应用集成BP神经网络模型预测土壤有机质空间分布

摘要

Based on the data collected from the project of soiltest-based formulated fertilization in Wannian county, Jiangxi province in 2014, a back propagation neural network ensemble model ( BPNN-Ada) was used to predict the spatial distribution of soil organic matter ( SOM ) which was then compared to those by back propagation neural network model (BPNN) and ordinary Kriging model (OK). The BPNN-Ada and BPNN model were trained using the geographical coordi-nates, elevation, slope and adjacent sampling points information as inputs. The prediction accuracy of three models followed the order of BPNN-Ada>BPNN>OK. BPNN-Ada model could help to produce the SOM map with higher accuracy and better effect, which was consistent with the true geographical information and actual situation of SOM. By overcoming the shortcom-ings of poor local search ability and easiness to fall into global optimum, BPNN-Ada improved the generalization ability of BPNN.%基于2014年江西省万年县测土配方施肥数据,以地理坐标、高程和坡度以及邻近样点信息作为网络的输入变量,采用集成BP神经网络模型(BPNN-Ada模型)预测土壤有机质的空间分布,并与未集成的BP神经网络模型(BPNN模型)和普通克里金模型(OK模型)进行比较.结果表明,3种模型的预测精度大小顺序为BPNN-Ada模型>BPNN模型>OK模型.集成BP神经网络模型预测精度最高,效果最好,比较符合土壤有机质地学分布规律及实际情况.BPNN-Ada模型克服了BP神经网络局部搜索能力差和易陷入全局最优的缺点,提高了BP神经网络的泛化能力.