Journal of Integrative Agriculture ›› 2024, Vol. 23 ›› Issue (8): 2820-2841.DOI: 10.1016/j.jia.2024.01.015

• • 上一篇    下一篇

应用时序遥感数据增强农田土壤有机质制图模型性能

  

  • 收稿日期:2023-07-25 接受日期:2023-10-17 出版日期:2024-08-20 发布日期:2024-07-29

Improving model performance in mapping cropland soil organic matter using time-series remote sensing data

Xianglin Zhang1, Jie Xue2, Songchao Chen1, 3, Zhiqing Zhuo4, Zheng Wang1, Xueyao Chen1, Yi Xiao1, Zhou Shi1, 5#   

  1. 1 Institute of Applied Remote Sensing and Information Technology, College of Environmental and Resource Sciences, Zhejiang University, Hangzhou 310058, China
    2 Department of Land Management, Zhejiang University, Hangzhou 310058, China
    3 ZJU-Hangzhou Global Scientific and Technological Innovation Center, Zhejiang University, Hangzhou 311215, China
    4 Institute of Digital Agriculture, Zhejiang Academy of Agricultural Sciences, Hangzhou 310021, China
    5 Key Laboratory of Spectroscopy Sensing, Ministry of Agriculture and Rural Affairs, Hangzhou 310058, China
  • Received:2023-07-25 Accepted:2023-10-17 Online:2024-08-20 Published:2024-07-29
  • About author:Xianglin Zhang, E-mail: zhangxianglin@zju.edu.cn; #Correspondence Zhou Shi, E-mail: shizhou@zju.edu.cn
  • Supported by:
    This study was supported by the National Natural Science Foundation of China (U1901601) and the National Key Research and Development Program of China (2022YFB3903503). 

PDF

摘要：

面对日益严重的全球土壤退化问题，明确农田土壤有机质的空间分布格局对于土壤碳库核算、农田质量评价和制定有效的管理政策具有重要意义。作为一种空间信息预测技术，数字土壤制图已被广泛应用于不同区域尺度的土壤信息空间制图上。然而，由于精确量化人为干扰因素存在较大的难度，针对农田尺度的土壤有机质制图反演模型的精度往往低于其它土地覆被类型。为解决该问题，本研究使用2021年广州采集的462个农田土壤样本系统评估“信息提取-特征选择-模型融合”框架在提升农田土壤有机质反演精度的潜力。本研究证明“信息提取-特征选择-模型融合”框架可以在有效提升最终反演结果的精度（R2：从0.48到0.53）并且不会对模型不确定性造成显著的负面影响。量化环境动态变化信息是一种产生与土壤有机质线性和非线性相关协变量的有效方法。应用该方法产生的环境协变量可将随机森林模型的R2从0.44提高到0.48，将极端梯度提升模型的R2从0.37提高到0.43。当环境协变量较少（< 200）时推荐使用前向递归特征筛选算法，当环境协变量较（> 500）时推荐使用Boruta特征筛选算法。Granger-Ramanathan模型融合方法可以组合不同初始预测模型的优势以提高预测结果精度并平均不确定性。当初始预测模型结构相似时，参与融合的初始预测模型数量的增加对最终预测没有显著的影响。鉴于上述优势，“信息提取-特征选择-模型融合”框架对提高不同区域尺度数字土壤制图精度方面具有较高的潜力，其制图结果可以为土壤保护政策的制定提供有效参考。

Abstract:

Faced with increasing global soil degradation, spatially explicit data on cropland soil organic matter (SOM) provides crucial data for soil carbon pool accounting, cropland quality assessment and the formulation of effective management policies.  As a spatial information prediction technique, digital soil mapping (DSM) has been widely used to spatially map soil information at different scales.  However, the accuracy of digital SOM maps for cropland is typically lower than for other land cover types due to the inherent difficulty in precisely quantifying human disturbance.  To overcome this limitation, this study systematically assessed a framework of “information extraction-feature selection-model averaging” for improving model performance in mapping cropland SOM using 462 cropland soil samples collected in Guangzhou, China in 2021.  The results showed that using the framework of dynamic information extraction, feature selection and model averaging could efficiently improve the accuracy of the final predictions (R2: 0.48 to 0.53) without having obviously negative impacts on uncertainty.  Quantifying the dynamic information of the environment was an efficient way to generate covariates that are linearly and nonlinearly related to SOM, which improved the R2 of random forest from 0.44 to 0.48 and the R2 of extreme gradient boosting from 0.37 to 0.43.  Forward recursive feature selection (FRFS) is recommended when there are relatively few environmental covariates (<200), whereas Boruta is recommended when there are many environmental covariates (>500).  The Granger-Ramanathan model averaging approach could improve the prediction accuracy and average uncertainty.  When the structures of initial prediction models are similar, increasing in the number of averaging models did not have significantly positive effects on the final predictions.  Given the advantages of these selected strategies over information extraction, feature selection and model averaging have a great potential for high-accuracy soil mapping at any scales, so this approach can provide more reliable references for soil conservation policy-making.


Key words: cropland , soil organic matter ,  digital soil mapping ,  machine learning ,  feature selection ,  model averaging