基于BP神经网络的土壤容重预测模型

The Model for Predicting Soil Bulk Density Based on the BP Neural Network

土壤容重是农业生产和研究的重要参数,但因成本高、工作量大等原因,其获取仍是一个紧迫的问题。数学建模技术使得科研人员尝试使用土壤传递函数PTFs间接获取土壤容重值。本研究以复合圆锥指数仪为工具,探讨应用BP神经网络建立PTFs预测土壤容重。选取粘土和粉质壤土作为实验对象,在MATLAB2008a上建立、并评价预测土壤容重的BP神经网络,即PTFs模型。研究中以均方根误差RMSE和决定系数R2为性能指标来评价所建BP神经网络。结果表明,针对复合圆锥指数仪的测量结果,应用BP神经网络算法建立PTFs可以有效预测土壤容重。粘土容重预测的决定系数R2达到0.6973,粉质壤土容重预测的决定系数R2达到0.6868。实验结果还证实土壤容重预测与测量深度无关,但与土壤类型显著相关。

Soil bulk density is becoming more and more significant in modern agricultural production andresearch. However, it has not drawn enough attention in both practical production and scientific researchbecause it is expensive and labor-some to obtain the exact value traditionally. Fortunately, the development ofmathematical modeling techniques enables researchers to get soil bulk density indirectly using PTFs（pedotransfer functions） recently. In order to study the feasibility of PTFs＇ prediction for soil bulk density, theauthor applied BP neural network algorithm to establish PTFs. The study materials were 2 kinds of soil sampleswith different physical parameter, which were made in lab environment. The data used included the bulkdensity, cone index and volumetric water content of the examined samples, to train and test BP networks. Theywere measured with the help of a dual-sensor penetrometer. In this study, 6 PTFs based on BP neural networkalgorithm were established and evaluated by the root mean square error（RMSE） and the determinationcoefficient（R2） in the platform of MATLAB2008 a. 2 types of soil samples（clay and silt-loam） with differentphysical parameters were made as study materials. According to the results, PTFs using BP neural networkalgorithm could effectively predict the soil bulk density with data from a dual-sensor penetrometer. The R2 ofclay PTF for bulk density reached 0.6973, and the R2 of silt-loam PTF for bulk density reached 0.6868. It alsosupported that soil bulk density prediction using BP networks was significantly related with soil type, while hadnothingto do with penetration depth.