[4]
52 … … … … … 5.
Google Scholar
[5]
47 Note: due to the limited length, list only part of data Model Building. 1. Network layer. Network fitting precision has close relationship with the number of network layer and each layer neurons. Neural network usually adopt three layer structures, and the fitting accuracy can increase by adding the number of each layer neurons. In this model, three layer structures was used, and tansig function was selected as the transfer function of the input and hidden layer, that is to say ƒ. The transfer function of the output layer is purelin function, namely.
DOI: 10.7717/peerj-cs.597/table-2
Google Scholar
[2]
Number of neurons. Input and output layers are respectively related to the input variables and output variables, so the number of input and output neurons is determined by the number of input and output variables. The input variables are crop physiological capacitance and physiological resistance, and the output variable is crop evapotranspiration in the model. Hidden layers include 11 neurons, and the network structure is three layer structures, namely 3-11-1.
DOI: 10.7717/peerj-cs.856/table-6
Google Scholar
[3]
Model Training. Firstly the input/output sample data are defined, and the parameters of the neural network are set, then the network was Trained and simulated until the objective error was reached. At the same time, the rationality of the network was verified by a set of experiment data. In the process of building model, Newff function of MATLAB neural network toolbox was used and initialized. The parameters were set as follows: training steps is 800, target error is 0. 001, learning rate is 0. 05 adopted, and momentum is 0. 95. When training process reached the target error, the training process will automatically stop.
DOI: 10.1016/0956-7135(93)90132-8
Google Scholar
[4]
Simulation result and analysis. Due to using L-M optimization algorithm (Trainlm), network can get faster convergence speed and very small training error. Below is the training results of the program, and figure 2 is the corresponding error change curve. From the result, when the network was trained to the 232 step, the network performance was achieved. The training error curve of the model is shown in figure 2. Fig. 2 Curve of model error From the results of the simulation, the neural network was trained by 60 group sample data, and prediction error and precision can reach the predetermined requirement.
DOI: 10.7554/elife.31557.029
Google Scholar
[5]
Experimental Verification. The BP neural network can overcome the limitation of the sensor network and linear neural network and realize the mapping of any linear or non-linear function. However, the BP neural network is based on the gradient descent error back propagation algorithm for learning, so network training speed is normally slow, and very easy to get into the local minimum point. Improvement optimization fast algorithm was adopted in the present model, can solve some problems. Training process was repeatedly implemented in the model design, and a set of data were usually taken to testify the design validity and improve network generalization ability. In this neural network system, a group of experimental data were taken (not the training sample data) to predict crop water requirement, and the rationality of the model was analyzed by comparisons. Contrasting of model testing result and Experiment values is shown in table 2. Table 2 Contrasting of model testing result and Experiment values serial number experimental data Physiological Physiological soil Resistance capacitance moisture MW RF /% predicted value /mm·d-1 experimental value /mm·d-1 relative error /% 1.
Google Scholar
[4]
70 From the training results of random 5 groups of experimental data, this optimization BP neural network has very good prediction performance, and can completely meet the need of irrigation management of crop growth, and especially it very easy to operate. Summary The results of this paper showed that crop water requirement could be predicted by crop physiology the electrical characteristics and soil humidity. The model has more precise and reasonable inference process. Due to using BP neural network technology, this will overcome the complexity and difficulty of the traditional calculation method was overcome. We can get crop water requirement as long as a set of data were input into the model by. sim(net, p) function. At the same time, this neural network model was trained by long-term experiment data, thus the big error defects was overcame. If the model was applied to irrigation system, the irrigation controller will irrigate crop at regular intervals and irrigation amount, and it is helpful for crop growth and yield increase Rreferences.
Google Scholar
[1]
Hecht-Nielsen, R Komogorovs mapping neural network existence theorem. Proceedings of the Intermational Conference on Neural Networks[C]. New York: IEEE Rress, 1987, 3: 11-13.
Google Scholar
[2]
Renato Silvio da Frota Ribeiro. Fuzzy logic based automated irrigation control system optimized via neural networks[D]. America：The university of Tennessee，(1998).
Google Scholar
[3]
Paul Dominic Colaizzi. Ground Based Remote Sensing for Irrigation Management in Precision Agriculture [D]. America：The university of Arizona，(2001).
Google Scholar
[4]
Quanxing Zhang, John and Ken Tilt. Application of Fuzzy Logic in an Irrigation Control System [A]. Proceedings of the IEEE International Conference on Industrial Technology, America, 1996: 593-597.
DOI: 10.1109/icit.1996.601660
Google Scholar
[5]
Adunias Dos Santos Teixerra. Dynamic programming model for quasi-real time reservoir operation and irrigation scheduling [A]. A dissertation presented for the PHD. The University of California Davis. America, 1998.
Google Scholar
