论文信息
论文解读
AI 辅助整理- 方法
- 光谱一阶/二阶导数与对数变换,SPA结合相关分析选特征波长,PLSR与SVM建模
- 发现
- CA+Log′(1/R)+PLSR模型最优,Rv²达0.8973、RPD为2.4664,线性模型即可
- 机会点
- 可探索非线性模型与多源数据融合,并验证跨区域迁移性以提升SOM反演泛化能力。
论文摘要
本研究基于可见光-近红外光谱与机器学习技术,对陕北横山区农田土壤有机质(SOM)的预测进行了研究。首先分析了土壤pH与SOM的统计分布特征,随后采集高光谱数据,并通过一阶导数(FD)、二阶导数(SD)及对数变换等数学变换改善光谱信息,结合连续投影算法与相关分析提取特征波长,最后构建并比较了偏最小二乘回归(PLSR)与支持向量机(SVM)模型。结果表明,SOM的变异系数超过50%,呈极高变异性,且SOM与光谱反射率呈显著负相关;一阶导数变换使相关系数由-0.55提升至-0.29。各模型中,CA+Log′(1/R)+PLSR组合模型表现最佳,验证决定系数(Rv²)为0.8973,相对分析误差(RPD)为2.4664,具有较优的精度与稳定性。本研究证实,在适当的导数或对数变换下,线性PLSR模型可实现土壤有机质的高精度反演,为区域精准农业与土地可持续利用提供技术支撑。Bangladesh J. Bot. 55(3): 741-747, 2026 (September) Special
全文
Bangladesh Journal of BotanyBangladesh J. Bot. 55(3): 741-747, 2026 (September) Special DOI: https://doi.org/10.3329/bjb.v55i30.94530
SOIL ORGANIC MATTER PREDICTION BASED ON VISIBLE-NEAR-
INFRARED SPECTROSCOPY AND MACHINE LEARNING
ZHENYU DONG1, NI WANG1*, JIANCANG XIE1 AND JUNGANG LUO1
State Key Laboratory of Eco-hydraulics in Northwest Arid Region, Xi’an University of Technology,
Xi’an, Shaanxi, 710048, China
Keywords: Inversion model, Logarithmic transformations, Machine learning
Abstract
This study investigates the prediction of soil organic matter (SOM) based on visible-near-infrared
spectroscopy and machine learning techniques in farmland soils of Hengshan district, northern Shaanxi. The
statistical distribution characteristics of soil pH and SOM were first analyzed. Hyperspectral data were then
collected, and spectral information was improved through mathematical transformations including first
derivative (FD), second derivative (SD), and logarithmic transformations. Feature wavelengths were
extracted using the successive projections algorithm combined with correlation analysis. Finally, PLSR
(Partial Least Squares Regression) and SVM (Support V ector Machine) were constructed and compared. The
results show that the coefficient of variation of SOM has exceeded 50%, indicating an extremely high
variability, and SOM showed a significant negative correlation with spectral reflectance. The first derivative
transformation increases the correlation coefficient from -0.55 to -0.29. Among the models, the
CA+Log′(1/R)+PLSR combined model performed the best, achieving a validation coefficient of
determination (Rv²) of 0.8973 and a residual predictive deviation (RPD) of 2.4664, demonstrating a superior
accuracy and stability. This study confirms that, with appropriate derivative or logarithmic transformations,
the linear PLSR model can achieve high-precision inversion of soil organic matter, providing technical
support for regional precision agriculture and sustainable land use.
Introduction
As a non-renewable resource with multiple strategic functions, arable land serves as a vital
foundation for ensuring national food security, maintaining ecological balance, and promoting
sustainable rural development (Du et al. 2022, Kong et al. 2023). The quality of arable land
directly determines agricultural production potential and has a profound impact on long-term,
stable grain output (Liu et al. 2024). Strengthening the development and management of arable
land quality is a key measure for implementing the precise arable land protection system. Rapid
industrialization and urbanization continue to put high-quality farmland at risk of irreversible
conversion, while the supply of developable reserve farmland has neared its ecological carrying
capacity, and the structural problem of replacing high-quality land with low-quality land has
become increasingly prominent (Shi et al. 2024). Under the combined influence of natural factors
and human activities, the overall trend of farmland quality degradation has not yet been curbed
(Qian et al. 2023). Issues such as soil pollution, declining soil fertility, and aging infrastructure are
intertwined, directly leading to the impairment of farmland’s productive and ecological functions
and severely constraining the sustainable utilization of farmland resource (Mao et al. 2025).
