Scholarlink

<p>Hyperspectral Remote Sensing (HRS) is emerging as a promising tool in natural resources management and is increasingly applied for soil health assessment. This study aims to evaluate the potential of hyperspectral remote sensing (HRS) integrated with machine learning (ML) as a non-destructive, rapid alternative to traditional wet chemistry methods for soil nutrient analysis. Traditional wet chemistry methods, though reliable, are often time-consuming, labour intensive, costly, reliant on hazardous reagents, and not easily scalable for large area monitoring, which limits their effectiveness in supporting rapid decision-making. Accurate and efficient assessment of soil nutrient status is essential for sustainable land management and precision agriculture and this research attempts to address the critical gap in scalable and timely soil health monitoring techniques across diverse agro-ecological regions in India. HRS provides a solution by capturing detailed soil spectral signatures that reflect nutrient variability.Field-level spectral data were collected from 159 soil samples across two micro-watersheds located in humid tropical and tropical regions. Spectral reflectance data covering wavelengths from 350 to 2500&#xa0;nm were used as predictors for estimating Nitrogen (N), Available Phosphorus (P), and Exchangeable Potassium (K). This range spans the visible–near infrared (VNIR) and shortwave infrared (SWIR) regions, which are highly sensitive to soil mineralogical and organic matter features that influence nutrient availability, thereby making it particularly suitable for detecting soil nutrients. Multiple pre-processing techniques (e.g., Savitzky-Golay Derivatives, Detrending) were applied to enhance spectral data quality. Machine learning models viz., Partial Least Squares Regression (PLSR), Support Vector Machine (SVM), Gaussian Process Regression (GPR) and Random Forest (RF)were employed and evaluated using R<sup>2</sup> and Root Mean Squared Error (RMSE).Savitzky-Golay Derivative pre-processing along with the SVM model with showed the best performance for nitrogen prediction (R<sup>2</sup> = 0.49; RMSE = 20.10). For phosphorus, GPR without transformation yielded the most accurate estimates (R<sup>2</sup> = 0.58; RMSE = 12.62). In predicting potassium, SVM with detrending performed relatively well (R<sup>2</sup> = 0.47; RMSE = 205.79). These moderately significant R<sup>2</sup> values and relatively high RMSE suggest that while the models capture meaningful soil nutrient variability, their predictive accuracy may be highly suitable for regional or watershed scale monitoring. This can be used for field level recommendations, where fertilizer recommendation is derived from range value of nutrients such as low, medium and high.Overall, the results demonstrated practical predictive power of hyperspectral signatures for key soil nutrients under field conditions.The study confirms the viability of using HRS combined with machine learning for soil nutrient prediction, offering a non-invasive and scalable approach for soil health assessment. These findings advance precision agriculture by enabling efficient monitoring of soil fertility dynamics, there by supporting sustainable land management and improved crop productivity in diverse agro-ecological settings.</p> Graphical Abstract <p>This study presents a field-level application of Hyperspectral Remote Sensing (HRS) integrated with machine learning techniques to assess soil nutrient levels enabling rapid nutrient diagnostics compared to conventional methods. Conducted across two micro-watersheds representing humid tropical and tropical agro-ecologies in India, the research utilized spectral reflectance data (350–2500&#xa0;nm) from 159 soil samples to predict Nitrogen (N), Available Phosphorus (P), and Exchangeable Potassium (K). A suite of machine learning algorithms including Support Vector Machine, Gaussian Process Regression, Random Forest, and Partial Least Squares Regression were evaluated in combination with eight spectral pre-processing techniques. The best-performing models were SVM for nitrogen (R<sup>2</sup> = 0.49, RMSE = 20.10) and exchangeable potassium (R<sup>2</sup> = 0.47, RMSE = 205.79), and GPR for phosphorus (R<sup>2</sup> = 0.58, RMSE = 12.62). The findings support the potential of HRS and machine learning for rapid, scalable soil fertility assessment, offering valuable tools for precision agriculture and sustainable soil health monitoring in diverse Indian agro-ecological regions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Soil Spectra for Smart Farming—A Scalable Alternative to Wet Chemistry Using Fusion of Hyperspectral and Machine Learning in Diverse Agro Ecologies of India

