| 李雨鸿,冯锐,纪瑞鹏,张霞,武晋雯,于文颖,王婷,李晶.基于CR-CARS-RF的黑土区土壤有机质含量高光谱估算[J].干旱地区农业研究,2025,(4):203~210 |
| 基于CR-CARS-RF的黑土区土壤有机质含量高光谱估算 |
| Hyper\|spectral estimation of soil organic matter content in black soil based on CR-CARS-RF |
| |
| DOI：10.7606/j.issn.1000-7601.2025.04.21 |
| 中文关键词: 黑土 土壤有机质 高光谱 竞争自适应重加权算法 随机森林模型 |
| 英文关键词:black soil soil organic matter hyper\|spectral competitive adaptive re\|weighting algorithm random forest model |
| 基金项目:辽宁省农业气象灾害重点实验室开放基金（2023SYIAEKFZD06）；中国气象局创新发展专项（CXFZ2023J059）；辽宁省自然科学基金（2023-MS-042）；中国气象局重点创新团队建设项目（CMA2024ZD02） |
|
| 摘要点击次数: 853 |
| 全文下载次数: 1055 |
| 中文摘要: |
| 以辽宁昌图和苏家屯区149个黑土样本为基础，结合地面高光谱数据与实验室土壤有机质（SOM）含量化学测定，系统分析了不同高光谱预处理与建模方法对SOM反演精度的影响，旨在构建快速且精准的SOM含量高光谱反演模型。通过Savitzky-Golay平滑法（SG）对光谱数据进行去噪处理后，对比分析了倒数（1/R）、倒数的对数（log(1/R)）、一阶导数（FDR）、标准化（SNV）及连续统去除（CR）五种光谱变换方法，并采用皮尔逊相关分析（PCC）和竞争自适应重加权（CARS）算法筛选特征波段，结合偏最小二乘（PLSR）、多元线性回归（MLR）和随机森林（RF）三种建模方法，构建了12种反演模型，对各模型的预测精度进行对比分析。结果表明：（1）CR变换显著增强了光谱与SOM的相关性，2 166 nm之后的128个波段相关系数绝对值超过0.5，最高达0.75；（2）CARS算法有效压缩了特征波段数量，将其控制在全波段的6%以下，与PCC-PLSR和PPC-RF相比，CARS-PLSR和CARS-RF模型预测精度显著提升，其决定系数（R2）分别提高13.4%和14.5%，均方根误差（RMSE）分别降低12.8%和11.9%；（3）非线性RF模型的预测精度最优，与MLR和PLSR相比，其R2分别提升32.1%和3.5%，RMSE分别降低34.9%和4.4%；（4）在12种预测模型中，CR-CARS-RF模型表现最佳，其建模R2为0.91，RMSE为1.76 g·kg-1；预测R2为0.79，RMSE为2.49 g·kg-1，表明CR-CARS-RF模型具有较高的预测精度和可靠性，可为辽宁黑土区土壤有机质含量的高效精准监测提供有力的技术支撑。 |
| 英文摘要: |
| This study, based on 149 black soil samples from Changtu and Sujiatun districts in Liaoning Province, combined ground hyperspectral data with laboratory chemical determination of soil organic matter (SOM) content, systematically analyzed the impacts of different hyperspectral preprocessing and modeling methods on SOM inversion accuracy, and aimed to construct a rapid and precise hyperspectral inversion model for SOM content. After denoising spectral data using the Savitzky\|Golay smoothing method, five spectral transformation methods were compared: reciprocal (1/R), logarithmic reciprocal (log(1/R)), first derivative (FDR), standard normal variate (SNV), and continuum removal (CR). Pearson correlation coefficient (PCC) analysis and competitive adaptive reweighted sampling (CARS) algorithm were used for feature band selection. Three modeling methods—partial least squares regression (PLSR), multiple linear regression (MLR), and random forest (RF)—were combined to construct 12 inversion models, with their prediction accuracies compared and analyzed. The results showed that: (1) CR transformation significantly enhanced the correlation between spectra and SOM, with the absolute values of correlation coefficients exceeding 0.5 for 128 bands after 2 166 nm, reaching up to 0.75. (2) The CARS algorithm effectively reduced the number of feature bands to below 6% of the full bands, and compared with PCC-PLSR and PCC-RF, the prediction accuracies of CARS-PLSR and CARS-RF models were significantly improved, with their coefficients of determination (R2) increasing by 13.4% and 14.5%, and root mean square errors (RMSE) decreasing by 12.8% and 11.9%, respectively. (3) The nonlinear RF model achieved the best prediction accuracy, with R2 increasing by 32.1% and 3.5%, and RMSE decreasing by 34.9% and 4.4% compared to MLR and PLSR, respectively. (4) Among the 12 prediction models, the CR-CARS-RF model performed optimally, with a modeling R2 of 0.91 and RMSE of 1.76 g·kg-1, and a prediction R2 of 0.79 and RMSE of 2.49 g·kg-1. These results indicate that the model has high prediction accuracy and reliability, providing a strong technical support for efficient and precise monitoring of SOM in black soil region of Liaoning. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| | |