Suggested Citation:
Daly, K.L., K. Kramer, A. Remsen. 2018. SIPPER: Oil-Marine Snow-Mineral Aggregate Interactions and Sedimentation during the 2010 Deepwater Horizon Oil Spill in the northeastern Gulf of Mexico from May-June 2010. Distributed by: Gulf of Mexico Research Initiative Information and Data Cooperative (GRIIDC), Harte Research Institute, Texas A&M University–Corpus Christi. doi:10.7266/N779437K
Data Parameters and Units:
Project name, Principal investigator, Institution, Location, Cruise, Date (local and GMT), Time (local and GMT), Latitude (decimal degrees N), Longitude (decimal degrees W), Station, Depth of images (m), Volume filtered (m^3), Marine snow particle size (area mm2), Abundance (count/m^3)
Methods:
A. SIPPER Field Deployments The SIPPER camera imaging system was developed by the University of South Florida, Center for Ocean Technology and described in Samson et al. (2001) and Remsen et al. (2004). The towed platform carried several environmental sensors (CTD, oxygen, chlorophyll fluorescence, transmissometer: see Supplemental Information- Instruments). SIPPER used a high speed Dalsa Piranha-2 line-scan camera and a pseudo-collimated LED generated light sheet to image the shadows and outlines of resolvable particles that passed through a 100 cm2 field of view. The operational optical resolution of the system is ~65 um. SIPPER was towed at speeds between 2-3 knots in an oblique profile through the water column, spending approximately equal amounts of time at each meter of depth between the surface and 300 m. At stations with a bottom depth shallower than 300 m, SIPPER was towed within approximately 5 m from the seafloor. Imaging and environmental data were stored internally on a Firewire hard drive and processed upon retrieval of the SIPPER instrument from a deployment using a customized software package called the Plankton Image Classification and Extraction Software (PICES). PICES was used to extract images of interest, classify them using user-specified training libraries and to manage the SIPPER images and environmental data collected. B. SIPPER Data Analyses SIPPER data, including concurrently collected environmental instrument data, were offloaded to a desktop PC after every deployment. Using the Plankton Instrument Classification and Extraction Software (PICES) (Kramer et al. 2011), images greater than or equal to 250 pixels in total area (~0.5 mm equivalent spherical diameter or ESD) were extracted from the raw SIPPER data, preliminarily classified and automatically entered into a MYSQL based image database. A custom database management application called PICES Commander was used to manage and classify the resulting dataset. Classification of SIPPER images involved several steps including both automated classification and manual labeling of the images. (1) Image Extraction and Preliminary Classification: During the image extraction step, features for each image are determined and entered into a feature table in the PICES database and an initial classification for each image is predicted using a comprehensive, multi-class feature selection (MFS) support vector machine (SVM) (Kramer et al., 2011). This classifier was initially developed from training images collected in the Gulf of Mexico prior to the Deepwater Horizon oil spill. Training images and new image classes were added throughout the sampling period and slowly replaced the images that were collected before the spill. We utilized a hierarchical naming format for the particle and plankton image classes to reflect the level of similarity between image classes. More closely related image classes shared one or more parts of their name separated by underscores and this naming schema is used in the classification effort. (2) Manual Validation: After image extraction and the first run of the comprehensive classifier, thumbnail images of the resulting classifications were then examined using the image browsing and validation capability of PICES Commander. Image classes that consisted of numerous true positive examples (for example calanoid copepods or detritus) were noted for inclusion in the final classification effort and then ignored while image classes that were less abundant or rare were closely examined. True positive image examples from these rarer image classes were validated, meaning that their image class label could not be relabeled during a subsequent classification. This was done for every predicted image class until all rare images were validated from the first comprehensive classification. Occasionally, a new image class was encountered. When this occurred, that class was added to the comprehensive classifier and any encountered examples would be validated and added to the training library (3) Run Final Comprehensive MFS-SVM Classification: After all deployments had their images extracted, initially classified, and new image classes encountered, labeled, and added to the training library, the final comprehensive MFS-SVM classifier was rebuilt and rerun on all deployments using all the new training examples that had been uncovered in Step 2. The purpose of this classification cycle was to identify which common classes to include in the ultimate dual classifier, identify deployments where less common image