Methods in Oceanography最新文献

Comparison of profiles from oxygen Optodes deployed on profiling floats with ship-based bottle casts taken at the time of deployment shows typical low biases of approximately 0 to −40 μmol kg⁻¹. Most proposed methods to correct these biases use linear or multiplicative corrections of the derived variable $O_2$. Some of these methods depend on specific reference data such as deployment casts or air measurements. Here, we describe a versatile in situ method to recalculate $O_2$ directly from Optode phase and temperature by recalibrating two coefficients of the modified Stern–Volmer equation. This method may be used to calibrate most floats deployed with Optodes to date, as well as present floats, including those equipped with air measurement capability. Reference data can be taken from historic ship casts, climatology, deployment casts, and/or air measurements, depending on availability.

In situ calibrations were performed on 147 Optodes floats deployed on UW floats between 2004 and 2015 using only World Ocean Database (WOD) references. Median differences to World Ocean Atlas (WOA) 2009 climatology were reduced from ∼6% to ∼1%. Deployment casts were used to estimate error for eight Argo floats deployed in the Indian and Pacific Oceans; the aggregate error was reduced from 8% to 0.3%.

Comparison of six pairs of Optodes deployed on the same float showed relative errors after in situ calibration of $0.1\pm 0.6\mu \mathrm{mol}{\mathrm{kg}}^{-1}$. WOD-calibrated surface air oxygen values for nineteen Optode floats with air-measurement capability were compared with expected oxygen levels from NCEP surface level pressures and showed typical errors of $<\pm 2\%$.

Using data from eight floats with deployment casts, comparison of phase-domain linear correction with oxygen-domain linear correction showed a difference of less than $\sim \pm 2\%$. Comparison of surface gain correction with deployment casts found gain-corrected values below the depth of the oxygen minimum to be consistently low, with residuals of approximately −0.5 to −4.5%.

Locating marine organisms in their natural habitats is important for understanding ocean biodiversity. Many species are often camouflaged in their surroundings, rendering them hard to detect. Our increasing ability to image large areas of the ocean floor produces millions of images, which must be inspected to spot the occasional organism. This calls for automation of camouflage detection. We investigate reliable detectability of marine camouflage by looking for structural regularities as cues to locating organisms in their natural settings. We study skates and flounder, which use different mechanisms to avoid detection. We introduce a simple edge-based criterion for detecting local structural regularity to reduce the image area to be inspected for likely presence of camouflaged organisms. This sets the stage for efficient use of more complex algorithms to confirm detections and aid in marine census. We also study the possibility of detecting octopuses based on a simple measure of texture applied to a hierarchical segmentation of octopus images.

Imaging systems were developed to explore the fine scale distributions of plankton (<10 m), but they generate huge datasets that are still a challenge to handle rapidly and accurately. So far, imaged organisms have been either classified manually or pre-classified by a computer program and later verified by human operators. In this paper, we post-process a computer-generated classification, obtained with the common ZooProcess and PlanktonIdentifier toolchain developed for the ZooScan, and test whether the same ecological conclusions can be reached with this fully automatic dataset and with a reference, manually sorted, dataset. The Random Forest classifier outputs the probabilities that each object belongs in each class and we discard the objects with uncertain predictions, i.e. under a probability threshold defined based on a 1% error rate in a self-prediction of the learning set. Keeping only well-predicted objects enabled considerable improvements in average precision, 84% for biological groups, at the cost of diminishing recall (by 39% on average). Overall, it increased accuracy by 16%. For most groups, the automatically-predicted distributions were comparable to the reference distributions and resulted in the same size-spectra. Automatically-predicted distributions also resolved ecologically-relevant patterns, such as differences in abundance across a mesoscale front or fine-scale vertical shifts between day and night. This post-processing method is tested on the classification of plankton images through Random Forest here, but is based on basic features shared by all machine learning methods and could thus be used in a broad range of applications.

In this article Dr. Jules S. Jaffe chronicles his development as a scientist and engineer. The story starts during his middle school years and continues up until the present day. Dr. Jaffe, as an inventor of technology for ocean exploration has played a role in a number of advances in ocean engineering. These range from the development of a planar laser imaging system for sensing fluorescent microstructure to swarms of underwater autonomous floats, to a current generation of underwater microscopes. The emphasis of the article is on career development and the process rather than the exact, and detailed, documentation of technology. Dr. Jaffe is also the Editor in Chief of Methods in Oceanography and he instituted these autobiographies for exactly this purpose: To give younger, aspiring, professionals an example of a career that has not been “straight through”, but rather a meandering path through a multitude of projects, proposals, and relationships with colleagues, students, and funding agencies.

