Spectral reflectance technology has emerged as a promising tool for estimating soil properties while offering a rapid, non-destructive, and cost-effective alternative to traditional methods. Free iron is an important soil property, and it reflects the occurrence and evolution of soil. An accurate and efficient determination of soil free iron content is important. To evaluate the feasibility of using spectral reflectance and machine learning methods to estimate soil free iron content, we collected the spectral reflectance of 540 soil samples from 135 locations. We looked at the original spectrum and transforms such as the first derivative (FD), standard normal variate (SNV), and continuum removed (CR). The full spectrum, correlated spectrum, and principal components from principal component analysis (PCA) were considered as model variable selection. We used machine learning algorithms, such as partial least squares (PLS), support vector machine (SVM), random forest (RF), and deep neural network (DNN) algorithms for model construction. We found that FD was a more efficient transform than the original, SNV and CR spectra. The average R2, RMSE, and RRMSE when using the FD transform for training were 0.797, 5.550 g/kg, and 25.1%, respectively. In testing models, CR had a higher accuracy than the other transforms and its R2, RMSE, and RRMSE were 0.644, 7.140 g/kg, and 32.7%. Variable selection based on PCA projection improved model accuracy compared to using full and correlated spectra. The average model R2, RMSE, and RRMSE following PCA were 0.821, 5.260 g/kg, and 23.9% in training and 0.692, 6.744 g/kg, and 30.9% in testing, which had a higher R2 and lower RMSE and RRMSE than when using the full and correlated spectra. Over-fitting may have occurred in our study when employing the CR transform and RF algorithm. Their models had high accuracy in training and low accuracy in testing. The model R2 using the DNN showed better performance than those using the PLS and SVM algorithm, but the DNN showed poorer performance in RMSE and RRMSE than that of the model utilizing the SVM and PLS algorithm. The best combination of spectral transform, variable selection, and modeling method was FD + PCA + SVM. The R2, RMSE and RRMSE of this combination were 0.876, 4.085 g/kg and 18.8%, respectively, in training; these reached 0.803, 5.203 g/kg and 23.9%, respectively, in testing. Hence, our study showed spectral reflectance and machine learning could be used to estimate soil free iron content rapidly, non-destructively, and economically. Given these valuable findings, the present study benefits soil properties mapping, crop nutrient management and improving environmental issues.

Estimation of soil free Iron content using spectral reflectance and machine learning algorithms

期刊论文
出版模式：

开放获取
发表日期：

2025年7月4日

查看全文

Wanzhu Ma,
Hongkui Zhou,
Hao Hu,
Zhiqing Zhuo,
Kangying Zhu,
Guangzhi Zhang

摘要

Spectral reflectance technology has emerged as a promising tool for estimating soil properties while offering a rapid, non-destructive, and cost-effective alternative to traditional methods. Free iron is an important soil property, and it reflects the occurrence and evolution of soil. An accurate and efficient determination of soil free iron content is important. To evaluate the feasibility of using spectral reflectance and machine learning methods to estimate soil free iron content, we collected the spectral reflectance of 540 soil samples from 135 locations. We looked at the original spectrum and transforms such as the first derivative (FD), standard normal variate (SNV), and continuum removed (CR). The full spectrum, correlated spectrum, and principal components from principal component analysis (PCA) were considered as model variable selection. We used machine learning algorithms, such as partial least squares (PLS), support vector machine (SVM), random forest (RF), and deep neural network (DNN) algorithms for model construction. We found that FD was a more efficient transform than the original, SNV and CR spectra. The average R², RMSE, and RRMSE when using the FD transform for training were 0.797, 5.550 g/kg, and 25.1%, respectively. In testing models, CR had a higher accuracy than the other transforms and its R², RMSE, and RRMSE were 0.644, 7.140 g/kg, and 32.7%. Variable selection based on PCA projection improved model accuracy compared to using full and correlated spectra. The average model R², RMSE, and RRMSE following PCA were 0.821, 5.260 g/kg, and 23.9% in training and 0.692, 6.744 g/kg, and 30.9% in testing, which had a higher R² and lower RMSE and RRMSE than when using the full and correlated spectra. Over-fitting may have occurred in our study when employing the CR transform and RF algorithm. Their models had high accuracy in training and low accuracy in testing. The model R² using the DNN showed better performance than those using the PLS and SVM algorithm, but the DNN showed poorer performance in RMSE and RRMSE than that of the model utilizing the SVM and PLS algorithm. The best combination of spectral transform, variable selection, and modeling method was FD + PCA + SVM. The R², RMSE and RRMSE of this combination were 0.876, 4.085 g/kg and 18.8%, respectively, in training; these reached 0.803, 5.203 g/kg and 23.9%, respectively, in testing. Hence, our study showed spectral reflectance and machine learning could be used to estimate soil free iron content rapidly, non-destructively, and economically. Given these valuable findings, the present study benefits soil properties mapping, crop nutrient management and improving environmental issues.