Spatial variability of soil nitrogen by BP neural network interpolation

Precise information about the spatial variability of soil nitrogen is essential in developing soil regionalization management and fertilization. A total of 523 soil samples were taken from top soil (0-20 cm) in the yellow brown soil in Shucheng county, Anhui province for testing the spatial variability of soil nitrogen and determining appropriate number of samples and interpolation method. Spatial variability of topsoil nitrogen was obtained using BP neural network interpolation at various number of samples. Kriging was conducted to compare with BP neural network interpolation under the same condition. Compare of results of BP neural network interpolation with Kriging indicated that G value of BP neural network interpolation was 7.75% higher than Kriging interpolation at 100 samples, and RMSE value was 0.1 lower than Kriging interpolation at the same number of samples, if higher than the 200 samples, G and RMSE value of neural networks interpolation and Kriging interpolation were basically the same. With the increased number of samples, interpolation accuracy also increased, the growth rate of G and RMSE was flattening. It can be concluded that neural network interpolation is a potential approach to spatial variability of soil nitrogen at small sample and large-scale random sampling.