基于辅助变量和神经网络模型的土壤有机质空间分布模拟  

SIMULATION OF DISTRIBUTION OF SOIL ORGANIC MATTER BASED ON AUXILIARY VARIABLES AND NEURAL NETWORK MODEL

作　　者：江叶枫 郭熙[1,2] 叶英聪 孙凯[1] 饶磊[1] 宋青励 

机构地区：[1]江西农业大学江西省鄱阳湖流域农业资源与生态重点实验室/国土资源与环境学院,江西南昌330045 [2]南方粮油作物协同创新中心,湖南长沙410000 [3]东华理工大学马克思主义学院,江西南昌330013

出　　处：《长江流域资源与环境》2017年第8期1150-1158,共9页Resources and Environment in the Yangtze Basin

基　　金：国家自然科学基金项目(41361049);江西省自然科学基金项目(20122BAB204012);江西省赣鄱英才"555"领军人才项目(201295)~~

摘　　要：为快速准确获取省域尺度下土壤有机质的空间分布状况。以江西省2012年测土配方施肥项目采集的16 582个耕地表层(0~20 cm)土壤样点数据,借助四方位搜索法、地统计学和遥感影像分析技术提取环境因子和邻近信息作为辅助变量,构建基于地理坐标与辅助变量的BP神经网络模型和普通克里金法结合的方法(BPNN_OK)、基于地理坐标与辅助变量的RBF神经网络模型和普通克里金法结合的方法(RBFNN_OK)和普通克里金法(OK法)3种方法,模拟省域尺度下耕地表层(0~20 cm)土壤有机质的空间分布。对2 416个验证样点进行独立验证的研究结果显示:基于辅助变量的神经网络模型较普通克里金法有较大提升。BPNN_OK法对土壤有机质预测结果的均方根误差、平均绝对误差、平均相对误差较OK法分别降低了2.76 g/kg、2.34 g/kg、9.83%,RBFNN_OK法较OK法分别降低了2.70 g/kg、2.29 g/kg、9.61%。研究显示,基于辅助变量的神经网络模型与OK法结合的方法明显地提高了土壤有机质空间分布模拟精度,并且存在改进和提高的空间。Accurate spatial information about soil organic matter （SOM） is critical for farmland use and soil environmental protection. In order to find the best interpolation method of SOM at the provincial scale, here we proposed there methods, back propagation neural network combined with ordinary kriging （BPNN_ OK, based on geographic coordinates, environmental factors and neighbor information as auxiliary variables） , radial basis function neural network with ordinary kriging （RBFNN_ OK, based on geographic coordinates, environmental factors and neighbor information as auxiliary variables） and ordinary kriging （OK） , to predict the distribution of SOM. Environmental factors were extracted by digital terrain and remote sensing image analysis technique. The four-direction search method was applied to get the neighbor information. To establish and validate this method, 16 109 soil samples were collected during the project of soil-test-based formulated fertilization in Jiangxi Province in 2012 and randomly divided into two groups, as modeling points （ 13 693） and validation points （2 416）. The results show that three methods produced the similar SOM maps. The error analyses indicated： Based on auxiliary variables and neural network model has greatly improved than OK method. Compare to OK, the root mean square errors （RMSE）, mean absolute errors （MAE） and mean relative errors （MRE） of BPNN_ OK were reduced 2.76 g/kg, 2.34 g/kg, 9.83%, RBFNN OK were reduced 2.70 g/kg, 2.29 g/ kg, 9.61%. This result suggested that it is helpful for improving the prediction accuracy to employ artificial neural network model in spatial prediction of SOM, and this model provides guidance how to select the model to predict soil nutrient at provincial scale, but could be improved in the future.

关 键 词：土壤有机质 辅助变量 神经网络模型 空间分布模拟 

分 类 号：S153.6[农业科学—土壤学]