Evaluation of three data mining methods to estimate reference evapotranspiration in Zanjan province

Introduction:

Reference evapotranspiration (ET0), a complex hydrological variable affecting crop water requirements and irrigation scheduling, is defined by a number of climatic factors that have an impact on water and energy balances. On the basis of accurate climatic data, conventional methods for calculating ET0 include a variety of empirical approaches. But there are a lots of locations where different climatic information might not be available for ET0 estimation.

The objective of this study is to evaluate different data mining methods to estimate ET0 with limited meteorological data. This study aims to answer the question: can reference evapotranspiration be estimated without reducing accuracy, regardless of the availability of all variables? In this research, the accuracy of data mining methods in estimating ET0 with respect to the plant water demand system (FAO Penman-Monteith standard method) was evaluated

 Data such as sunshine hour, air temperature, wind speed, and relative humidity from thirteen climatology stations in the Zanjan province over a ten-year period (2010-2021) were collected. The ET0 was calculated using the FAO56 Penman-Mantith method on a daily time scale (as refrence method) and the estimated values obtained by data mining methods (Artificial Neural Network (ANNs), Random Forest (RF) and Support Vector Machine (SVM)) were evaluated. The data from each station were divided into two sets: training (two-thirds of the data) and testing (one-third of the data) in order to calibrate and validate the proposed methods. Finally, based on NRMSE, RMSE, MBE, and EF criteria, the generalizability of the aforementioned methods for estimating ET0 was examined.

 According to the results, ANNs performed better than SVM and RF methods. The mean values of, RMSE, EF and NRMSE criteria for the ANNs method in the training and testing steps were 0.49, 0.94 and 0.14, respectively. The mean values of these criteria for RF method in the training step were 0.49, 0.94 and 0.14 and in the testing step was 0.52, 0.94 and 0.15, respectively. The mean values of these criteria for the SVM method for both (training and testing) steps were 0.52, 0.94 and 0.15, respectively.The average air temperature is the most significant and effective parameter to estimate ET0, according to more than 92 percent (12 stations) of the results obtained from two ANNs and RF methods. The sunshine hours is the second-most crucial and useful input in estimating ET0, according to more than 84 percent (11 stations) of the results. As a result, using four meteorological variables such as average air temperature, average relative humidity, wind speed, and sunshine hours as input, excellent performance can be achieved. The NRMSE values obtained from ET0 estimation did not exhibit regular variations with the average values of parameters (temperature, humidity, wind speed, sunshine hours, slope percentage).

It was found that the average air temperature was the most crucial and useful parameter as a result of the sensitivity analysis of the ANNs method and the Predictor Importance of the RF method. According to the current study, Pari and Zanjan stations outperformed than the other stations in Zanjan province, probably due to their plainer conditions. The results of the current study will help to estimate ET0 for semi-arid climates where ET0 is critical for agricultural water resource management.