Cameras can be considered measurement devices complementary to acoustic sensors when it comes to surveying marine environments. When calibrated and used correctly, these visual sensors are well-suited for automated detection, quantification, mapping, and monitoring applications and when aiming at high-accuracy 3D models or change detection. In underwater scenarios, cameras are often set up in pressure housings with a flat glass window, a flat port, which allows them to observe the environment. In this contribution, a geometric model for image formation is discussed that explicitly considers refraction at the interface under realistic assumptions like a slightly misaligned camera (w.r.t. the glass normal) and thick glass ports as common for deep sea applications. Then, starting from camera calibration, a complete, fully automated 3D reconstruction system is discussed that takes an image sequence and produces a 3D model. Newly derived refractive estimators for sparse two-view geometry, pose estimation, bundle adjustment, and dense depth estimation are discussed and evaluated in detail.

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 ($R^2=0.97$) 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.

The automated processing of images for scientific analysis has become an integral part of projects that collect large amounts of data. Our recent study of cuttlefish camouflaging behavior captured ∼12,000 images of the animals’ response to changing visual environments. This work presents an automated segmentation and classification workflow to alleviate the human cost of processing this complex data set. The specimens’ bodies are segmented from the background using a combination of intensity thresholding and Histogram of Oriented Gradients. Subregions are then used to train a texton-based classifier designed to codify traditional, manual methods of cuttlefish image analysis. The segmentation procedure properly selected the subregion from ∼95% of the images. The classifier achieved an accuracy of ∼94% as compared to manual annotation. Together, the process correctly processed ∼90% of the images. Additionally, we leverage the output of the classifier to propose a model of camouflage display that attributes a given display to a superposition of the user-defined classes.

This paper demonstrates how computer vision can be applied for the automatic detection of shrimp in smaller areas of interest with a high temporal resolution for long time periods. A recorded sequence of digital HD camera images from fixed underwater observatories provides unique opportunities to study shrimp behavior in their natural environment, such as number of shrimp and their abundance at different locations (micro habitats) over time. Temporal color contrast features were applied to enable the detection of the semi-transparent shrimp. To study the spatial–temporal characteristics of the shrimp, pseudo-color visualizations referred to as shrimp abundance maps (SAM) are introduced. SAMs for different time periods are presented, to show the potential of the methodology.

Underwater image analysis is a new field for computational pattern recognition. In academia as well as in the industry, it is more and more common to use camera-equipped stationary landers, autonomous underwater vehicles, ocean floor observatory systems or remotely operated vehicles for image based monitoring and exploration. The resulting image collections create a bottleneck for manual data interpretation owing to their size.

In this paper, the problem of measuring size and abundance of poly-metallic nodules in benthic images is considered. A foreground/background separation (i.e. separating the nodules from the surrounding sediment) is required to determine the targeted quantities. Poly-metallic nodules are compact (convex), but vary in size and appear as composites with different visual features (color, texture, etc.).

Methods for automating nodule segmentation have so far relied on manual training data. However, a hand-drawn, ground-truthed segmentation of nodules and sediment is difficult (or even impossible) to achieve for a sufficient number of images. The new ES4C algorithm (Evolutionary tuned Segmentation using Cluster Co-occurrence and a Convexity Criterion) is presented that can be applied to a segmentation task without a reference ground truth. First, a learning vector quantization groups the visual features in the images into clusters. Secondly, a segmentation function is constructed by assigning the clusters to classes automatically according to defined heuristics. Using evolutionary algorithms, a quality criterion is maximized to assign cluster prototypes to classes. This criterion integrates the morphological compactness of the nodules as well as feature similarity in different parts of nodules. To assess its applicability, the ES4C algorithm is tested with two real-world data sets. For one of these data sets, a reference gold standard is available and we report a sensitivity of 0.88 and a specificity of 0.65.

Our results show that the applied heuristics, which combine patterns in the feature domain with patterns in the spatial domain, lead to good segmentation results and allow full automation of the resource-abundance assessment for benthic poly-metallic nodules.