Prediction of riverflow using bivariate extreme value distribution with composite likelihood approach

Geographically, there is not much difference between the annual maximum streamflow between neighbouring sites. Consequently, inter-site dependency should be taken into account for the analysis.The objective of this study is to model the extreme river flow considering the dependence between sites. This study uses an annual maximum river flow (m3s-l) data at five (5) selected stations in Sabah with a sample size of 29 observed from 1988 to 2016. We applied the multivariate extreme value distributions to capture the dependencies of riverflows at neighbouring sites. We build a joint model of the bivariate logistic extreme distribution of the full likelihood function based on the pairwise joint estimation using the composite likelihood approach. The result shows that the bivariate logistics distribution is an appropriate model for the extreme river flow. We also forecast that the return level of the maximum river flow is increasing consistently but is not expected to exceed the maximum level once every 100 years.