Hyperspectral determination of heavy metal content in soil is a rapid, non-destructive optical technique that estimates concentrations of elements such as cadmium, lead, arsenic, chromium, copper, and zinc from high-resolution reflectance spectra. By combining laboratory, field, UAV-borne, or satellite hyperspectral data with chemometric models calibrated against standard chemical analysis, this method enables large-scale, low-cost screening and mapping of soil contamination. This article explains the technical principles, sample requirements, target analytes, analytical workflow, applications, and frequently asked questions related to hyperspectral soil heavy metal determination.
Technical Overview
Hyperspectral measurement records soil reflectance across hundreds of contiguous, narrow spectral bands, typically covering the visible, near-infrared, and shortwave-infrared regions (VNIR–SWIR, approximately 350–2500 nm) with spectral sampling of 1–10 nm. Each measurement therefore produces a quasi-continuous spectrum in which subtle absorption features and curve-shape characteristics encode information about soil composition, in contrast to the single-value results of conventional wet-chemical analysis.
The core scientific principle is indirect estimation. Trace-level heavy metal cations such as Cd, Pb, Cu, and Zn exhibit no diagnostic absorption features of their own in the VNIR–SWIR range. Instead, these metals are adsorbed onto or occluded within spectrally active soil constituents—iron and manganese oxides, clay minerals, organic matter, and carbonates—whose spectral signatures can be measured. Iron oxides produce a strong charge-transfer absorption below approximately 600 nm and a crystal-field band near 900 nm; hydrous minerals show OH and water features near 1400 nm and 1900 nm; Al–OH absorption occurs near 2200 nm; carbonates absorb near 2340 nm; and organic matter modifies the overall spectral slope and albedo. Elements with a strong geochemical affinity for Fe–Mn oxides and organic matter, such as As, Pb, Cu, Zn, Ni, and Cr, are generally modeled most successfully.
Measurements can be performed at multiple scales: laboratory benchtop or contact-probe spectrometers, in-situ field-portable spectroradiometers, UAV-borne imaging spectrometers with centimeter-to-meter ground resolution, airborne campaigns, and satellite missions such as GF-5 (AHSI), ZY-1 02D, PRISMA, and EnMAP, which provide approximately 30 m resolution suitable for regional mapping.
From a regulatory perspective, hyperspectral results support rapid screening against soil quality benchmarks such as GB 15618-2018 (risk control standard for agricultural land) and GB 36600-2018 (risk control standard for development land), as well as soil monitoring programs designed according to HJ/T 166-2004. Confirmatory compliance analysis, however, is still based on reference chemical methods such as ISO 11047 and ISO 22036.
Testing Samples
Laboratory Prepared Soil Samples
Samples are air-dried at room temperature (below 40 °C to preserve volatile species such as mercury), gently crushed, homogenized, and sieved to 2 mm for spectral measurement, while companion sub-samples are milled to 0.149 mm for chemical digestion. Typically 100–200 g of soil is required per sample. Spectral acquisition is performed under standardized illumination against a low-reflectance background to eliminate the confounding effects of moisture and surface roughness.
In-Situ Field Soil Samples
Surface soil (0–20 cm plow layer, or deeper intervals for profile studies) is measured directly in the field with a full-range spectroradiometer and fiber-optic probe. Measurement points are GPS-tagged and arranged according to grid-sampling designs consistent with HJ/T 166-2004. Because field moisture, aggregation, and roughness influence spectra, in-situ data require dedicated field calibrations or moisture-correction procedures.
Archived Survey Samples
Legacy soil samples with documented or certified chemical concentration data can be re-measured spectrally to enlarge calibration libraries, particularly valuable for elements with limited fresh sampling coverage.
Image-Supporting Ground-Truth Samples
Soil samples collected synchronously with UAV, airborne, or satellite hyperspectral overpasses, with sampling plots sized to match image pixels, are used for upscaling point-based models and validating image-derived concentration maps.
Testing Items
The principal deliverables of hyperspectral soil heavy metal determination include:
- Estimated total contents of Cd, Pb, Cr, As, and Hg—the five priority heavy metals in soil environmental standards—plus Cu, Zn, Ni, Co, Mn, V, and Fe, and optionally Tl, Sb, and Se
- Co-estimated auxiliary properties that drive the spectral response, including soil organic matter, iron oxide content, clay mineral abundance, and moisture proxies
- Contamination screening and grading against the risk screening values of GB 15618-2018 and GB 36600-2018
- Spatial distribution maps, exceedance probability maps, and contamination hot-spot delineation across study areas
- Quantitative accuracy and uncertainty metrics, including coefficient of determination (R²), root mean square error (RMSE), ratio of performance to deviation (RPD), and bias
Testing Methods
Reference Chemical Analysis as Calibration Basis
All hyperspectral models must be calibrated and validated against rigorous chemical reference data. Standard approaches include aqua regia or total digestion followed by ICP-OES or ICP-MS determination in accordance with ISO 22036, HJ 766-2015, and HJ 680-2013; flame or graphite-furnace atomic absorption spectrometry for Pb and Cd in accordance with ISO 11047 and GB/T 17141-1997; and atomic fluorescence spectrometry for As and Hg in accordance with GB/T 22105.1-2008. These reference values anchor the spectral inversion models.
