Enhanced long-term and snow-based streamflow forecasting by artificial intelligent methods using satellite imagery and seasonal information

This paper investigates the simultaneous use of in-situ hydrologic measurements, such as discharge, temperature, precipitation and snowfall information derived from satellite imagery in combination with two different AI methods, namely, Adaptive Neuro Fuzzy Inference System (ANFIS) and Artificial Neural Network (ANN), for developing enhanced long-term streamflow forecasting models. To enhance the reliability of the proposed models’ outputs, the number of input data used for their training and testing is increased using a sub-basin method. Furthermore, to accelerate the training process and achieve more accurate handling of seasonal changes, a parameter representing seasonal variations is introduced. A regionalization approach is also proposed to overcome the problem of deficiency and inappropriate distribution of hydro-meteorological stations in poor data regions. To obtain the most principal input variable set to be used in developing the models, a gradual model development approach is proposed and followed. In summary 12 streamflow forecasting models based on ANFIS, ANN and using three different model structures and two forecast time intervals (monthly, seasonal) are developed. The models are applied to data collected from the mountainous Talezang basin located in the southwestern Iran, which consists of 14 years of monthly measurements including streamflow, precipitation, temperature, and snow water equivalent (SWE) records and snow cover area obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS). The results indicate that the use of the sub-basin approach significantly improves both models’ performances, as indicated by the improvement of the correlation coefficient index (R) from 0.44, to 0.77 in the testing phase. Moreover, it is deduced that including additional input parameters in the model structure, as well as using seasonal information and satellite data, has a great impact on the model’s performance and accuracy, evident by the reduction of the scatter index (SI) by 35% on average. Comparing the long-term flow forecasts of both models showed that ANFIS is superior to ANN. It is concluded that the ANFIS method, developed based on data from the proposed sub-basin method and seasonal parameter, is capable of providing high quality streamflow forecasts, particularly for rivers and streams located in data poor regions.