Bayesian model for semi-automated zooplankton classification with predictive confidence and rapid category aggregation

Abstract

Zooplankton play a critical role in aquatic ecosystems and are commonly used as bioindicators to assess anthropogenic and climate impacts. Nevertheless, traditional microscopebased identification of zooplankton is inefficient. To overcome the low efficiency, computer-based methods have been developed. Yet, the performance of automated classification remains unsatisfactory because of the low accuracy of recognition. Here we propose a novel framework for automated plankton classification based on a naïve Bayesian classifier (NBC). We take advantage of the posterior probability of NBC to facilitate category aggregation and to single out objects of low predictive confidence for manual re-classifying in order to achieve a high level of final accuracy. This method was applied to East China Sea zooplankton samples with 154 289 objects, and the Bayesian automated zooplankton classification model showed a reasonable overall accuracy of 0.69 in unbalanced and 0.68 in balanced training for 25 planktonic and 1 aggregated non-planktonic categories. More importantly, after manually checking 17 to 38% of the objects of low confidence (depending on how one defines 'low confidence'), the final accuracy increased to 0.85−0.95 in the unbalanced training case, and after checking 18 to 42% of the low-confidence objects in the balanced training case, the final accuracy increased to 0.84−0.95. Our semi-automated approach is significantly more accurate than automated classifiers in recognizing rare categories, thereby facilitating ecological applications by improving the estimates of taxa richness and diversity. Our approach can make up for the deficiencies in current automated zooplankton classifiers and facilitates an efficient semi-automated zooplankton classification, which may have a broad application in environmental monitoring and ecological research. KEY WORDS: Automated classification · Naïve Bayesian classifier · Predictive confidence · Rapid category aggregation · Zooplankton community · ZooScan Resale or republication not permitted without written consent of the publisherMar Ecol Prog Ser 441: [185][186][187][188][189][190][191][192][193][194][195][196] 2011 the traditional approach is labor-intensive and timeconsuming, limiting our ability to analyze zooplankton samples and understand processes controlling aquatic ecosystem dynamics (Suthers & Rissik 2008, Gorsky et al. 2010. Therefore, an important issue in aquatic ecology concerns how to improve the efficiency in plankton analysis and make the data more comparable at different spatial and temporal scales (Wiebe & Benfield 2003, Perry et al. 2004.Automated or semi-automated computer-aided systems can improve the efficiency in sample analyses and may even deliver more accurate and consistent results than human taxonomists in some cases (Culverhouse et al. 2003, Benfield et al. 2007, MacLeod et al. 2010. Recently, a large amount of effort has been invested in developing automated plankton identification systems (Ortner et al. 1979, Balfoort et al. 1992, Boddy e...

Automated otolith image classification with multiple views: an evaluation onSciaenidae

“…Model aggregation has been suggested for classification if the training set is of small sample size because models trained from a small data set generally give higher variances (Beleites & Salzer, ). Conversely, a non‐aggregated model is suggested for large data sets to increase computational speed (Ye et al , ). In rejecting uncertain predictions, this study shows that the aggregated model performs better than the non‐aggregated model, in agreement with the results of Beleites & Salzer ().…”

Automated otolith image classification with multiple views: an evaluation onSciaenidae

“…Model aggregation has been suggested for classification if the training set is of small sample size because models trained from a small data set generally give higher variances (Beleites & Salzer, ). Conversely, a non‐aggregated model is suggested for large data sets to increase computational speed (Ye et al , ). In rejecting uncertain predictions, this study shows that the aggregated model performs better than the non‐aggregated model, in agreement with the results of Beleites & Salzer ().…”

Mesozooplankton size structure in response to environmental conditions in the East China Sea: How much does size spectra theory fit empirical data of a dynamic coastal area?

“…With continued advances in automated imaging systems (e.g. Beaulieu et al, 1999;Benfield et al, 2007;Ye et al, 2011), information on plankton size structure may be collected with relatively high efficiency.…”

Importance of prey size on investigating prey availability of larval fishes

“…We measured the body width of each individual under the microscope and omitted individuals bigger than 200 μm. For each mesozooplankton sample, we analyzed approximately 1000 individuals (with appropriate subsampling using Folsom splitter) using the ZooScan integrated system and followed the semi-automatic classification method [ 24 ] to identify the taxa. To facilitate comparisons between small-size zooplankton and mesozooplankton, we considered only copepod taxa in mesozooplankton samples (hereafter we used the term “mesozooplankton” and “copepods” interchangeably).…”

Automated otolith image classification with multiple views: an evaluation onSciaenidae

“…Model aggregation has been suggested for classification if the training set is of small sample size because models trained from a small data set generally give higher variances (Beleites & Salzer, ). Conversely, a non‐aggregated model is suggested for large data sets to increase computational speed (Ye et al , ). In rejecting uncertain predictions, this study shows that the aggregated model performs better than the non‐aggregated model, in agreement with the results of Beleites & Salzer ().…”

Mesozooplankton size structure in response to environmental conditions in the East China Sea: How much does size spectra theory fit empirical data of a dynamic coastal area?

“…With continued advances in automated imaging systems (e.g. Beaulieu et al, 1999;Benfield et al, 2007;Ye et al, 2011), information on plankton size structure may be collected with relatively high efficiency.…”

Importance of prey size on investigating prey availability of larval fishes

“…We measured the body width of each individual under the microscope and omitted individuals bigger than 200 μm. For each mesozooplankton sample, we analyzed approximately 1000 individuals (with appropriate subsampling using Folsom splitter) using the ZooScan integrated system and followed the semi-automatic classification method [ 24 ] to identify the taxa. To facilitate comparisons between small-size zooplankton and mesozooplankton, we considered only copepod taxa in mesozooplankton samples (hereafter we used the term “mesozooplankton” and “copepods” interchangeably).…”

Automated otolith image classification with multiple views: an evaluation onSciaenidae

“…Model aggregation has been suggested for classification if the training set is of small sample size because models trained from a small data set generally give higher variances (Beleites & Salzer, ). Conversely, a non‐aggregated model is suggested for large data sets to increase computational speed (Ye et al , ). In rejecting uncertain predictions, this study shows that the aggregated model performs better than the non‐aggregated model, in agreement with the results of Beleites & Salzer ().…”

Mesozooplankton size structure in response to environmental conditions in the East China Sea: How much does size spectra theory fit empirical data of a dynamic coastal area?

“…With continued advances in automated imaging systems (e.g. Beaulieu et al, 1999;Benfield et al, 2007;Ye et al, 2011), information on plankton size structure may be collected with relatively high efficiency.…”

Importance of prey size on investigating prey availability of larval fishes

“…We measured the body width of each individual under the microscope and omitted individuals bigger than 200 μm. For each mesozooplankton sample, we analyzed approximately 1000 individuals (with appropriate subsampling using Folsom splitter) using the ZooScan integrated system and followed the semi-automatic classification method [ 24 ] to identify the taxa. To facilitate comparisons between small-size zooplankton and mesozooplankton, we considered only copepod taxa in mesozooplankton samples (hereafter we used the term “mesozooplankton” and “copepods” interchangeably).…”