中国农业科学 ›› 2022, Vol. 55 ›› Issue (19): 3738-3750.doi: 10.3864/j.issn.0578-1752.2022.19.005

• 耕作栽培·生理生化·农业信息技术 • 上一篇    下一篇

基于CARS-BPNN的江西省土壤有机碳含量高光谱预测

吴俊1(),郭大千3,李果2,4,郭熙1,2(),钟亮1,朱青1,国佳欣1,叶英聪1   

  1. 1江西农业大学国土资源与环境学院/江西省鄱阳湖流域农业资源与生态重点实验室,南昌 330045
    2中科生态修复(江西)创新研究院,南昌 330045
    3江西省国土空间调查规划研究院,南昌 330045
    4江西省地质局912大队,南昌 330045
  • 收稿日期:2021-12-07 接受日期:2022-03-30 出版日期:2022-10-01 发布日期:2022-10-10
  • 通讯作者: 郭熙
  • 作者简介:吴俊,E-mail: JuneWu6667@163.com
  • 基金资助:
    国家自然科学基金(42071068)

Prediction of Soil Organic Carbon Content in Jiangxi Province by Vis-NIR Spectroscopy Based on the CARS-BPNN Model

WU Jun1(),GUO DaQian3,LI Guo2,4,GUO Xi1,2(),ZHONG Liang1,ZHU Qing1,GUO JiaXin1,YE YingCong1   

  1. 1College of Land Resources and Environment, Jiangxi Agricultural University/Key Laboratory of Poyang Lake Watershed Agricultural Resources and Ecology of Jiangxi Province, Nanchang 330045
    2Ecological Restoration and Innovation Research Institute of Jiangxi Province, Nanchang 330045
    3The National Land and Space Survey and Planning Research Institute of Jiangxi Province, Nanchang 330045
    4912 Brigade, Geological Bureau of Jiangxi Province, Nanchang 330045
  • Received:2021-12-07 Accepted:2022-03-30 Online:2022-10-01 Published:2022-10-10
  • Contact: Xi GUO

摘要:

【目的】探讨光谱变量选择及依据土壤类型进行分层校准两种方法对高光谱预测土壤有机碳(SOC)精度的影响。【方法】以江西省为研究区,490个土壤样本为研究对象,对研究区内的所有样本以及不同土壤类型样本分别通过竞争性自适应重加权采样(CARS)算法筛选特征波段,并采用偏最小二乘回归(PLSR)、支持向量机(SVM)、随机森林(RF)、反向传播神经网络(BPNN)4种模型,对比不同土壤类型下SOC在全波段以及CARS算法筛选后特征波段的预测精度。进而,还对比了全局校准和分层校准下SOC在全波段以及CARS算法筛选后特征波段的预测精度。【结果】(1)红壤筛选的特征波段为484、683—714和2 219—2 227 nm,水稻土筛选的特征波段为484、689—702和2 146—2 156 nm。红壤采用CARS-BPNN模型预测效果最佳(R 2=0.82),较全波段建模验证集R 2提升0.07。水稻土采用CARS-RF模型预测效果最佳(R 2=0.83),较全波段建模验证集R 2提升0.13。(2)在总体样本上,分层校准相比全局校准精度有所提升。采用CARS-BPNN进行分层校准预测效果最佳(R 2=0.82),较全局校准验证集R 2提升0.06。【结论】采用CARS-BPNN进行分层校准能够较好地预测江西省土壤有机碳含量,本研究可为其他类似地区预测土壤属性提供科学依据。

关键词: 土壤有机碳, 竞争适应重加权采样, 分层校准, 随机森林, 反向传播神经网络

Abstract:

【Objective】 This study explored the roles of spectral variable selection and stratified calibration based on soil type in visible-near-infrared (Vis-NIR) spectroscopy for predicting soil organic carbon (SOC) content on a large spatial scale. 【Method】 A total of 490 samples were collected in Jiangxi province (Southeast China) and used for modeling with partial least squares regression (PLSR), support vector machine (SVM), random forests (RF), and back-propagation neural network (BPNN). The competitive adaptive reweighted sampling (CARS) procedure was used to select the feature bands of different soil types and total samples (i.e., sum of red soils and paddy soils). The prediction accuracy of models incorporating full bands or feature bands was evaluated for the different soil types. Further, the prediction accuracy of these models based on their global and stratification calibration was compared for the total samples. 【Result】 (1) The feature bands of red soils were 484, 683-714, and 2 219-2 227 nm, while those of paddy soils were 484, 689-702, and 2 146-2 156 nm. The CARS-BPNN model showed the best prediction performance for red soils (validation set R2 = 0.82), being 0.07 higher than that of BPNN with full bands. The CARS-RF model also had the best prediction performance for paddy soils (validation set R2 = 0.83), being 0.13 higher than that of RF with full bands. (2) Based on the stratified calibration, the best prediction performance was obtained using the CARS-BPNN model (validation set R2 = 0.82), which was 0.06 higher than that of the model based on global calibration. 【Conclusion】 The CARS-BPNN model combined with stratified calibration based on soil type could accurately predict SOC content in the study area.

Key words: soil organic carbon, competitive adaptive reweighted sampling, stratified calibration, random forest, back propagation neural network