Spatial Interpolation Of Soil Moisture And Nutrients Using Bayesian Maximum Entropy And Neural Networks Ensemble

Precision agriculture emphasizes utilizing production potential evenly and reasonably by means of distributed management of soil moisture and nutrients so as to obtain higher economic yields or significant social benefits. It is prerequisite for distributed management to understand spatial distribution patterns of soil moisture and nutrients. and the effective approach of understanding the variation pattern of soil moisture and nutrients is spatial interpolation based on limited soil data, which is important to realization of precise irrigation and variable fertilization.This paper based on data collected from 161 sampling points in a farmland at northern Yangzhou, Jiangsu province. The data include soil moisture content, total nitrogen content, organic matter, available nitrogen, readily available potassium, and readily available phosphorus. Focusing on the characteristic of spatial variability of soil moisture and nutrients in the research area, combined features of the spatial interpolation methods, the following researches were conducted: ①using the characteristic that significant difference of soil properties exists in different fields but not in the same piece of field, and the uncertainty of estimation of the expression of the distributed characteristic of soil variables in the same field (building soft data), this method combined these soft data with measured data to simulate the spatial distribution of soil variables by utilizing the modern Geostatistics method-Bayesian maximum entropy (BME), and hereinafter it referred to as MVBME;②using non-linear expression of neural network method, and with the aid of the uncertainty of estimation of neural network ensemble(building soft data), this method integrated the above results into BME, and simulated the spatial distribution of soil variables using BME integrated with soft data, and hereinafter it referred to as NNEBME.③adopting several random sampling plans (train sample and test sample), compare the above spatial interpolation results with the estimated results from the Radial Basis Function (RBF) method, the Neural Network Ensemble (NNE) method, the Ordinary Kriging (OK) method, and the Residual Kriging (RK). The main findings are as follows:1) The spatial variability of moisture content in the research area was weak, and the spatial variability of nutrients was moderate. The spatial variability of the whole experimental field was much stonger than which of a single piece of field, and the spatial variability of soil properties in different fields was highly significant, while no significant difference existed in the same field.2) Before the outlier was corrected, the soil moisture content, total nitrogen content, organic matter, and available nitrogen followed a gaussian distribution, readily available potassium and readily available phosphorus did not follow a gaussian distribution. After outlier was corrected, the soil moisture content and nutrients both followed an approximate gaussian distribution.3) Spatial interpolation was carried out on the soil content and nutrients in the research area using MVBME, and the result was compared with the interpolation results of RBF, RK, and OK. The results showed that the mean error (ME) of MVBME was the smallest among these four spatially interpolation methods, and that MVBME reduced mean square error (MSE) by 23.77%-69.14% and 0.41%-56.17% reduction compared with RBF and RK respectively. The results also indicated that MVBME can reduce the MSE of soil organic matter, available nitrogen, readily available potassium and readily available phosphorus by 6.24%-52.37% in comparison with OK, and that in most cases, the MSE of soil moisture content and total nitrogen content was reduced by 10.25%-38.18%. MVBME was closest to unbiased estimation amongst four spatial interpolation methods, and it reflected the degree of fluctuation of variables most precisely.4) Spatial interpolation was carried out on the soil content and nutrients in the research area using NNEBME, and the result was compared with the interpolation results of NNE, RK, and OK. The results showed that ME of the value estimated by NNEBME was the most close to zero, which was approximated to unbiased estimation, and that NNEBME reduced MSE by 1.64%-45.20% in comparison with NNE. It is also demonstrated by the results that, except for soil moisture, NNEBME reduced the MSE by 0-40.05% compared with OK and RK. With the decreasing of the number of known points (sample capacity of modeling data), the advantage of NNEBME in spatial interpolation was more prominent. The MSE decomposition showed that NEBME has better ability to estimate mean and reproduce the fluctuation degree of soil variables.This paper employed the characteristics of spatial variation of farmland and the features of the spatial interpolation methods to build "soft data" which can be effectively used in Bayesian maximum entropy. It not only expanded the application of modern Geostatictic BME in Agricultural Soil and Water Sciences field, but also introduced a new thought for improving the interpolation precision of soil moisture content and nutrients.