引用本文:谢恩泽,赵永存,陆访仪,史学正,于东升.不同方法预测苏南农田土壤有机质空间分布对比研究[J].土壤学报,2018,55(5):1051-1061.
XIE Enze,ZHAO Yongcun,LU Fangyi,SHI Xuezheng,YU Dongsheng.Comparison Analysis of Methods for Prediction of Spatial Distribution of Soil Organic Matter Contents in Farmlands South Jiangsu, China[J].Acta Pedologica Sinica,2018,55(5):1051-1061
【打印本页】   【HTML】   【下载PDF全文】   查看/发表评论  【EndNote】   【RefMan】   【BibTex】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览 578次   下载 871 本文二维码信息
码上扫一扫!
分享到: 微信 更多
不同方法预测苏南农田土壤有机质空间分布对比研究
谢恩泽,赵永存,陆访仪,史学正,于东升
土壤与农业可持续发展国家重点实验室(中国科学院南京土壤研究所),土壤与农业可持续发展国家重点实验室(中国科学院南京土壤研究所),土壤与农业可持续发展国家重点实验室(中国科学院南京土壤研究所),土壤与农业可持续发展国家重点实验室(中国科学院南京土壤研究所),土壤与农业可持续发展国家重点实验室(中国科学院南京土壤研究所)
摘要:
准确把握土壤有机质(SOM)的空间分布规律对于土壤资源的高效持续利用具有重要意义。以江苏南部为研究区,以辅助因子与SOM的相关性强弱及辅助因子的可获取性为切入点,运用普通克里格(OK)、回归克里格(RK)和随机森林(RF)方法,结合地形、气候、土壤类型、土壤理化性质和施肥、碳投入等辅助数据预测了苏南地区农田SOM含量(0~20 cm)的空间分布。结果表明,三种方法预测的SOM空间分布总体趋势相似,表现为东高西低,但局部分异还存在差异;OK预测的精度最低,100次预测的均方根误差(RMSE)均值为6.97 g·kg-1。RK和RF的预测精度则均高于OK方法,表现为整合与SOM相关性最强的辅助因子全氮(TN)时,RK和RF预测的RMSE分别降低至5.25 g·kg-1和4.97 g·kg-1,而移除相关性最强的辅助因子TN后,RK和RF预测的RMSE亦较OK方法低,分别为6.21 g·kg-1和6.29 g·g-1;移除TN后,RK的预测精度稍高于RF,表明在其他辅助数据与SOM相关性相对较弱的条件下,RK方法有助于提高本研究区SOM预测精度;同时,尽管RK和RF的预测精度依然较OK高,但RK和RF对SOM方差的解释度则分别由51%和55%降低至了29%和28%。这表明,目前容易获取且相对廉价的辅助数据,对本研究区的SOM空间预测方面,还面临着数据质量低、预测精度不足等问题。
关键词:  空间预测  克里格插值模型  随机森林预测模型  有机质
DOI:10.11766/trxb201711250563
分类号:
基金项目:国家自然科学基金项目(41471177)和国家重点研发计划(2017YFA0603002)项目
Comparison Analysis of Methods for Prediction of Spatial Distribution of Soil Organic Matter Contents in Farmlands South Jiangsu, China
XIE Enze,ZHAO Yongcun,LU Fangyi,SHI Xuezheng and YU Dongsheng
Institute of Soil Science, Chinese Academy of Sciences,Institute of Soil Science, Chinese Academy of Sciences,Institute of Soil Science, Chinese Academy of Sciences,Institute of Soil Science, Chinese Academy of Sciences,Institute of Soil Science, Chinese Academy of Sciences
Abstract:
【Objective】Soil organic matter(SOM) plays a key role in maintaining soil quality and functions and crop productively and a critical role in global C recycling. Therefore, knowledge and understandings of spatial distribution patterns of SOM are important to sustainable utilization of soil resources, guarantee of food security, and mitigation of the momentum of global climate change.【Method】A total of 413 cropland topsoil samples (0~20 cm) were collected from the southern parts of Jiangsu Province for analysis of SOM contents, and ordinary kriging (OK), regression kriging (RK), and random forest (RF) methods were employed for mapping spatial distribution of SOM contents. Auxiliary data such as soil type, topographic factors derived from DEM, climate, soil temperature and moisture, soil properties, and agricultural management practices (N fertilization rate and C input by crop residues) varying in spatial resolution were first scaled into grids 300 m in resolution using either the kriging interpolation or neighborhood averaging method, and then the auxiliary factors screened by a stepwise-regression process were used in RK and RF predictions of SOM. Moreover, to identify impacts of the correlations between auxiliary factors and SOM content on spatial prediction of SOM contents in accuracy, the root mean square errors (RMSE) derived by RK and RF methods were also compared between the situations of removing and retaining the auxiliary factor with highest correlation coefficient.【Result】The SOM spatial distribution patterns derived with the OK, RK and RF methods were quite similar, that is to say, SOM contents in the eastern parts of the study area are relatively high, whereas those in the western parts are low. But local differences did exist in detail of SOM distribution prediction between the methods can be intuitively observed. The OK method was the lowest in prediction accuracy, with mean RMSE being 6.97 g·kg-1 and lower than the RK and RF methods, of which the mean RMSE of the RK and RF methods was lowered down to 5.25 and 4.97 g·kg-1, respectively, when total nitrogen (TN) of the auxiliary factors that were most closely related to SOM was integrated. However, when TN of these auxiliary factors was removed, the RMSE predicted with RK and RF was lower than that with OK, being 6.21 and 6.29 g·kg-1, respectively, while the average explained variance was decreased to 29% and 28%, respectively. However, the RK and RF methods are still better than the OK method, as the RMSE derived by OK was as high as 6.97 g·kg-1, and the explained variance of OK was only 9.7%.【Conclusion】RK and RF are both higher than OK in prediction accuracy, however, the difference in prediction accuracy between RK and RF depends on degree of the correlation between the auxiliary factors and SOM. When auxiliary factors most closely related to SOM, such as TN was included in the prediction, RF was better than RK; while those were excluded, , RK was slightly better than RF, indicating that RK is still promising due to the relatively high-cost of TN measurement. In addition, prediction accuracy of RF largely depends on degree of the correlation between the auxiliary data and SOM, when TN was removed from the RF prediction, the predicted RMSE increased significantly, indicating that the current easily attainable and available low-cost auxiliary are still facing many challenges in improving SOM prediction accuracy in plain regions with strong anthropogenic influences. Consequently, developments of new scaling methodology for raw auxiliary data or new higher resolution auxiliary data for quantifying relationships between auxiliary data and SOM are critical for improving accuracy of the prediction of SOM in plain areas with intensifying anthropogenic influences.
Key words:  Spatial prediction  Kriging  Random forest models  Soil organic matter