Using artificial intelligence for CyanoHAB niche modeling: discovery and visualization of Microcystis–environmental associations within western Lake Erie

Abstract

Cyanobacterial harmful algal blooms (CyanoHABs), mainly composed of the genus Microcystis, occur frequently throughout the Laurentian Great Lakes. We used artificial neural networks (ANNs) involving 31 hydrological and meteorological predictors to model total phytoplankton (as chlorophyll a) and Microcystis biomass from 2009 to 2011 in western Lake Erie. Continuous ANNs provided modeled-measured correspondences (and modeling efficiencies) ranging from 0.87 to 0.97 (0.75 to 0.94) and 0.71 to 0.90 (0.45 to 0.88) for training-cross-validation and test data subsets of chlorophyll a concentrations and Microcystis biovolumes, respectively. Classification ANNs correctly assigned up to 94% of instances for Microcystis presenceabsence. The influences of select predictors on phytoplankton and CyanoHAB niches were visualized using biplots and threedimensional response surfaces. These then were used to generate mathematical expressions for the relationships between modeled CyanoHAB outcomes and the direct and interactive influences of environmental factors. Based on identified conditions (ϳ40 to 50 g total phosphorus (TP)·L −1 , 22 to 26°C, and prolonged wind speeds less than ϳ19 km·h −1 ) underlying the likelihood of occurrence and accumulation of phytoplankton and Microcystis, a "target" concentration of 30 g TP·L −1 appears appropriate for alleviating blooms. ANNs generated robust ecological niche models for Microcystis, providing a predictive framework for quantitative visualization of nonlinear CyanoHAB-environmental interactions. Résumé: Les fleurs d'eau de cyanobactéries nuisibles (CyanoHAB), constituées principalement du genre Microcystis, sont fréquentes dans tous les Grands Lacs laurentiens. Nous avons utilisé des réseaux neuronaux artificiels (« ANNs ») incluant 31 prédicteurs hydrologiques et météorologiques pour modéliser la biomasse de phytoplancton totale (chlorophylle a) et de Microcystis, de 2009 à 2011, dans la partie occidentale du lac Érié. Des ANNs continus ont fourni des correspondances modèle-mesures (et des valeurs d'efficacité de la modélisation) allant de 0,87 à 0,97 (0,75 à 0,94) et de 0,71 à 0,90 (0,45 à 0,88) pour des sous-ensembles de données d'entraînement-de contrevalidation et d'essai des concentrations de la chlorophylle a et des biovolumes de Microcystis, respectivement. Les ANNs de classification ont affecté correctement jusqu'à 94 % des cas d'absence ou de présence de Microcystis. Les influences de prédicteurs sélectionnés sur les niches de phytoplancton et de CyanoHAB ont été visualisées à l'aide de diagrammes de double projection et de surfaces de réponse tridimensionnelles, qui ont ensuite été utilisés pour produire des expressions mathématiques pour les relations entre les CyanoHAB modélisées et les influences directes et interactives de facteurs environnementaux. À la lumière des conditions cernées (ϳ40 à 50 g phosphore total (PT)·L −1 , de 22 à 26°C et des vitesses du vent soutenues inférieures à ϳ19 km·h −1 ) sous-tendant la probabilité de présence et d'accumulation de phytoplancto...

Tracking cyanobacteria blooms: Do different monitoring approaches tell the same story?

“…suggesting that higher critical meteorological thresholds might need to be crossed for surface scum to form. Especially high surface water temperatures develop during prolonged low-mixing, calm summer conditions, which are known to enhance scum formation (Ibelings et al, 2003; series (Millie et al, 2014) and more generally with observations of Microcystis growth rates reaching maxima at temperatures ≥25 °C (Butterwick et al, 2005;Reynolds, 2006).…”

“…For each meteorological and hydrological predictor (wind velocity, wind stress, irradiance, temperature, and river flow), three time-lagged variables were included as candidate predictors in the model. Specifically, we calculated average values for the 2 days, 8 days, and 30 days preceding each sampling date, to capture potential effects of each variable at different time scales leading up to each bloom observation (Millie et al, 2014) while minimizing collinearity among time-lagged variables and preventing model over-fitting. For example, wind forcing is expected to exert both a short-term (hourly/daily) impact on bloom dynamics by modulating mixing of the water column as well as a longer-term effect by driving lake circulation, currents, and subsequently water/nutrient residence time (Michalak et al, 2013;Zhou et al, 2015).…”

“…Microcystis spp. in recent years (Bridgeman et al, 2013;Millie et al, 2014), the contribution of eukaryotic phytoplankton to total chlorophyll can be substantial especially in the beginning and the end of the CHAB season (Bridgeman et al, 2012;Millie et al, 2009), thereby leading to higher chlorophyll values estimated by MO TOT and SW TOT relative to NCCOS. Indeed, when comparing the three satellite products in terms of bloom areal extent only (not shown), without quantifying chlorophyll content, some of the observed discrepancies are mitigated (see Shuchman, 2014).…”

Microstructure metrics for quantitative assessment of particle size and dispersion: Application to metal-matrix composites

“…Structural equation method is more appropriate for cases where there are many input parameters that have interaction and causality relation between these parameters is not totally known [48]. Artificial neural networks (ANNs), while being very useful for modeling complicated systems, are considered a black-box method due to their lack of ability in explaining the relationship between parameters of a system [49,50]. Also, training ANNs is computationally expensive, and it may be necessary to use complicated algorithms to obtain desirable results [50,51].…”

A State-Based Sensitivity Analysis for Distinguishing the Global Importance of Predictor Variables in Artificial Neural Networks

“…Model linearity and independence between/among multiple input variables often are not valid assumptions; as a consequence, altering the value of one predictor variable will affect (the values of) other inputs and, in turn, impact the predictive power (and uncertainty) of a response variable. For example, assume that the meteorological and hydrological variables, ambient (air) and water temperatures, respectively, are used as predictors within an ANN attempting to model a water-quality response (e.g., water clarity arising from biotic abundance [16,17]). Because ambient and water temperatures typically would be highly correlated (yet not perfectly related due to distinct physical properties of water and air), analyzing the network's predictive uncertainty in regard to ambient temperature alone, while keeping the value of water temperature static, is neither appropriate nor logical.…”