Moritz S. Schmid, PhD
Underwater imaging, deep learning, and plankton ecology at Oregon State University
Publications
Filter by type:
Prey and predator overlap at the edge of a mesoscale eddy: fine-scale, in-situ distributions to inform our understanding of oceanographic processes
Eddies can enhance primary as well as secondary production, creating a diverse meso- and submesoscale
seascape at the eddy front which can affect the aggregation of plankton and particles.
Due to the coarse resolution provided by sampling with plankton nets, our knowledge of plankton
distributions at these edges is limited. We used a towed, undulating underwater imaging system to
investigate the physical and biological drivers of zoo- and ichthyoplankton aggregations at the edge of
a decaying mesoscale eddy (ME) in the Straits of Florida. Using a sparse Convolutional Neural Network
we identified 132 million images of plankton. Larval fish and Oithona spp. copepod concentrations were
significantly higher in the eddy water mass, compared to the Florida Current water mass, only four
days before the ME’s dissipation. Larval fish and Oithona distributions were tightly coupled, indicating
potential predator-prey interactions. Larval fishes are known predators of Oithona, however, Random
Forests models showed that Oithona spp. and larval fish concentrations were primarily driven by
variables signifying the physical footprint of the ME, such as current speed and direction. These results
suggest that eddy-related advection leads to largely passive overlap between predator and prey, a
positive, energy-efficient outcome for predators at the expense of prey.
The intriguing co-distribution of the copepods Calanus hyperboreus and Calanus glacialis in the subsurface chlorophyll maximum of Arctic seas
Studying the distribution of zooplankton in relation to their prey and predators is challenging, especially in situ. Recent developments in underwater imaging enable such fine-scale research. We deployed the Lightframe On-sight Keyspecies Investigation (LOKI) image profiler to study the fine-scale (1 m) vertical distribution of the copepods Calanus hyperboreus and C. glacialis in relation to the subsurface chlorophyll maximum (SCM) at the end of the grazing season in August in the North Water and Nares Strait (Canadian Arctic). The vertical distribution of both species was generally consistent with the predictions of the Predator Avoidance Hypothesis. In the absence of a significant SCM, both copepods remained at depth during the night. In the presence of a significant SCM, copepods remained at depth in daytime and a fraction of the population migrated in the SCM at night. All three profiles where the numerically dominant copepodite stages C4 and C5 of the two species grazed in the SCM at night presented the same intriguing pattern: the abundance of C. hyperboreus peaked in the core of the SCM while that of C. glacialis peaked just above and below the core SCM. These distributions of the same-stage congeners in the SCMs were significantly different. Lipid fullness of copepod individuals was significantly higher in C. hyperboreus in the core SCM than in C. glacialis above and below the core SCM. Foraging interference resulting in the exclusion from the core SCM of the smaller C. glacialis by the larger C. hyperboreus could explain this vertical partitioning of the actively grazing copepodite stages of the two species. Alternatively, specific preferences for microalgal and/or microzooplankton food hypothetically occupying different layers in the SCM could explain the observed partitioning. Investigating the observed fine-scale co-distributions further will enable researchers to better predict potential climate change effects on these important Arctic congeners.
Use of Machine Learning for Predicting and Analyzing Ecological and ‘Presence Only’ Data: An Overview of Applications and a Good Outlook
2.1 Introduction
Over a decade ago, Leo Breiman (2001a) wrote: “There are two cultures in the use
of statistical modeling to reach conclusions from data. One assumes that the data
are generated by a given stochastic data model. The other uses algorithmic models
and treats the data mechanism as unknown. The statistical community has been
committed to the almost exclusive use of data models. This commitment has led to
irrelevant theory, questionable conclusions, and has kept statisticians from working
on a large range of interesting current problems. Algorithmic modeling, both in
theory and practice, has developed rapidly in fields outside statistics.”
Lipid load triggers migration to diapause in Arctic Calanus copepods—insights from underwater imaging
Copepod lipids fuel the Arctic marine ecosystem, but information on the fine-scale distribution of copepods and lipids is nonexistent. This study investigated the fine-scale (1 m) vertical distribution of the copepods Calanus hyperboreus , Calanus glacialis and Metridia longa during a Lagrangian drift in the North Water Polynya using the Lightframe On-sight Keyspecies Investigation (LOKI) imaging system. A copepod species- and stage-specific automatic identification model based on machine learning, a subcategory of artificial intelligence, was used to identify images taken by LOKI. Lipids were measured from images of copepods taken over the whole water column (1m resolution). Diel vertical migration (DVM) in all three species was detected. In C. hyperboreus and C. glacialis C4-females as well as M. longa C5-females lipid load of deep copepod individuals was significantly higher than that of shallower individuals. Vertical distribution profiles and individual lipid loads suggested that individuals with lower lipid load continued DVM, while others with high lipid load ceased migrating, remaining at depth. Calanus hyperboreus individuals seemed to migrate to diapause at lower lipid fullness (50%) than C. glacialis (60%). A bioenergetics model showed that Calanus females had enough lipids to diapause for over a year, highlighting the significant lipid overhead they use for capital breeding.
