Increased sampled volume improves Microcystis aeruginosa complex (MAC) colonies detection and prediction using Random Forests. (August 2017)

• Record Type: Journal Article
• Authors: Segura, A.M.; Piccini, C.; Nogueira, L.; Alcántara, I.; Calliari, D.; Kruk, C.
• Is Part Of: Ecological indicators. Volume 79(2017)
• Page Start: 347
• Page End: 354
• Publication Date: 2017-08
• Subjects: Cyano-Hab -- Predictability -- Machine learning -- South America -- Microcystis aeruginosa complex; Environmental monitoring -- Periodicals; Environmental management -- Periodicals; Environmental impact analysis -- Periodicals; Environmental risk assessment -- Periodicals; Sustainable development -- Periodicals; 333.71405
• DOI: 10.1016/j.ecolind.2017.04.047
• Languages: English
• ISSNs: 1470-160X

Abstract

Abstract: Limited predictability of cyanobacteria and algal harmful blooms (CyanoHABs) impairs the development of adequate water management programs. The Microcystis aeruginosa complex (MAC) is ubiquitous worldwide. Their large colony size and relatively low numerical abundance imply that MAC abundance and presence are usually underestimated in traditional phytoplankton quantifications, which are based on samples of small volume. The objective of this work was twofold: (a) evaluate four sampling strategies of increasing sampling size to detect MAC organisms and (b) asses the predictability of MAC presence using easy-to-measure environmental variables. Sampling strategies were (I) 5–25 mL sedimented water samples inspected under inverted light microscope; (II) 20 L of water samples inspected by on-board naked-eye; (III) samples collected by towing a 25 μm pore size net inspected under light microscope; (IV) naked-eye inspection of 1000–7000 L concentrated water samples collected using a 115 μm-pore plankton net. We evaluated these objectives in a large environmental gradient (800 km) from freshwater to marine water (salinity range = 0–33) covering a wide range of temperatures (10–33 °C), underwater turbidity (0–158 NTU) and wind intensity (0–8 ms −1 ). Classification Random Forest models (presence/absence of MAC organisms) were constructed and evaluated for each strategy by randomly partitioning data into training (2/3) and test (1/3) sets. A systematic increase in average accuracy (from 51 to 90%) and sensitivity (from 45 to 94%) towards methods with larger sampling size was found ( i.e. I–IV). The best obtained model showed a high accuracy (90%) and sensitivity (94%) to detect MAC presence. These results suggest that the presence of MAC organisms can be accurately predicted using easy-to-measure environmental variables once sampling size is adequate. The proposed methodology demands very low costs and could be readily incorporated in most water monitoring plans to provide early warning of MAC occurrence, even when there is a low biomass of organisms.