Prediction of soil moisture with complex-valued neural network

Abstract

Soil moisture is a critical state affecting a variety of land surface and subsurface processes. In this paper we report investigation results of multilayer neural network with multi-valued neurons (MLMVN), it is a distinct type of complex-valued neural network with derivative-free back-propagation algorithm. We examined the proposed method by using 4752 soil moisture data set at 30cm underground and environmental factors (rainfall, temperature and wind speed) collected respectively. Firstly, in order to smooth the data, outliers and missing values were replaced by the mean values of the neighbors. Meanwhile, from the autocorrelation and timing diagram can be seen that soil moisture was non-stationary nonlinear time series, and rainfall, temperature and wind speed had significant influence on soil moisture according to the correlation analysis. Secondly, principal component analysis (PCA) was used to eradicate the correlation of initial input parameters (soil moisture, rainfall, temperature and wind speed), and the first three principal components were nominated to restructure the samples into a lower dimensions, to reduce the scale of network and improve network performance. Finally, transformed the restructured samples into complex values as inputs and outputs of MLMVN network. The experimental results show that, in multi-step ahead soil moisture prediction, two hidden layers PCA_MLMVN network outperforms the MLMVN network in term of prediction accuracy with the average prediction accuracy reached 92.8%, enhanced 4.5% compared with the MLMVN network. The result shows that PCA_MLMVN comfirm a good performance in the long-term prediction of soil moisture and show little accumulating errors for multi-step ahead predictions.