Reflectance data collected using ground-based spectrometers often contain noise; however,
applying various mathematical transformations can improve the signal-to-noise ratio and
significantly enhance the correlation between spectral data and soil properties. Studied the soil of
cultivated land in the Bortala Mongol Autonomous Prefecture of Xinjiang. Li et al. (2024) used
the ASD Field4 hyperspectral imager to acquire spectral data (350-2500 nm) and applied
transformations such as first-order differentiation (FD), the first derivative of the logarithm, the
*Author for correspondence: wangni@xaut.edu.cn. 1Institute of Water Resources and Hydro-electric Engineering,
Xi’an University of Technology, Xi’an, Shaanxi, 710048, China
742 DONG et al.
first derivative of the inverse, and multivariate scatter correction. Applied fi rst-order differential
transformation, second -order differential transformation, inverse logarithmic transformation,
inverse logarithmic first-order differential transformation, and envelope removal transformation to
preprocess spectral reflectance data acquired using an ASD hyperspectral imager (He et al. 2024).
They found that mathematical transformations of different orders enhanced spectral information to
some extent. They identified 600, 1300, 1600, 1900, and 2300 nm as characteristic bands for soil
total nitrogen. Based on this, the PLSR -GRNN hyperspectral inversion model established using
second-order differential transformation (SDR) achieved the highest inversion accuracy, with
model validation coefficients of determination R² and RMSE reaching 0.90 and 0.51, respectively,
enabling rapid and accurate inversion of total nitrogen content in calcareous soils within the study
area. They employed a random forest feature importance algorithm to select 40 feature bands with
importance values ranging from 0.21 to 0.47. with 25% of the feature bands in the visible region
and 75% in the mid - and near-infrared regions, primarily distributed around 510 -600, 820, 1000-
1100, 1400, and 1600 nm (Zhang et al . 2024). In a study of the Loess hilly and gully regions,
compared PLSR, MLR, and SVM and found that the SPA -SVM model, which extracts feature
wavelengths based on the continuous projection algorithm, exhibits good universality and high
inversion accuracy, making it suitable for experiments with small datasets (Zhang et al. 2018).
To address the aforementioned key bottlenecks, this study focuses on Hengshan County in the
agro-pastoral transition zone of Northern Shaanxi. Taking soil organic matter in cultivated land as
the subject of investigation, the study employs a comprehensive technical framework
encompassing data collection-spectral analysis-feature selection-model development-system
integration to carry out the following main tasks: (1) analyzing the distribution of soil organic
matter in the study area, (2) identifying the spectral bands characteristic of soil organic matter and
(3) developing a prediction model suitable for the study area.
Materials and Methods
Hengshan is located in northern Shaanxi, slightly west of the central part of Y ulin City, within
the coordinates 37°22′ to 38°14′ north latitude and 108°56′ to 110°02′ east longitude. The region
lies at the intersection of the southern edge of the Mu Us Desert and the northern tip of the Loess
Plateau, situated within the ecological transition zone where the Ordos Grassland meets the Loess
Hills. The terrain of the entire region slopes from the southwest and northwest toward the
northeast. Elevations in the southwest generally range from 1300 to 1530 meters, while in the river
valley areas of the northeast, they drop to between 900 and 1,200 meters. The region has a
temperate continental monsoon semi-arid climate, with an annual average temperature of 8.6°C
and marked interannual and intermonthly fluctuations. Extreme high temperatures can reach
38.4°C, while extreme lows can drop to -29°C.
In winter, the region is dominated by the Mongolian cold high-pressure system, and in
summer, it is influenced by continental low-pressure systems; the annual average atmospheric
pressure is approximately 890.5 millibars. Regarding precipitation, the annual average ranges from
382 to 398 millimeters, with a highly uneven seasonal distribution; the majority of the annual
rainfall is concentrated between July and September. The frost-free period lasts approximately 146
to 172 days, with significant diurnal temperature variations and relatively abundant solar and
thermal resources. The parent material of the soil in this area is primarily loess, with most soil
profiles belonging to the A-C type. The texture is predominantly sandy loam, with a strong lime
reaction throughout; the pH ranges from 8.0 to 8.8, indicating an alkaline reaction. Taking the
sandy loam of the plateau as an example, the topsoil layer is approximately 22 centimeters thick,
exhibiting a yellow-orange, clumpy structure that is loose and porous. It contains about 0.84%
organic matter and 0.057% total nitrogen, with available phosphorus at only 6 ppm, resulting in
SOIL ORGANIC MA TTER PREDICTION BASED ON VISIBLE-NEAR-INFRARED 743
relatively weak water and nutrient retention capacity. This soil pattern is highly coupled with the
distinctive features of the region’s topography and landforms, jointly shaping the ecologically
sensitive natural resource base of the Hengshan region, where agriculture and animal husbandry
coexist.
