Determination of sensitive variables regardless of hydrological alteration in artificial neural network model of chlorophyll a: Case study of Nakdong River

• Published in: Ecological modelling Vol. 398; pp. 67 - 76
• Main Authors: Kim, Hyo Gyeom, Hong, Sungwon, Jeong, Kwang-Seuk, Kim, Dong-Kyun, Joo, Gea-Jae
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 24.04.2019
• Subjects: Algal blooms; Artificial neural network; Chlorophyll a; Regulated river; River modification; Sensitivity analysis; Algal blooms; Regulated river; River modification; Sensitivity analysis; Artificial neural network; Chlorophyll a
• ISSN: 0304-3800; 1872-7026
• DOI: 10.1016/j.ecolmodel.2019.02.003

•Artificial neural networks trained by the dataset in the pre-hydrological modification period can predict the chlorophyll a in the post-period.•Data analysis by the network modelling indicated that water quality variables are dominant forcing factors shaping algal dynamics.•The model performance in the post-period allowed to infer new patterns caused by river regulation.•Profile sensitivity analyses could elucidate a causal relationship among each descriptor and the output.•The reduction of nitrogen would be effective than adaptive management of dam discharge to control algal blooms. The Nakdong River has suffered from hydrological alterations in the river channel and riverine area during the Four Major Rivers Restoration Project (FMRRP). As these anthropogenic modifications have induced intensive algal blooms, the prediction of algal abundances has become an important issue for securing a source of drinking water and ecosystem stability. This study aimed to assess the changed river system in terms of chlorophyll a concentrations using artificial neural network (ANN) models trained for the pre-FMRRP period and tested for the post-FMRRP period in the middle reaches of such a river-reservoir system, and identify the descriptors that consistently affect algal dynamics. A total of 19 variables representing biweekly water-quality and meteo-hydrological data over 10 years were used to develop models based on different ANN algorithms. To identify the major descriptor to the algal dynamics, sensitivity analyses were performed. The best and most feasible model incorporating five parameters (wind velocity, conductivity, alkalinity, total nitrogen, and dam discharge) based on the topology of a probabilistic neural network with a smoothing parameter of 0.028 showed satisfactory results (R = 0.752, p < 0.01). Some mismatches were found in the post-FMRRP period, which may be due to a discrete event with a newly adapted over-wintering species and different causes of the summer growth of cyanobacteria owing to the river alteration. Based on the lowest sensitivity of dam discharge and the combination results of environmental management with total nitrogen, ANN modelling indicated that short-term water quality variables are persistent factors shaping algal dynamics.