KEYWORDS: copepod lipids; DVM; diapause; fine scale vertical distribution; underwater imaging; machine learning; automatic zooplankton identification model; North Water Polynya; Arctic Ocean
Journal of Plankton Research is attributed as the original place of publication: https://doi.org/10.1093/plankt/fby012
Ensembles of Ensembles: Combining the Predictions from Multiple Machine Learning Methods
Growth and reproduction of the chaetognaths Eukrohnia hamata and Parasagitta elegans in the Canadian Arctic Ocean: Capital breeding versus income breeding
In Arctic seas, primary production and the availability of food for zooplankton are strongly pulsed over the short productive summer. We tested the hypothesis that Eukrohnia hamata and Parasagitta elegans, two similar and sympatric arctic chaetognaths, partition resources through different reproductive strategies. The two species had similar natural longevities of around 2 years. Eukrohnia hamata, which occurred at epi- and meso-pelagic depths, spawned two distinct broods in autumn and spring. Offspring production coincided with drops in the frequency of E. hamata with visible lipid reserves, characteristic of capital breeders. Growth was positive from April to January and negative in February and March. Growth and maturation were similar for the two broods. Storage reserves contained in an oil vacuole may allow E. hamata to reproduce and grow outside the short production season. Parasagitta elegans produced one brood in summer–autumn during peak production in near-surface waters, characteristic of income breeders. In winter, P. elegans co-inhabited meso-pelagic waters with E. hamata, where it neither grew nor reproduced. As the Arctic warms, the development of an autumn phytoplankton bloom could favour the summer–autumn brood of P. elegans. (Journal of Plankton Research is attributed as the original place of publication: https://doi.org/10.1093/plankt/fbx045)
The LOKI underwater imaging system and an automatic identification model for the detection of zooplankton taxa in the Arctic Ocean
Highlights
• An automatic zooplankton identification model has been developed for 114 taxonomic categories.
• The model successfully distinguishes species and stages.
• Various model validations show high model performance for identifying key zooplankton taxa.
• The model makes unprecedented insights into the fine scale vertical distribution of taxa possible.
Abstract
We deployed the Lightframe On-sight Keyspecies Investigation (LOKI) system, a novel underwater imaging system providing cutting-edge imaging quality, in the Canadian Arctic during fall 2013. A Random Forests machine learning model was built to automatically identify zooplankton in LOKI images. The model successfully distinguished between 114 different categories of zooplankton and particles. The high resolution taxonomical tree included many species, stages, as well as sub-groups based on animal orientation or condition in images. Results from a machine learning regression model of prosome length (R2=0.97) were used as a key predictor in the automatic identification model. Model internal validation of the automatic identification model on test data demonstrated that the model performed with overall high accuracy (86%) and specificity (86%). This was confirmed by confusion matrices for external testing results, based on automatic identifications for 2 complete stations. For station 101, from which images had also been used for training, accuracy and specificity were 85%. For station 126, from which images had not been used to train the model, accuracy and specificity were 81%. Further comparisons between model results and microscope identifications of zooplankton in samples from the two test stations were in good agreement for most taxa. LOKI’s image quality makes it possible to build accurate automatic identification models of very high taxonomic detail, which will play a critical role in future studies of zooplankton dynamics and zooplankton coupling with other trophic levels.
Also see researchgate at goo.gl/ZctMep for the publication!