In this study, soil spectral data were collected from pretreated soil samples using an ASD
FieldSpec® 4 Hi-Res high-resolution spectroradiometer. The spectrometer has a wavelength range
of 350-2500 nm. Within the 350-1000 and 1000-2500 nm ranges, the sampling bandwidths were
1.4 and 1.1 nm, respectively, with spectral resolutions of 3 nm (at 700 nm) and 8 nm (at 1400/2100
nm). The spectral resampling interval was 1 nm, and the number of output bands was 2151.
Preprocessing includes classical methods such as SG smoothing, differential transformation,
MSC, and SNV , aimed at eliminating noise, scattering, and background interference. The SPXY
method was employed to simultaneously account for spectral information and nutrient distribution
characteristics, thereby enhancing the representativeness of the training samples. Feature band
selection methods include correlation analysis, applied in a multi-stage, progressive manner to
reduce data redundancy. Regarding modeling, PLSR and SVM machine learning models were
compared, and model performance was evaluated using metrics such as R², RMSE, RPD, and
RPIQ, thereby establishing a comprehensive evaluation system.
Results and Discussion
The descriptive statistics for soil nutrient content in Hengshan District are shown in Table 1
and Fig. 1. The soil pH ranged from 8.12 to 9.62, with a mean of 8.79 and a standard deviation of
0.19. The coefficient of variation was less than 10%, indicating low variability. The skewness of
soil pH was 0.28, and the kurtosis was 1.65; the data distribution exhibited a very slight right-
skewed, flat-peaked characteristic, which generally met the requirements of a normal distribution.
The SOM content ranged from 1.25 to 66.70 g/kg, with a mean of 6.12 g/kg and a standard
deviation of 4.89 g/kg. The coefficient of variation was greater than 50%, indicating extremely
high variability. The skewness of SOM was 7.32, and the kurtosis was 75.51, indicating that the
data distribution exhibited extremely significant, strongly right-skewed and peaked characteristics.
Table 1. Descriptive statistics of soil nutrient content in Hengshan district.
Type Maximum Minimum Mean Std. Dev. CV Skewness Kurtosis
pH 9.62 8.12 8.79 0.19 2.17(%) 0.28 1.65
SOM 66.70 1.25 6.12 4.89 79.93(%) 7.32 75.51
By comparing the shapes of the spectral curves and the characteristics of spectral absorption
peaks at different wavelengths, it was observed that the reflectance generally increases from the
visible to the short-wave infrared region (Fig. 2). Specifically, this can be divided into three
distinct regions: in the visible region (350-780 nm), the slope of the reflectance curve is relatively
steep, showing a rapid increase; Upon entering the near-infrared region (780-1300 nm), the rate of
increase slows significantly, and the curve flattens; whereas in the short-wave infrared region
(1300-2500 nm), the curve exhibits multiple sharp fluctuations due to second-harmonic and sum-
frequency absorption from molecular bonds in moisture and organic matter. In terms of
distribution patterns, for fixed wavelength bands, the reflectance of different organic matter
samples exhibits an ordered, stratified aggregation. Samples with higher soil organic matter
(SOM) content are concentrated in the low-reflectance region, while those with lower SOM
content are concentrated in the high-reflectance region, with minimal overlap. This indicates that
distinguishing soils with different organic matter levels based on spectral characteristics is highly
feasible. The range of soil spectral reflectance varies between 0.13 and 0.65. Among these, soils
744 DONG et al.
with SOM content in Grade III exhibit the lowest spectral reflectance, ranging from 0.13 to 0.53;
soils in Class IV had spectral reflectance ranging from 0.18 to 0.59; soils in Class V had spectral
reflectance ranging from 0.22 to 0.61; and soils in Class V had spectral reflectance ranging from
0.28 to 0.65. In the 350-430 nm range, the spectral curve showed a steep downward trend; in the
431-608 nm range, it rose sharply; Spectral reflectance increases rapidly in the 609-790 nm range
and rises slowly in the 791-2500 nm range. During this period, seven reflectance peaks appear at
617, 788, 1360, 1724, 2128, 2301, and 2375 nm, while five absorption valleys occur at 671, 890,
1052, 1170, and 1266 nm, and five relatively distinct absorption troughs at 430, 1411, 1919, 2203,
and 2352 nm. As dark-colored matter, organic matter strongly absorbs incident light, thereby
reducing soil reflectance. Overall, higher soil organic matter content corresponds to lower spectral
reflectance, indicating a negative correlation between soil organic matter content and spectral
reflectance.
