our method has the best performance and is capable of rapidly and accurately extracting particles' regions on a group of synthetic and real holograms (natural oceanic particles), respectively (mean IoU: 0.81 and 0.76; standard-deviation IoU: 0.18 and 0.15). The processing time is near in-time: 15 s for 1,000 synthetic images and 1 s for 40 real images. This performance of the proposed method is also evaluated on the five natural oceanic particle images in each class of twenty classes: the best result comes from Class Fecal pellets with 0.93 of mean IoU, and 0.05 of standard-deviation IoU; the worst result is from Class Chainthin, but the mean IoU and standard-deviation IoU are still 0.58 and 0.16. Lastly, we use this method to analyze the size distributions of two vertical profiles of particle images recorded in the ocean.They exhibit similar trends in particle size distribution with a general decrease in particle size with increasing depth.