The effect of particle properties on the depth profile of buoyant plastics in the ocean
Most studies on buoyant microplastics in the marine environment rely on sea surface sampling. Consequently, microplastic amounts can be underestimated, as turbulence leads to vertical mixing. Models that correct for vertical mixing are based on limited data. In this study we report measurements of the depth profile of buoyant microplastics in the North Atlantic subtropical gyre, from 0 to 5 m depth. Microplastics were separated into size classes (0.5–1.5 and 1.5–5.0 mm) and types (‘fragments’ and ‘lines’), and associated with a sea state. Microplastic concentrations decreased exponentially with depth, with both sea state and particle properties affecting the steepness of the decrease. Concentrations approached zero within 5 m depth, indicating that most buoyant microplastics are present on or near the surface. Plastic rise velocities were also measured, and were found to differ significantly for different sizes and shapes. Our results suggest that (1) surface samplers such as manta trawls underestimate total buoyant microplastic amounts by a factor of 1.04–30.0 and (2) estimations of depth-integrated buoyant plastic concentrations should be done across different particle sizes and types. Our findings can assist with improving buoyant ocean plastic vertical mixing models, mass balance exercises, impact assessments and mitigation strategies.
Available here: http://www.nature.com/articles/srep33882
or at researchgate: https://goo.gl/E3CY9N
Bridging the gap from student to senior scientist: recommendations for engaging early career scientists in professional biological societies
ZOOMIE v1.0 (Zooplankton Multiple Image Exclusion)
1. Introduction
ZOOMIE is an image treatment tool developed to ensure optimal quality for images collected with the Lightframe On-sight Keyspecies Investigation (LOKI) System, an underwater zooplankton camera system. ZOOMIE does that by identifying cases where multiple pictures of the same specimen have been taken (hereafter referred to as double images), a phenomenon that frequently occurs when imaging plankton in a constrained volume during vertical deployments. The process of identifying double pictures can be carried out manually but is very time consuming. By applying ZOOMIE, the time needed to identify double images is substantially reduced. It is essential to account for double images when representative distributions of images are wanted
ZOOMIE can automatically filter thousands of images based on previously extracted image parameters (e.g. area, mean grey pixel value, kurtosis; here extracted using the LOKI browser software (Isitec GmbH; http://www.isitec.de/start.htm)). The filtering is based on a set of rules that compares the image parameters of multiple images in order to detect double images and exclude them. The set of rules can be changed easily in the ZOOMIE scripts so that researchers can easily adapt the thresholds for finding double images necessary for their LOKI settings. After running the actual script to find double images, other scripts can be executed to automatically transfer images flagged for exclusion to a new folder.
Finally, the results can be visualized on an internal homepage, using the actual images which are linked to the database. Here we can validate the outcome of the processing and we can manually adapt the outcome through dragging and dropping of images to verify if any images were wrongly allocated to a double image group.
Although ZOOMIE was developed for LOKI images and the exclusion of double images, ZOOMIE could easily be adapted to handle other tasks requiring the handling and comparison of large numbers of images.
Assessments of Carbon Stock Hotspots in Nicaragua and Costa Rica
Abstract
Climate change is negatively affecting tropical regions through increasing temperatures and decreased precipitation leading to changes in local hydrology and decreasing water supply among others. In order to make accurate future predictions of carbon stock and forest health it is necessary to better understand the current underlying baseline carbon stock and how it may vary across space. Here we adapted an existing carbon stock assessment method and applied it to two tropical regions in Nicaragua and Costa Rica managed by the Maderas Rainforest Conservancy. Carbon stock was calculated based on 1) above-ground tree biomass, 2) above-ground sapling biomass, 3) leaf litter, herb and grass biomass, 4) soil organic carbon, 5) below-ground biomass, 6) stumps and deadwood and 7) regenerating plants. Our results show a strata-pooled average of 234.09 ± 379 Mg C ha-1 (n=40) carbon at the Costa Rican site and 209.20 ± 216 Mg C ha-1 (n=40) at the Nicaraguan site. These values are much higher than those available on a biome-wide scale, highlighting the extent of carbon stock loss outside these study areas as a result of anthropogenic disturbances, in comparison to more pristine areas. Local investigations into carbon stocks in the tropics are necessary to better estimate the current state of carbon content in the tropics. By adapting existing sampling protocols to local conditions this can be achieved efficiently. Furthermore, local estimates of carbon stock enable non-governmental organizations (NGOs) to participate in the Reducing Emissions from Deforestation and forest Degradation (REDD) program led by the United Nations.
A Short Introduction to Tropical Land- and Seascapes and Their Wildlife Conservation Management
There is more to the picture than meets the eye.
—a common saying
1.1 Introduction
The tropics eternally fascinate us. But tropical land- and seascapes mean many
things for many people (Forsyth and Miyata 1987; Kricher and Plotkin 1999). For
some, they can be a great home, a wonderful holiday, and a study site, while for
others they constitute a miserable life (with an average daily income of US$ 4) in a
life-threatening habitat (Collier 2007; Davis 2007). It is not an overstatement to say
that in the tropics, one can die easily. To the rest of the world, however, the tropics
still represent a land of opportunity (a “lebensraum”; Figs. 1.2 and 1.3)…..
