{"title":"利用区域可见光和近红外光谱库及机器学习预测热带耕地中的土壤有机碳组分","authors":"","doi":"10.1016/j.still.2024.106297","DOIUrl":null,"url":null,"abstract":"<div><p>Soil organic carbon (SOC) is not a single and uniform entity, therefore understanding SOC fractions, particularly particulate organic carbon (POC) and mineral-associated organic carbon (MAOC), offers valuable insights into SOC dynamics. However, traditional laboratory measurements of SOC fractions are labor-intensive and costly. Therefore, leveraging rapid and cost-effective soil spectroscopy holds significant promise for addressing this challenge. While previous studies have concentrated on predicting SOC fractions using mid-infrared (MIR) spectroscopy, the potential of visible and near-infrared (VNIR) spectroscopy remains relatively unexplored, especially for tropical soils. To fill this gap, we evaluated six machine learning approaches, including three global models (Cubist, random forest (RF), partial least squares regression (PLSR)) and three local models (memory-based learning fitted by applying partial least squares regression (MBL-PLSR) and Gaussian process local regressions (MBL-GPR), non-linear memory-based learning (N-MBL)), for predicting POC and MAOC (g C kg<sup>−1</sup> soil) based on a regional soil VNIR spectral library (224 samples) from lateritic red soil in the tropical region of Guangdong Province, China. We also assessed the impact of variable selection on improving model performance by iteratively evaluating and removing insignificant predictor variables to determine the optimal number of predictors. The results showed that: (1) MBL-PLSR and N-MBL demonstrated commendable predictive performance, attaining coefficients of determination (R<sup>2</sup>) of 0.73 and 0.72 for POC, and 0.53 and 0.55 for MAOC on the validation set, respectively, outperforming Cubist and PLSR; (2) variable selection simplified predictive models by identifying the best spectral bands, leading to improved predictive accuracy for both POC (R<sup>2</sup> increased from 0.68 to 0.73) and MAOC (R<sup>2</sup> increased from 0.49 to 0.55); (3) the overall predictive performance of VNIR spectroscopy was higher for POC (R<sup>2</sup> of 0.73) compared to MAOC (R<sup>2</sup> of 0.55), while MAOC could be predicted more accurately by subtracting POC predictions from SOC observations (R<sup>2</sup> of 0.73). The favorable predictive accuracy underscores VNIR spectroscopy's viability for POC predictions. Additionally, MAOC can be well predicted by subtracting the predicted POC from the measured SOC. The outcomes of this study offers valuable insights for predicting SOC fractions using VNIR spectroscopy.</p></div>","PeriodicalId":49503,"journal":{"name":"Soil & Tillage Research","volume":null,"pages":null},"PeriodicalIF":6.1000,"publicationDate":"2024-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.sciencedirect.com/science/article/pii/S0167198724002988/pdfft?md5=93afa345294be3fa43d3480f6d4b430d&pid=1-s2.0-S0167198724002988-main.pdf","citationCount":"0","resultStr":"{\"title\":\"Prediction of soil organic carbon fractions in tropical cropland using a regional visible and near-infrared spectral library and machine learning\",\"authors\":\"\",\"doi\":\"10.1016/j.still.2024.106297\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><p>Soil organic carbon (SOC) is not a single and uniform entity, therefore understanding SOC fractions, particularly particulate organic carbon (POC) and mineral-associated organic carbon (MAOC), offers valuable insights into SOC dynamics. However, traditional laboratory measurements of SOC fractions are labor-intensive and costly. Therefore, leveraging rapid and cost-effective soil spectroscopy holds significant promise for addressing this challenge. While previous studies have concentrated on predicting SOC fractions using mid-infrared (MIR) spectroscopy, the potential of visible and near-infrared (VNIR) spectroscopy remains relatively unexplored, especially for tropical soils. To fill this gap, we evaluated six machine learning approaches, including three global models (Cubist, random forest (RF), partial least squares regression (PLSR)) and three local models (memory-based learning fitted by applying partial least squares regression (MBL-PLSR) and Gaussian process local regressions (MBL-GPR), non-linear memory-based learning (N-MBL)), for predicting POC and MAOC (g C kg<sup>−1</sup> soil) based on a regional soil VNIR spectral library (224 samples) from lateritic red soil in the tropical region of Guangdong Province, China. We also assessed the impact of variable selection on improving model performance by iteratively evaluating and removing insignificant predictor variables to determine the optimal number of predictors. The results showed that: (1) MBL-PLSR and N-MBL demonstrated commendable predictive performance, attaining coefficients of determination (R<sup>2</sup>) of 0.73 and 0.72 for POC, and 0.53 and 0.55 for MAOC on the validation set, respectively, outperforming Cubist and PLSR; (2) variable selection simplified predictive models by identifying the best spectral bands, leading to improved predictive accuracy for both POC (R<sup>2</sup> increased from 0.68 to 0.73) and MAOC (R<sup>2</sup> increased from 0.49 to 0.55); (3) the overall predictive performance of VNIR spectroscopy was higher for POC (R<sup>2</sup> of 0.73) compared to MAOC (R<sup>2</sup> of 0.55), while MAOC could be predicted more accurately by subtracting POC predictions from SOC observations (R<sup>2</sup> of 0.73). The favorable predictive accuracy underscores VNIR spectroscopy's viability for POC predictions. Additionally, MAOC can be well predicted by subtracting the predicted POC from the measured SOC. The outcomes of this study offers valuable insights for predicting SOC fractions using VNIR spectroscopy.