Prediction of Dew Point Temperature Using Tree and Kernel Based Models

Iran’s agricultural sector faces unique obstacles due to the diversity of its climatic conditions, underscoring the crucial importance of safeguarding crops against the impacts of climate change. One effective strategy to mitigate damage to the agricultural industry is the ability to accurately predict dew point temperature. The dew point refers to the temperature at which water vapor in the air condenses into liquid water or dew, given a constant air pressure. Notably, an excessively high dew point can adversely affect the performance of air conditioning systems and reduce the efficiency of coolant-based ventilation mechanisms. The formation of dew in ecosystems is influenced by a triad of key factors: radiative exchange between the Earth’s surface and the atmosphere, turbulent heat transfer, and vapor pressure. While the dew point is typically measured using a moisture meter, there also exist empirical equations that relate air temperature and humidity. However, reliable dew point forecasting often requires common meteorological parameters, such as relative humidity and precipitation, which are not consistently measured at many weather stations or may be subject to significant error. As such, regression-based estimation methods are frequently employed. Recognizing the importance of data-driven approaches in dew point estimation, this study explores the use of several predictive models, including support vector regression (SVR), the M5P tree model, the M5Rules rule-generation algorithm, Gaussian process regression (GPR), linear regression (LR), random forest (RF), and random tree (RT) models. These models were applied to data collected from two stations in Gorgan and Shahrekord, Iran, to estimate dew point temperature.
Methodology
The Gorgan basin is geographically situated between 25°54’ east longitude and 36°50’ north latitude, while the Shahrekord basin is located between 50°49’ east longitude and 32°20’ north latitude. Topographically, Shahrekord lies in the eastern segment of the Zagros Mountain range, along the Zagros fault margin. The input parameters for this study were obtained from the Iran Meteorological Organization, covering the period from 1990 to 2021. These parameters include daily maximum temperature (T_max), daily minimum temperature (T_min), daily average temperature (T_m), sunshine hours (sshn), average wind speed (ffm), average relative humidity (RH_m), maximum relative humidity (RH_max), and minimum relative humidity (RH_min). To assess the accuracy of the input parameters and models, the dew point (DP) was extracted from the testing data and evaluated using regression and tree-based models. Additionally, eight potential scenarios were defined to estimate the daily DP.
Findings
The study findings revealed that scenarios 1, 2, and 3 exhibited the highest correlation with dew point temperature, while scenarios 4, 5, 6, 7, and 8 displayed the lowest correlation. However, upon evaluating the scenarios based on the established criteria, it was determined that scenarios 1, 2, and 3 performed less effectively than the other scenarios. Consequently, the placement of the majority of parameters (scenarios 6, 7, and 8) led to a decrease in the models’ errors. The results obtained for the Gorgan models showed that the R-value ranged from 1 to 0.952. In the Gorgan station, the highest RMSE was observed for RT-3 at 2.0307, and the lowest RMSE was for SVR-8 at 0.222. Furthermore, the best fit for SVR-8 was characterized by RMSE: 0.222, NSE: 0.999, MBE: 0.092, MAE: 0.147, WI: 1, and SI: 0.017, while the worst fit was for RT-3 with RMSE: 2.307, NSE: 0.882, MBE: 0.875, MAE: 1.745, WI: 0.971, and SI: 0.179. The results for the Shahrekord models indicated that the R-value ranged between 0.996 and 0.615. In the Shahrekord station, the highest RMSE was observed for RT-2 at 4.952, and the lowest RMSE was for SVR-7 at 0.550. Additionally, the best fit for SVR-7 was characterized by RMSE: 0.550, NSE: 0.989, MBE: 0.374, MAE: 0.346, WI: 0.997, and SI: -0.15, while the worst fit was for RT-2 with RMSE: 4.957, NSE: 0.131, MBE: 1.43, MAE: 3.914, WI: 0.754, and SI: -1.352.
Conclusion
In this study, meteorological data were fitted using various modeling techniques, including Gaussian Process Regression (GPR), Linear Regression (LR), M5P, M5Rules, Random Forest (RF), Regression Tree (RT), and Support Vector Regression (SVR), to obtain the dew point temperature at the Gorgan and Shahrekord weather stations. The performance of these models was then evaluated under different scenarios, and the best-performing models were selected for estimating the dew point temperature. The results indicate that the estimation accuracy of the models using a single input parameter, such as minimum temperature (T_min), was lower compared to the other models in both the Shahrekord and Gorgan stations. The SVR model with scenario 8 and the M5P model with scenario 7 demonstrated the best performance in estimating the dew point temperature for the Gorgan and Shahrekord stations, respectively. Furthermore, the comparison of the selected models revealed that the SVR model had the highest accuracy among the models evaluated. For the Gorgan station, the models were ranked from high to low accuracy as follows: SVR, M5P, M5Rules, GPR, RF, and LR. For the Shahrekord station, the ranking was M5P, M5Rules, GPR, RF, RT, and LR. Additionally, the comparison of the tree-based and regression-based models showed that the regression models, such as SVR and LR, had higher accuracy in estimating the dew point temperature compared to the tree-based models, such as M5P, M5Rules, and RT.