classes might be abundant enough to include in the ultimate dual classifier, and validate any new rare image examples that may have been identified during the second run of the comprehensive classifier. (4) Build Deployment-Specific MFS-SVM Classifiers: For each deployment, after rare and uncommon groups were searched for true positives and validated in Step 3, a MFS_SVM classifier was pared down to the most common classes found in that deployment and rebuilt. New MFS classifiers were, therefore, built for each deployment with classes specific to that deployment. Multiple deployments could share the same final MFS classifier if they shared the same common groups. A final feature selection was then run for all the image classes in the final MFS Classifiers, using Amazon High Performance Computing, to create the most appropriate feature set for the classifiers. (5) Build Deployment-Specific BFS-SVM Classifiers: Another set of SVM classifiers was also built using binary feature selection (BFS) as described in Kramer et al. (2011), where features and parameters are tuned for each class pair combination. These BFS classifiers used the same image classes as the MFS classifiers so that they were specific for the particular conditions encountered during each deployment. By using pair-wise feature selection, overall classification accuracy can be improved, and feature selection and SVM training time can be reduced. (6) Run Dual Classification: Lastly, these two sets of classifiers (MFS-SVM and BFS-SVM) were used to run a modified version of the dual classifier described for the video plankton recorder (VPR) by Hu and Davis (2006). As the name implies, a dual classifier makes use of two classifiers where each classifier makes its own separate prediction and a particle image is only labeled as a specific image class if both classifiers agree, otherwise the particle image is labeled “other”. In the case of the VPR, a SVM classifier using texture based features and a neural net using shaped-based features were used as the two classifiers. For the SIPPER data, a neural net performed poorly and it was determined that a SVM BFS using separate feature sets for each pairwise classification was the best alternative. Along with the SVM MFS classifier, the SVM BFS made up the SIPPER dual classifier. A further modification took advantage of the hierarchical naming structure of SIPPER image classes by allowing for predictions from partial matches between the two classifiers. For example if one classifier predicted that an unknown image was crustacean_copepod_calanoid and the other classifier predicted it was crustacean_copepod_oithona, it was very likely that the image was in fact a copepod. Rather than labeling this disagreement as “other” as the traditional dual classifier would have done, our modified classifier would label the image as crustacean_copepod since both classifiers agreed to that partial match. Only when the two classifiers predictions shared no common root in the SIPPER class names (e.g. one classifier predicts an unknown image is detritus_snow and the other predicts larvacean) would an image be labeled “other”. Dual classifiers are used because they improve specificity (reduce the rate of false positives) and thereby give more accurate abundance estimations, especially in regions of low relative abundance. Dual classifiers were run on each deployment using the most common image classes in that deployment as determined in Step 3. A total of 32 individual image classes were common enough to be used in at least one deployment and are listed below. Validated images from rare and uncommon classes that were removed from the classifier retained their validated labels in the classifier output.
Instruments:
SIPPER Environmental Sensors. Environmental data were collected simultaneously with the SIPPER imaging system during each deployment. Sensors included a Seabird 19Plus CTD, Seabird SBE43 oxygen sensor, and WET Labs FLNTURTD chlorophyll fluorescence and turbidity, and a transmissometer. AWET Labs CDOM sensor also was used on a few cruises. Sensors were calibrated at Seabird and WET Labs and then integrated into the SIPPER towed platform.
Error Analysis:
Samson, S., Hopkins, T., Remsen, A., Langebrake, L., Sutton, T., and J. Patten (2001). A System for High-Resolution Zooplankton Imaging. IEEE Journal of Oceanic Engineering, 26(4): 671-676. doi: 10.1109/48.972110 Remsen, A., Hopkins, T.L., S. Samson (2004). What you see is not what you catch: a comparison of concurrently collected net, Optical Plankton Counter, and Shadowed Image Particle Profiling Evaluation Recorder data from the northeast Gulf of Mexico. Deep Sea Research Part I: Oceanographic Research Papers, 51(1): 129-151. doi: 10.1016/j.dsr.2003.09.008 Kramer, K., Goldgof, D.B., Hall, L.O., A. Remsen (2011). Increased classification accuracy and speedup through pair-wise feature selection for support vector machines. IEEE Symposium on Computational Intelligence and Data Mining (CIDM), Paris, 2011, pp. 318-324. doi: 10.1109/CIDM.2011.5949457 Q. Hu and C. Davis (2006). Accurate automatic quantification of taxa-specific plankton abundance using dual classification with correction. Marine Ecology Progress Series, 306: 51-61. doi: 10.3354/meps306051
View Metadata