Gap infilling of daily streamflow data using a machine learning algorithm (MissForest) for impact assessment of human activities

Abstract

Machine learning algorithm has been increasingly used to fill missing daily streamflow data from neighboring gauges in data-scarce regions. However, how human activities, especially reservoir construction, may affect the performance of these gap-filling algorithms has not been explicitly assessed. This study applied the MissForest algorithm (MF) to infill daily streamflow gaps at 58 selected stations in the north area of Tianshan Mountains, an arid region in northwest China. The stations without reservoir impacts (i.e., with natural flow regimes) and with reservoir impacts (i.e., with high impacts of human activities) have been separately used to implement the infilling such that the impact of human activities on the performance of MF can be analyzed. Results show that MF using the station without reservoir impacts performed very well in filling daily streamflow gaps. Its performance had been insignificantly influenced by an increasing amount of missing data, different numbers of stations, and different lengths of observed data at each station. Without human activities, MF performance obviously improved with an increasing of 2.2%-3.2% in mean of R2 and a reducing of 21%-40% in mean of SMAPE. However, the performance of MF exhibited a noticeable degradation as a result of human activities that mean of R2 reduced by 1.1%-3.4% and mean of SMAPE increased by −19.5%-30.6%. This could be attributed to the fact that the MF algorithm is unable to fully capture the relevant information regarding human activities. The reconstructed daily streamflow series showed little flow regime change at the stations without reservoir impacts, but an advanced peak flow from July to June and a decrease of peak flow by 13% at the stations with reservoir impacts during the period of 2006–2011 compared with the baseline period of 1964–1989. It is recommended to first fill in the data gaps at stations without human activity, and then fill in the data gaps at stations influenced by human activities. We recommend exercising caution when using the MF method in rivers where human activities are significant or where low flow and intermittent flow events occur frequently.