Performance evaluation of random forest and boosted tree in rainfall-runoff process modeling for sub-basins of Lake Urmia

Abstract

This study aimed to develop rainfall-runoff (P-Q) modeling using machine learning models in the sub-basins of Lake Urmia, Iran. In this research, chronological records of hydrological parameters and meteorological inputs at a regional scale were analyzed using Random Forest (RF) and Boosted Tree (BT) heuristic methods. This study compared the performance of these two models for the Urmia Basin over the period from 1976 to 2019. The results showed that the RF model provided better estimates in Akhula, Daryan, and Ghermez Gol stations in the eastern sub-basin and Miandoab, Pole Ozbak, Abajalu Sofla, Nezam Abad, and Pole Bahramlu stations in the western sub-basin. In contrast, the BT model performed better at Pole Senikh, Shishvan, Gheshlagh Amir, Shirin Kandi, and Khormazard stations in the eastern sub-basin and Babarud, Keshtiban, and Yalghoz Aghaj stations in the western sub-basin. Additionally, the time series analysis showed changes in yearly rainfall frequency and a decreasing trend in flow discharge in most years. These findings highlight a significant reduction in inflow to Lake Urmia over the past 43 years, with a particularly sharp decline in recent years.