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • S. Rama Subramoniam,
  • K. Naveena,
  • Muhamed Fasil,
  • B. Chandrasekaran,
  • Girish Gopinath,
  • R. Hebbar,
  • U. Surendran

摘要

Hyperspectral Remote Sensing (HRS) is emerging as a promising tool in natural resources management and is increasingly applied for soil health assessment. This study aims to evaluate the potential of hyperspectral remote sensing (HRS) integrated with machine learning (ML) as a non-destructive, rapid alternative to traditional wet chemistry methods for soil nutrient analysis. Traditional wet chemistry methods, though reliable, are often time-consuming, labour intensive, costly, reliant on hazardous reagents, and not easily scalable for large area monitoring, which limits their effectiveness in supporting rapid decision-making. Accurate and efficient assessment of soil nutrient status is essential for sustainable land management and precision agriculture and this research attempts to address the critical gap in scalable and timely soil health monitoring techniques across diverse agro-ecological regions in India. HRS provides a solution by capturing detailed soil spectral signatures that reflect nutrient variability.Field-level spectral data were collected from 159 soil samples across two micro-watersheds located in humid tropical and tropical regions. Spectral reflectance data covering wavelengths from 350 to 2500 nm were used as predictors for estimating Nitrogen (N), Available Phosphorus (P), and Exchangeable Potassium (K). This range spans the visible–near infrared (VNIR) and shortwave infrared (SWIR) regions, which are highly sensitive to soil mineralogical and organic matter features that influence nutrient availability, thereby making it particularly suitable for detecting soil nutrients. Multiple pre-processing techniques (e.g., Savitzky-Golay Derivatives, Detrending) were applied to enhance spectral data quality. Machine learning models viz., Partial Least Squares Regression (PLSR), Support Vector Machine (SVM), Gaussian Process Regression (GPR) and Random Forest (RF)were employed and evaluated using R2 and Root Mean Squared Error (RMSE).Savitzky-Golay Derivative pre-processing along with the SVM model with showed the best performance for nitrogen prediction (R2 = 0.49; RMSE = 20.10). For phosphorus, GPR without transformation yielded the most accurate estimates (R2 = 0.58; RMSE = 12.62). In predicting potassium, SVM with detrending performed relatively well (R2 = 0.47; RMSE = 205.79). These moderately significant R2 values and relatively high RMSE suggest that while the models capture meaningful soil nutrient variability, their predictive accuracy may be highly suitable for regional or watershed scale monitoring. This can be used for field level recommendations, where fertilizer recommendation is derived from range value of nutrients such as low, medium and high.Overall, the results demonstrated practical predictive power of hyperspectral signatures for key soil nutrients under field conditions.The study confirms the viability of using HRS combined with machine learning for soil nutrient prediction, offering a non-invasive and scalable approach for soil health assessment. These findings advance precision agriculture by enabling efficient monitoring of soil fertility dynamics, there by supporting sustainable land management and improved crop productivity in diverse agro-ecological settings.

Graphical Abstract

This study presents a field-level application of Hyperspectral Remote Sensing (HRS) integrated with machine learning techniques to assess soil nutrient levels enabling rapid nutrient diagnostics compared to conventional methods. Conducted across two micro-watersheds representing humid tropical and tropical agro-ecologies in India, the research utilized spectral reflectance data (350–2500 nm) from 159 soil samples to predict Nitrogen (N), Available Phosphorus (P), and Exchangeable Potassium (K). A suite of machine learning algorithms including Support Vector Machine, Gaussian Process Regression, Random Forest, and Partial Least Squares Regression were evaluated in combination with eight spectral pre-processing techniques. The best-performing models were SVM for nitrogen (R2 = 0.49, RMSE = 20.10) and exchangeable potassium (R2 = 0.47, RMSE = 205.79), and GPR for phosphorus (R2 = 0.58, RMSE = 12.62). The findings support the potential of HRS and machine learning for rapid, scalable soil fertility assessment, offering valuable tools for precision agriculture and sustainable soil health monitoring in diverse Indian agro-ecological regions.