A first overview of open access digital data for the Ross Sea: complexities, ethics, and management opportunities.
Abstract
It is now understood that the Ross Sea stands as one of the last relatively pristine (ocean) areas. Many decades of international research have been carried out under the Antarctic Treaty System stipulating that data acquired under this scheme must be shared with the global community. In line with Carlson (Nature 469:293, 2011, Polar Research 10.3402/polar.v32i0.20789, 2013), we find little evidence of enforcement towards making digital geographic information systems (GIS) project data available online for the wider Ross Sea ecosystem. While it is possible to find easily >40 digital datasets for most areas and pixels worldwide, despite many decades of research in the Ross Sea, only app. 100 digital datasets can be found for the study area. It simply shows that data from many studies in the region are not available. High-quality population and trend data explicit in space and time are mostly missing in the public realm, e.g., from the Commission for the Conservation of Antarctic Marine Living Resources (CCAMLR.org). This presents an ethical dilemma because it still appears that sufficient data exist for a pro-active and pre-cautionary management of this region. No coherent and efficient management scheme truly exists and is applied for this precious part of the world now heavily affected by global stressors and mismanagement of data and resources.
Open access data and machine learning models of 53 charismatic species in the Antarctic Ocean
Climate Change in the Arctic
Chapter 6.3 Zooplankton
Abstract of the feasibility study:
The research described in this feasibility report indicates that The Ocean Cleanup Array is a feasible and viable method to remove large amounts of plastic pollution from a major accumulation zone known as the Great Pacific Garbage Patch. Computer simulations have shown that floating barriers are suitable to capture and concentrate floating plastic debris. Combined with ocean current models to determine how much plastic would encounter the structure, a cleanup efficiency of 42% of all plastic within the North Pacific gyre can be achieved in ten years using a 100 km Array. In collaboration with offshore experts, it has been determined that this Array can be made and installed using current materials and technologies. The estimated costs are €317 million in total, or €31.7 million per year when depreciated over ten years, which translates to €4.53 per kilogram of collected ocean debris.
9.1. Climate change and predictions of pelagic biodiversity components
One of the powerful figures in the article. Here we see figure 4 a which shows predicted change from 2010 to 2100 based on future CanESM 2 data. The values shown here are mean predicted relative occurrence indeces (ROI) pooled over all 38 species that were modeled out. Warm colours show high predicted change and cool colours show lower change. The general trends of our study indicate a decline in ROI predictions for 2100. We think this represents an indication for a declining habitat quality and decreasing distribution range for traditional Antarctica species. One can see that eastern Antarctic waters are predicted to be among the most affected regions of change.
Seasonal observations and machine-learning-based spatial model predictions for the common raven (Corvus corax) in the urban, sub-arctic environment of Fairbanks, Alaska
Abstract
The common raven (Corvus corax) is an abundant generalist of the northern hemisphere, known to congregate and roost near human-related food sources. Due to a growing human-footprint and associated anthropogenic food subsidies, raven populations have increased dramatically over the past several years throughout the USA. The sub-arctic region has also witnessed increased urbanization and industrialization, and ravens have taken advantage of these changes. During 2004 and 2006, we surveyed parking lots on a bi-weekly basis in the city of Fairbanks in interior Alaska, showing an influx of ravens in winter. Between 2010 and 2012, we documented the presence and absence of ravens at a permanent set of 30 suspected raven locations and 21 randomized locations within the city limits of Fairbanks. We used machine learning (RandomForests) and 12 spatial GIS datasets from the Fairbanks North Star Borough to accurately model-predict the relative occurrence of ravens during winter and summer in Fairbanks. Our research showed a positive correlation between raven occurrence and commercial and residential zones in both winter and summer, as well as an inverse geographic relationship between ravens and the waste transfer station in the study area in winter, and a direct correlation near restaurants in summer. These results emphasize the link that ravens have with commercial, anthropogenic food sources, and how Fairbanks and its subsidized, urban habitat may be shaping part of the wider sub-arctic biodiversity landscape.
Keywords Common raven Fairbanks Alaska Distribution model Machine learning Subsidized predator
Marine Ecosystems and Climate Change
Schmid, MS (2012) One hundred seventy environmental GIS data layers for the circumpolar Arctic Ocean region
- Data includes for instance mixed layer depths, phytoplankton fluorescence, as well as predicted plankton distributions.