基于BP神经网络插值的土壤全氮空间变异

Spatial variability of soil nitrogen by BP neural network interpolation

大尺度土壤养分空间变异研究可以为土壤改良分区治理提供基础数据。寻求合适的取样数和插值方法是进行土壤养分空间变异研究的关键。以安徽省舒城县为例,共取得0～20cm土壤表层样品523个,土壤全氮的空间变异由BP神经网络插值方法在不同取样数条件下获得,通过与克里格插值法进行比较得出:样本数在100个时,神经网络插值的预测吻合度(G)比克里格插值高7.75%,均方根误差(RMSE)低0.1,总体精度优于克里格;样本数大于200时,神经网络插值和克里格插值精度基本相同,随着采样数量增加,两种方法的插值精度也在提高,并逐步趋于平稳。在大尺度土壤养分空间变异研究中,在小样本情况下,神经网络插值具有优势。

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.