تخمین تبخیروتعرق مرجع روزانه به کمک مدل یادگیری عمیق، جنگل تصادفی و درخت تصمیم (مطالعه موردی: دشت سیستان)

In water scarcity conditions, water resources management and accurate use of available water for sustainable production of required products are the essential approches in arid and semi-arid areas. To achieve this management, the evapotranspiration plays a very important role that is one of the main components of the hydrological cycle and successes of water projects. In the other words, sustainable production depends on the correct estimation of the evapotranspiration. Determination of the water requirement of crops and gardens will increase the irrigation efficiency and improve water management in the field. In order to estimate the evapotranspiration, it is necessary to accurately calculate reference evapotranspiration. So far, many experimental methods have been proposed to estimate the reference evapotranspiration. The FAO Penman Montieth method is one of the most precise methods to determine of this parameter. But, the limitation in using this method is calibrating for areas that determined to be studied. Over the last few years, new tools and innovative methods have been widely used to estimate the reference evapotranspiration. Therefore, it is necessary to check the correctness of these models and methods for different regions.
Sistan is one of the fertile regions of Iran, which, unfortunately, over the years, the phenomenon of drought has been swept away the active farming in the area and destroyed a boom in agricultural production in this region. Nevertheless, the main occupation of rural people in this region is still agriculture. Any neglect on the water category in this area, can be a serious damage to it. Winds for more than 120 days are the main climate features of this region that has extremely been increase the evaporation. Therefore, it is necessary to extract the highest efficiency from the limited water in the area. So, in this research, the effective climatic parameters to calculate of daily reference evapotranspiration in Sistan plain were identified and analyzed, using different scenarios of the combination of meteorological parameters. These scenarios were used as inputs models of decision trees, random forest and deep learning. The best scenario was extracted by results of models. Then, the accuracy of the results was compared with the experimental method of FAO Penman Montieth as the basis method. Scenarios were M1 - M25. The data used in these scenarios were: max, min and average of two temperature and humidity parameters, rainfall, sunshine, wind speed, and pan evaporation. The data were provided from the Zabul meteorological stations for 2009-2018. For validation, 80% of the data have been use for training and the remaining 20% for testing the models.
Results of these models were compared with the estimated results by the FAO Penman Montieth model and shown, in decision tree model, M10 scenario with the lowest error rate (RMSE=1.003) and the highest correlation coefficient ( R=0.983 ) was the best scenario. In random forest model, M10 scenario with the lowest error rate (RMSE=1.003), MAE=0.8 and the highest correlation coefficient (R=0.983) was the best scenario. In deep learning model, M5 scenario with the lowest error rate (RMSE=0.517), MAE=0.399 and the highest correlation coefficient (R=0.996) was the best scenario. In results of the supporting machine model and selecting the best model, by evaluating the comparison of decision tree and deep learning models, it was shown that a specific pattern can be introduced for Sistan region with proper accuracy for estimating daily reference evapotranspiration and should introduce a superior scenario for each model. In order to deliberation the importance of various meteorological parameters in the results of the mention models, among all the parameters which have used in decision tree model, average temperature, wind speed, maximum temperature and minimum temperature, respectively, had the most importance. In random forest model, maximum temperature, wind speed, average temperature and pan evaporation had the most importance, respectively. Finally in deep learning model, average temperature, maximum temperature, wind speed and minimum humidity, respectively, had the most effect on calculation of daily reference evapotranspiration in case study region. As a general conclusion it can be expressed that deep learning model is the best model among selected models and M5 scenario in deep learning with the highest correlation coefficient (R=0.996) and the lowest error (RMSE=0.517), had a good accuracy to modeling of daily reference evapotranspiration.