Fig. 1. Statistical distribution map of soil nutrient content in Hengshan district.
Fig. 2. Average reflectance curves for varied SOM content.
SOIL ORGANIC MA TTER PREDICTION BASED ON VISIBLE-NEAR-INFRARED 745
With regard to the raw spectral reflectance of soil, a negative correlation was observed in the
350-2500 nm wavelength range, peaking at 626 nm, with a correlation coefficient of only -0.29
(Fig. 3a). At a significance level of p <0.05, significant correlations were observed in the 510-554,
796-847, and 2465-2500 nm ranges, with 130 features and a maximum correlation coefficient of -
0.19. At a significance level of p <0.01, a highly significant correlation was observed in the 555-
797 nm range, with 242 features and a maximum correlation coefficient of -0.29. After applying a
first-order differential transformation to the raw spectra, at a significance level of p <0.05,
significant correlations were observed in the 427-502, 598-611, 634-661, 698-740, 1054-1069,
1102-1116, 1665-1670, 1910-1921, and 2353-2384 nm showed significant correlations at a
significance level of p <0.05, with 232 features and a maximum correlation coefficient of -0.39.
(Fig. 3b). At a significance level of p <0.01, highly significant correlations were observed in the
503-597, 662-697, and 741-994 nm ranges, with 385 features and a maximum correlation
coefficient of -0.55. After applying a second-order differential transformation to the raw
reflectance and performing a correlation analysis with SOM, it was found that at a significance
level of p <0.05, significant correlations were observed in the 416 nm, 424-429, 443-446, 450-
451, 459-671, 475-478, 507, 520-525, 555-558, 778-802, 829-831, 930-936, 941-943, 954-970,
1735-1736, 1923-1924, 1943-1944, 2219-2339, and 2487-2492 nm, with a total of 90 features
(Fig. 3c). At a significance level of p <0.01, significant correlations were observed in the 441-442,
447-463, 473-506, 526-669, 779-801, 832-969, 1925-1942, 2120-2121, and 2331-2338 nm, with
320 features and a maximum correlation coefficient of -0.47.
Fig. 3. Correlations of spectral reflectance and its derivative transformations with SOM across different wavelengths.
Fig. 4. Predict scatterplots based on SOM content of CA.
746 DONG et al.
The prediction model based on raw spectral reflectance (R) exhibited the lowest accuracy and
poorest stability, whereas models based on transformations such as FD, Log′(1/R), or Log(1/R)
performed better (Table 2). In the PLSR models, the Rc² values for models based on the six
spectral transformations ranged from 0.2961 to 0.8586, and the RMSEC values ranged from
1.2835 to 4.8708; Based on a comparison of model accuracy validation metrics, the accuracy of
the six spectral models, ranked from highest to lowest, was FD > Log′(1/R) > SD > Log″(1/R) >
Log(1/R) > R. The Rv², RMSEP , RPD, and RPIQ values for FD were 0.8973, 1.1587, 2.4664, and
3.1320, respectively. In the SVM models, the Rc² values for the models based on the six spectral
transformations ranged from 0.2395 to 0.8594, and the RMSEC values ranged from 1.2793 to
5.6294; The models were ranked in descending order of accuracy as follows: Log′(1/R) > FD >
SD > Log″(1/R) > Log(1/R) > R. Among them, Log′(1/R), which exhibited the highest accuracy,
had Rv², RMSEP , RPD, and RPIQ values of 0.8983, 1.2912, 2.2171, and 2.8154, respectively.
Comparing the two machine learning models above, the CA + Log′(1/R) + PLSR model exhibits
the highest accuracy and strongest stability, with Rc² of 0.8586, RMSEC of 1.2835, Rv² of 0.8973,
RMSEP of 1.1587, RPD of 2.4664, and RPIQ of 3.1320. SVM ranks second (with Rv² = 0.8983
under the FD transformation). The results indicate that after appropriate derivative or logarithmic
transformations, the linear PLSR model also achieves good predictive accuracy and stability
(Fig. 4).
Table 2. Comparison of CA-based soil organic matter prediction accuracy.
