Development of a model for predicting mussel weight: a comparison of traditional and artificial intelligent methods

Abstract

The relationship between length and weight is non-linear. Predictive modelling using linear regression methods subjects these variables to transformation which results in models of poor predictive value. Hence, a comparative study on developing a predictive model using traditional (length-weight relationship, LWR; multiple linear regression, MLR) and artificial intelligent (artificial neural networks, ANN) methods was conducted. Specimens (n = 320) of the horse mussel Modiolus modulaides were randomly collected from October 2018 to March 2019 at the coastal area of Dumangas, Iloilo, Philippines. Shell length, shell width and shell height were used as predictor variables for total weight. A multi-layer perceptron architecture model was used and the values were determined by the ANNs model using the actual data. In addition, LWR and MLR models were generated from the same data after log-transformation. The results indicated superiority of the ANN model to predict mussel weight to traditional LWR and MLR models. The ANNs model had the highest correlation coefficient and lowest errors among the predictive models. The ANNs model generated from this study can be a good alternative to existing models and may be useful in sustainable fisheries management.

Keywords

• Artificial neural networks
• fisheries biology
• length-weight relationship
• multilayer perceptron
• predictive modelling