The Estimation Of Reference Crop Evapotraspiration Based On Data Mining And Artificial Intelligence

In China, not only the total available water resources is shortage, but also exists great geographical variety with large discrepancy between north and south and the unbalance of distribution in seasons. Also agricultural water consumption is heavy and the using efficiency of agricultural water is lower. The estimation of the reference crop evapotraspiration with high accuracy can help increase agricultural water use efficiency and it is absolutely necessary and significant to plan and manage the irrigation system and design irrigation schedule and make water resources planning and so on. Because the direct method to measure the crop evapotraspiration is highly complicated and maintaining the measurement tool is hugely expensive. Usually, the actual crop evapotraspiration can be calculated by multiplying crop coefficient and reference crop evapotraspiration, which is a convenient and effective. Among all the equations for calculating reference crop evapotraspiration using meteorological data, the Penman-Monteith recommended by FAO has been utilized globally due to the fact that this equation can be applied widely with high accuracy and without regional restriction. However, this equation need relatively much weather data and it cannot be used in some areas which lack completed and enough weather data, especially in remote and undeveloped places. Therefore, it is imperative to find a relatively simple method to estimate reference crop evapotraspiration using less weather data. In this research, fuzzy C-means algorithm, genetic algorithm, Gamma test algorithm, extreme learning machine neural network and other data mining and artificial intelligence techniques and models. The main results and conclusions of this research were listed following:(1) This study evaluated seven models (Hargreaves equation, Hargreaves-M1 equation, Hargreaves-M2 equation, Irmark-Allen equation, Priestley-Taylor equation, Makkink equation and Turc equation) commonly used to estimate daily reference crop evapotranspiration (ET0) values and was to determine the model used to estimate ETo with small data requirements and high accuracy for eighteen synoptic stations in Zhejiang province. The results showed that the Priestley-Taylor model was the best suited model in estimating ETo, and Hargreaves equation, Irmark-Allen equation and Makkink equation turned out to be suited under some areas.(2) If a trained model can be applied in some areas under similar climate, it could give good performance in these regions and improved the generalization ability of the models. This research assume that the weather station belonged to same cluster which was determined by the fuzzy C-means algorithm (FCM) according to the longitude, latitude and altitude and confirm this hypothesis by actual modelling. Clustering result of FCM is sensitive to the initial center, and often cannot achieve the result of the global optimal because the existence of the local minimum points of the objective function. So an improved FCM clustering algorithm combined with genetic algorithm was proposed to solve the problem that converges at minimal value. The results showed that the improved FCM clustering algorithm can achieve the greater stability, while the running time of algorithm was longer than the normal FCM clustering algorithm. In this paper, the clustering features of weather station were recognized by using the improved FCM clustering analysis method. As a result, the eighteen meteorological stations were divided into three categories, which can provide the necessary information for modeling. (3) Before establishing models to estimate ETo, it is always difficult to choose the proper input variables to build the model. Gamma test is a data analysis routine that can estimate the best Mean Squared Error that can be achieved by any continuous or smooth data model constructed using the data This can help choose the proper input variables. In this paper, the Levenberg-Marquardt neural network using different variable combination to train was used to verify the feasibility of Gamma test. The results illustrated that Gamma test can choose the best variable combination for building the model. Then Gamma test was to analyze the training data set in Hangzhou station, Jinhua station and Ruian station and choose the suited inputs for modeling.(4) Extreme Learning Machine (ELM), as a fast machine learning algorithm, it has the high learning efficiency, conceptual simplicity, and the good generalization capability. In this research, the ELM neural network was utilized to estimate ETo. The training strategy was designed according to the result of clustering analysis and Gamma test, which can provide necessary information. In a cluster, there are more than one weather station and chose one from them as the training station (the training data set derived from this station). The observation data during the period from 2007 to 2010 at the chosen station used as training data set and validating data set included the weather data from 2011 to 2013 from all the stations in the same cluster. After trained and validated the ELM neural network, the results showed that all the trained model with different variables achieved a satisfied performance, while the models performed well when they are validated in different stations except Dongtou station, Yuhuan station, Shengsi station, Dachendao station and Shipu station. Therefore, it is feasible to estimate ETo using the ELM neural network. Besides, the information from results of clustering analysis and Gamma test did help increase the generalization capability.