Evaluation of Time Series Models in Prediction of Seasonal Precipitation Based on Remote Sensing Data (Case Study: Arid and Semi-arid Climates)

Rainfall modeling is essential in water resources and agriculture management, especially in arid and semi-arid climates. Problems in generalizing precipitation from a point to a region are the reasons for alternative methods to estimate precipitation, and one of these methods is the use of remote sensing. Across the globe, research is being conducted to evaluate the accuracy of remote sensing precipitation data. Accordingly, the aim of this study is to model the prediction of TRMM Multi-satellite Precipitation Analysis (TMPA-3B43) data in arid and semi-arid climates of Iran, using the SARIMA simulation model for the period of 1998-2017 in seasonal scale.  

For this study, monthly precipitation data from TMPA 3B43-v7 were selected from two different climates. To ensure the reliability of the results, a completely randomized selection of four regions was made from different provinces and topographic conditions. The TMPA satellite data (3B43-V7) was prepared by the NASA databases during the period of 1998-2017. These data have a time-stamping time resolution and spatial resolution of 0.25 degrees, covering from 50 degrees south to 50 degrees north of latitudes. Satellite pixel data was corrected using the corresponding synoptic station data. Monthly data after correction were aggregated to seasonally data. Spring data from April, May and June, the summer data from July, August, and September, autumn data from October, November, and December, and winter data were obtained from January, February, and March. Using the autocorrelation function (ACF) and partial autocorrelation function (PACF), time series models were determined. The normality of the residuals, the independence of the residuals and the constant of the variance of the residuals was checked out. A general model was fitted to the data to check the accuracy of the selected model. The result showed that the selected model has the least error compared to other models. After calibration and evaluation of the selected models, seasonal precipitation data was predicted based on the monthly scale for the period (2018-2021) and compared with the base period (1998-2017). The evaluation was measured using some statistical indices including AIC, R, MBE, MAE, and RMSE.  

The correlation coefficient based on monthly data changes from 0.81 to 0.9 in the studied areas which confirms the high accuracy of modeling based on SARIMA technique in monthly data for seasonal prediction. Best model for corresponding pixel of Behbahan station was SARIMA(0,0,1)(1,1,1)12 and SARIMA(0,0,0)(2,1,1)4, Zarghan station was SARIMA(2,0,0)(1,1,1)12 and SARIMA(0,0,0)(2,1,2)4, Kashan station SARIMA(0,0,0)(1,1,1)12 and SARIMA(0,0,1)(2,1,1)4 and Sirjan station SARIMA(1,0,1)(1,1,1)12 and SARIMA(0,0,0)(2,1,1)4 for monthly and seasonal modeling respectively. The mean bias error in the monthly and seasonal periods of the Kashan station corresponding pixel (0.2 and 0.5 respectively) and the monthly period of the Sirjan station corresponding pixel (-0.3) was obtained. Seasonal modeling based on monthly data was compared with seasonal prediction data. In the autumn, the same as the spring season, seasonal precipitation increases in three regions and decreases in one station. In the summer, it was observed relative stability in seasonal precipitation at all points. In winter, unlike the spring and autumn seasons, seasonal precipitation was decreased in three areas.  

The results showed that if the predicted monthly data turned into seasonal data, they would be more accurate for seasonal forecasting. Generally, based on annual precipitation, the studied areas will have a relative precipitation reduction of 1.8 mm per year. Regarding the average annual rainfall in the four studied areas, the average annual precipitation is estimated to be decreased by about 1 percent in the study period. Separately, annual precipitation changes are -1.3%, 1.1%, 1%, and 1.8% in Kashan, Sirjan, Behbahan, and Zarghan areas respectively.