</p></div>\",\"PeriodicalId\":49503,\"journal\":{\"name\":\"Soil & Tillage Research\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":6.1000,\"publicationDate\":\"2024-09-09\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"https://www.sciencedirect.com/science/article/pii/S0167198724002988/pdfft?md5=93afa345294be3fa43d3480f6d4b430d&pid=1-s2.0-S0167198724002988-main.pdf\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Soil & Tillage Research\",\"FirstCategoryId\":\"97\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0167198724002988\",\"RegionNum\":1,\"RegionCategory\":\"农林科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"SOIL SCIENCE\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Soil & Tillage Research","FirstCategoryId":"97","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0167198724002988","RegionNum":1,"RegionCategory":"农林科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"SOIL SCIENCE","Score":null,"Total":0}
Prediction of soil organic carbon fractions in tropical cropland using a regional visible and near-infrared spectral library and machine learning
Soil organic carbon (SOC) is not a single and uniform entity, therefore understanding SOC fractions, particularly particulate organic carbon (POC) and mineral-associated organic carbon (MAOC), offers valuable insights into SOC dynamics. However, traditional laboratory measurements of SOC fractions are labor-intensive and costly. Therefore, leveraging rapid and cost-effective soil spectroscopy holds significant promise for addressing this challenge. While previous studies have concentrated on predicting SOC fractions using mid-infrared (MIR) spectroscopy, the potential of visible and near-infrared (VNIR) spectroscopy remains relatively unexplored, especially for tropical soils. To fill this gap, we evaluated six machine learning approaches, including three global models (Cubist, random forest (RF), partial least squares regression (PLSR)) and three local models (memory-based learning fitted by applying partial least squares regression (MBL-PLSR) and Gaussian process local regressions (MBL-GPR), non-linear memory-based learning (N-MBL)), for predicting POC and MAOC (g C kg−1 soil) based on a regional soil VNIR spectral library (224 samples) from lateritic red soil in the tropical region of Guangdong Province, China. We also assessed the impact of variable selection on improving model performance by iteratively evaluating and removing insignificant predictor variables to determine the optimal number of predictors. The results showed that: (1) MBL-PLSR and N-MBL demonstrated commendable predictive performance, attaining coefficients of determination (R2) of 0.73 and 0.72 for POC, and 0.53 and 0.55 for MAOC on the validation set, respectively, outperforming Cubist and PLSR; (2) variable selection simplified predictive models by identifying the best spectral bands, leading to improved predictive accuracy for both POC (R2 increased from 0.68 to 0.73) and MAOC (R2 increased from 0.49 to 0.55); (3) the overall predictive performance of VNIR spectroscopy was higher for POC (R2 of 0.73) compared to MAOC (R2 of 0.55), while MAOC could be predicted more accurately by subtracting POC predictions from SOC observations (R2 of 0.73). The favorable predictive accuracy underscores VNIR spectroscopy's viability for POC predictions. Additionally, MAOC can be well predicted by subtracting the predicted POC from the measured SOC. The outcomes of this study offers valuable insights for predicting SOC fractions using VNIR spectroscopy.
期刊介绍:
Soil & Tillage Research examines the physical, chemical and biological changes in the soil caused by tillage and field traffic. Manuscripts will be considered on aspects of soil science, physics, technology, mechanization and applied engineering for a sustainable balance among productivity, environmental quality and profitability. The following are examples of suitable topics within the scope of the journal of Soil and Tillage Research:
The agricultural and biosystems engineering associated with tillage (including no-tillage, reduced-tillage and direct drilling), irrigation and drainage, crops and crop rotations, fertilization, rehabilitation of mine spoils and processes used to modify soils. Soil change effects on establishment and yield of crops, growth of plants and roots, structure and erosion of soil, cycling of carbon and nutrients, greenhouse gas emissions, leaching, runoff and other processes that affect environmental quality. Characterization or modeling of tillage and field traffic responses, soil, climate, or topographic effects, soil deformation processes, tillage tools, traction devices, energy requirements, economics, surface and subsurface water quality effects, tillage effects on weed, pest and disease control, and their interactions.