發展浮游動物影像自動化辨認系統 – 以東海生態系為例

本研究目的在於發展一套浮游動物影像自動化辨認系統。以東海的浮游動物為材料，探討此系統在這個西太平洋陸棚海域的實用性。東海的水文環境狀況複雜，從沿岸地區受到長江淡水注入的影響，直至東海陸棚外受黑潮的影響都包括在其中。在這樣複雜的水文狀況下，我想要測試水體訓練樣本的準確率會高於全域訓練樣本準確率的假設（水體專一性的影響）。在記錄了兩兩測站之間的交叉辨認率之後，發現辨認率隨著兩兩測站之間環境因子異質性的遞增而遞減；例如，同屬近岸地區的測站交互辨認率會高於以近岸測站去辨認遠岸測站。另外，交叉辨認率的矩陣結構與環境異質性的矩陣結構有顯著地相關性。這些結果確認了水體專一性的影響，因此假設成立。不符合預期的是，全域訓練樣本的正確率（平均75％）皆不比最好的水體訓練樣本來得差，我們發現這是因為大部分觀察到的正確率皆受到優勢類群的影響，顯示出水體訓練樣本在當下的分類解析度之中效能有限。但重點是，由自動化辨認所得到的群聚結構仍然可以解釋大部分的環境變異，說明了這套系統實用的可能性。

關鍵字

水體訓練樣本；全域訓練樣本；影像分析；機器學習；水文環境；浮游動物群聚結構

並列摘要

We developed an automatic classification system for mesozooplankton in the East China Sea, a region with complicated environmental variations ranging from coastal areas affected by river runoff to the shelf break influenced by the Kuroshio. Considering the large variation of water masses, we test the hypothesis that a water mass-specific training set (WTS) would perform better than a regional training set (RTS, which is randomly assembled from all stations). To test the water mass specificity, we evaluated pair-wise cross predictions for all stations. We found that cross-prediction accuracy decreased with an increase in environmental dissimilarity; for example, mutual predictions between coastal stations performed better than those using coastal stations to predict Kuroshio stations. Furthermore, the cross-prediction matrix is significantly correlated with the similarity matrix derived from environmental variables. These results suggest clear water mass specificity in training sets. However surprisingly, the prediction accuracy (with an average of 75%) of RTS performs equally well as the best WTS results for each station. This is mainly due to the contribution of dominant zooplankton categories to the overall accuracy rate. This suggests the benefit of WTS is limited for the coarse taxonomic resolution (order or higher) employed here. The machine-identified species composition (albeit containing ~38.5% error) still explained a significant amount of variance associated with the environmental gradient, demonstrating the potential capability of the automatic classification system.

並列關鍵字

water mass-specific training set ； regional training set ； image analysis ； machine-learning ； hydrography ； zooplankton community

參考文獻

Beaugrand, G., P. C. Reid, F. Ibanez, J. Alistair, and M. Edwards. 2002. Reorganization of North Atlantic marine copepod biodiversity and climate. Science 296: 1692-1694.

Bell, J. L., and R. R. Hopcroft. 2008. Assessment of ZooImage as a tool for the classification of zooplankton. Journal of Plankton Research 30: 1351-1367.

Benfield, M. C. and others 2007. RAPID- Research on Automated Plankton Identification. Oceanography 20: 12-26.

Culverhouse, P. F., R. Ellis, R. G. Simpson, R. Williams, R. W. Pierce, and J. T. Turner. 1994. Automatic categorization of 5 species of Cymatocylis (Protozoa, Tintinnida) by artificial neural network. Marine Ecology Progress Series 107: 273-280.

Culverhouse, P. F. and others 1996. Automatic classification of field-collected dinoflagellates by artificial neural network. Marine Ecology Progress Series 139: 281-287.

國際替代計量

發展浮游動物影像自動化辨認系統 – 以東海生態系為例

全文下載

主題瀏覽