本文已被:浏览 414次 下载 885次 |
![本文二维码信息](/download_upload_file.aspx?file_name=/uploadfile/two_dimensional_code_img/zgnydxxb/2023/3/20230318.png) 码上扫一扫! |
基于CWT-RF模型估算博斯腾湖湖滨绿洲土壤有机碳含量 |
孟珊1,2,李新国1,2*,江远东1,2,麦麦提吐尔逊·艾则孜1,2
|
|
(1.新疆师范大学 地理科学与旅游学院, 乌鲁木齐 830054;2.新疆干旱区湖泊环境与资源实验室, 乌鲁木齐 830054) |
|
摘要: |
为实现博斯腾湖湖滨绿洲土壤有机碳含量的快速估算,结合实测的土壤高光谱数据与土壤有机碳数据,通过连续小波变换(CWT)进行土壤光谱数据预处理,利用相关系数法(CC)、连续投影算法(SPA)、竞争自适应重加权采样(CARS)、遗传算法(GA)筛选的特征波段作为建模输入量,构建随机森林(RF)模型。结果表明:研究区土壤有机碳含量平均值随土层深度增加由12.36 g/kg降低至10.49 g/kg,变异系数平均值为69.62%,空间异质性较强;CWT变换可以有效提高不同土层深度土壤有机碳含量与光谱反射率间的相关性,不同土层深度相关系数均值平均提升约22.41%;光谱数据经过CWT变换构建的模型精度明显提升,RF模型验证集R2与RPD分别平均提高7.09%、10.06%。CC、CARS、SPA、GA方法能消除光谱信息冗余,有效降低CWT-RF模型的输入量与RMSE值,土层深度0~20、20~40、40~60和60~80 cm筛选的特征波段平均压缩至全波段数目分别为8.51%、5.38%、2.21%和3.67%;RMSE值分别平均降低111.67%、135.61%、12.25%和74.96%,有效提升了建模速率与模型精度。利用CWT-SPA-RF模型对博斯腾湖湖滨绿洲0~80 cm土壤有机碳含量进行估算的效果最佳,构建的模型验证集R2≥0.77,RMSE≤3.06 g/kg,RPD≥2.07。 |
关键词: 土壤有机碳含量 土壤高光谱数据 连续小波变换 随机森林 特征波段筛选 湖滨绿洲 |
DOI:10.11841/j.issn.1007-4333.2023.03.18 |
投稿时间:2022-08-05 |
基金项目:新疆维吾尔自治区自然科学基金项目(2022D01A214);国家自然科学基金项目(41661047,U2003301) |
|
Estimation of soil organic carbon content in lakeside oases based on CWT-RF model |
MENG Shan1,2,LI Xinguo1,2*,JIANG Yuandong1,2,EZIZ·Mamattursun1,2
|
(1.College of Geographic Sciences and Tourism, Xinjiang Normal University, Urumqi 830054, China;2.Xinjiang Laboratory of Lake Environment and Resources in Arid Zone, Urumqi 830054, China) |
Abstract: |
To achieve rapid estimation of soil organic carbon content in the lakeshore oasis of Bosten lake, both the soil hyperspectral data and soil organic carbon data measured were combined, the soil spectral data were preprocessed by continuous wavelet transform(CWT). The feature bands screened by correlation coefficient analysis(CC), continuous projection algorithm(SPA), competitive adaptive reweighted sampling(CARS), and genetic algorithm(GA)methods were used in modeling. A random forest(RF)model was constructed using the feature bands screened by continuous projection algorithm(SPA), competitive adaptive reweighted sampling(CARS)and genetic algorithm(GA). The results showed that: The average soil organic carbon content in the study area decreased from 12. 36 g/kg to 10. 49 g/kg with the increasing of soil depth The average coefficient of variation was 69. 62% displaying strong spatial heterogeneity. The continuous wavelet transform effectively improved the correlation between soil organic carbon content and spectral reflectance at different soil depths, and the average value of correlation coefficient at different soil depths increased by 22. 41%. The accuracy of the model constructed by continuous wavelet transform of spectral data was significantly improved, and the RF model validation set R2 and RPD were improved by 7. 09% and 10. 06% on average, respectively. CC, CARS, SPA and GA algorithms eliminated the redundancy of spectral information and effectively reduced the input and RMSE values of the CWT-RF model. The characteristic bands screened at soil depths of 0-20, 20-40, 40-60 and 60-80 cm were compressed to 8. 51%, 5. 38%, 2. 21% and 3. 67% of the full spectral number on average, respectively. The RMSE values were reduced by 111. 67%, 135. 61%, 12. 25%, 74. 96% on average, which effectively improved the modeling rate and model accuracy. In conclusion, the CWT-SPA-RF model was the best for estimating the organic carbon content of 0-80 cm soil in the lakeside oasis, with the constructed model validation set R2≥0. 77, RMSE≤3. 06 g/kg and RPD≥2. 07. |
Key words: soil organic carbon content soil hyperspectral data continuous wavelet transform random forest characteristic band screening lakeside oasis |