Modeling Type Calibration Validation
Rc2 RMSEC Rv2 RMSEP RPD RPIQ
CA
PLSR
R 0.2961 4.8708 0.3159 4.3313 0.6492 0.8253
FD 0.8586 1.2835 0.8973 1.1587 2.4664 3.1320
SD 0.7099 2.0218 0.7297 1.7979 1.5657 1.9883
LOG 0.6618 2.2610 0.6816 2.0105 1.4001 1.7780
FDL 0.7379 1.8850 0.7577 1.6762 1.6794 2.1326
SDL 0.6800 2.1671 0.6998 1.9271 1.459 1.8550
SVM
R 0.2395 5.6294 0.3071 5.0906 0.5617 0.7141
FD 0.8594 1.2793 0.8718 1.1640 2.4745 3.1424
SD 0.7526 1.8134 0.8202 1.6398 1.7457 2.2169
LOG 0.6841 2.1492 0.7517 1.9436 1.4729 1.8704
FDL 0.8307 1.4278 0.8983 1.2912 2.2171 2.8154
SDL 0.7304 1.9193 0.798 1.7356 1.6474 2.0945
Soil pH in Hengshan District exhibits low variability (mean 8.79, CV <10%), generally
conforming to a normal distribution. Soil organic matter shows extremely high variability (mean
6.12 g/kg, CV >50%) with a strongly right-skewed, leptokurtic distribution. Spectrally, reflectance
generally increases from visible to short-wave infrared regions. High-SOM samples are
concentrated in low-reflectance regions, while low-SOM samples occupy high-reflectance regions,
indicating a negative correlation. After first derivative transformation, the maximum correlation
coefficient increases from -0.29 to -0.55, demonstrating that derivative transformations effectively
enhance spectral features. Among the models compared, the CA+Log′(1/R)+PLSR combined
model achieves the highest accuracy and stability (Rv²=0.8973, RPD=2.4664), followed by the
SVM model. In conclusion, with appropriate derivative or logarithmic transformations, the linear
PLSR model can achieve satisfactory predictive accuracy and stability.
SOIL ORGANIC MA TTER PREDICTION BASED ON VISIBLE-NEAR-INFRARED 747
Acknowledgement
This work was supported by the National Natural Science Foundation of China (Grant
No.51979221), the Research Fund of the State Key Laboratory of Eco-hydraulics in Northwest
Arid Region, Xi’an University of Technology (Grant No.2019KJCXTD-5), Doctoral Dissertation
Innovation Fund of Xi'an University of Technology (104-252072305) and Shaanxi Water
Conservancy Science and Technology Program Project (Grant No. 2025slkj-10).
References
Du GM, Liang CA and Li NN 2022. Reflections on establishing national strategic reserve system of
cultivated land. Agric. Econ. Manag. 3: 20-27.
He W , Li YQ, Yu L, Wang JY , Ni LK and Li N 2024. Hyperspectral inversion of total nitrogen content in
calcareous soil in karst areas. Carsologica Sin. 43(5): 1047-1056.
Kong XB, Chen WG and Dang YX 2023. Current situation, challenges and transformation of cultivated land
protection in China. J. Soc. Sci. Hunan Norm. Univ. 52(5): 31-41.
Li JQ, Feng YH, Chen SH, Wang ZA, Liu P, Liang ZY , Sun FF, Chen R and Geng QL 2024. Estimation of
soil organic matter and total nitrogen based on hyperspectral technology. Xinjiang Agric. Sci. 61(10):
2491-2499.
Liu R and Fang YW 2024. Practice and improvement of the arable land requisition-compensation balance
system. Mod. Gov. 7: 40-50.
Mao W , Su S, Chen M, Liu XL, Xu XY , Zhang YQ, Li TP , Zhao HT, Zhang JD and Wang HY 2025.
Conceptual analysis and evaluation indicator system development of cultivated land quality. J. Plant
Nutr. Fertil. 31(11): 2133-2155.
Qian JC, Shi N, Zhao HF, Wu KN and Feng Z 2023. Theoretical analysis and research framework of
resilience management of cultivated land in China: From single-objective trade-offs to multi-objective
synergy. China Land Sci. 37(3): 38-47.
Shi YC, Jin XB, Han B, Liang XY , Wang SL and Sun R 2024. Progress and future directions in cultivated
land assessment research: A focus on changes in resource cognition. J. China Agric. Univ. 29(8): 237-
250.
Zhang C, Liu YM, Sun YN, Wang L and Liu JH 2018. Hyperspectral prediction model of soil nutrient
content in the loess hilly-gully region, China. J. Ecol. 29(9): 2835-2842.
Zhang XQ, Ma SX, Li ZW , Zheng DC and Song HY 2024. Prediction of total nitrogen content in brown soil
based on hyperspectral and combined prediction model. Spectrosc. Spectr. Anal. 44(8): 2310-2317.
(Manuscript received on 02 June 2026; revised on 26 August 2026)
论文原文与出版信息
Bangladesh Journal of Botany