Spectral Data Acquisition
Reflectance spectra are collected with a full-range 350–2500 nm spectroradiometer, calibrated against a white reference panel, typically averaging 10–30 scans per measurement to improve signal-to-noise ratio. For imaging systems, radiometric calibration, atmospheric correction (for airborne and satellite data), and geometric orthorectification are applied before spectral inversion.
Spectral Preprocessing and Feature Extraction
- Savitzky–Golay smoothing to suppress high-frequency instrument noise
- First and second derivative transforms and logarithmic reflectance conversion to resolve overlapping absorption features
- Continuum removal to normalize absorption feature depth, area, and position
- Standard normal variate (SNV) and multiplicative scatter correction to mitigate particle-size and baseline effects
- Feature extraction using band-depth parameters, principal component analysis, and wavelet decomposition to reduce dimensionality
Modeling and Validation
Calibration datasets are typically split 70:30 or 2:1 between calibration and validation sets, supplemented by k-fold cross-validation. Widely used linear models include partial least squares regression (the methodological workhorse), principal component regression, and stepwise multiple linear regression. Nonlinear alternatives include support vector regression with radial basis kernels, random forest, and gradient boosting, while deep learning architectures such as one-dimensional convolutional neural networks are increasingly applied to full-range spectra. Model quality is interpreted using RPD thresholds: values of 2.0 or above indicate robust quantitative capability, 1.4–2.0 indicates acceptable screening accuracy, and values below 1.4 support qualitative discrimination only. Exceedance-classification outputs are additionally evaluated through confusion matrices and receiver operating characteristic analysis.
Upscaling and Spatial Mapping
Validated models are applied pixel-by-pixel to corrected hyperspectral imagery after masking non-soil surfaces such as vegetation and water. Image-derived estimates are frequently fused with ground measurements using geostatistical methods such as ordinary or regression kriging, producing continuous concentration surfaces and risk-zoning maps aligned with applicable GB screening values.
Applications
- Screening of agricultural land and delineation of exceedance zones near smelting, mining, and wastewater-irrigated areas
- Brownfield and industrial site investigation supporting land redevelopment and human-health risk assessment
- Regional and national soil pollution detailed surveys, providing dense spatial coverage at a fraction of conventional laboratory costs
- Dynamic monitoring of remediation projects, verifying the decline of contaminant concentrations over treatment cycles
- Precision agriculture, guiding variable-rate application of amendments based on mapped micronutrient and trace-element patterns
- Emergency environmental response, rapidly delineating contaminated footprints following accidental spills or atmospheric deposition events
In practice, hyperspectral determination functions as a front-end screening tool: by identifying hot spots in advance, it can reduce the number of confirmatory laboratory analyses required in a survey by a substantial margin while concentrating accredited third-party laboratory verification where it is most needed.
FAQ
Can hyperspectral spectra directly detect heavy metals in soil?
No. Heavy metals at trace concentrations have no diagnostic absorption features in the 350–2500 nm range. Estimation is indirect, relying on the correlation between metal concentrations and spectrally active carriers such as iron and manganese oxides, clay minerals, and organic matter. This is why site-specific calibration against chemical reference data is indispensable.
What accuracy can be expected?
Validation studies commonly report R² values of 0.6–0.9 for As, Cu, Zn, Pb, Ni, and Cr at regional scales. Cd and Hg are more difficult, often yielding lower accuracy because of their very low concentrations and weaker association with spectrally active constituents. Hyperspectral determination is therefore positioned as a quantitative screening and mapping tool rather than a substitute for regulatory-grade chemical analysis.
How many samples are needed to build a reliable calibration model?
As a practical guideline, at least 60–100 representative samples spanning the full concentration gradient of the study area are recommended, with larger libraries required for geochemically heterogeneous regions or multi-element regional models.
What sample preparation is required for laboratory spectral measurement?
Samples should be air-dried, gently crushed, sieved to 2 mm, and thoroughly homogenized before measurement under standardized illumination. For in-situ field measurements, soil moisture and surface roughness must be recorded and corrected for, since both strongly influence reflectance.
Which spectral region is most informative?
The full 350–2500 nm range is preferred. The visible and near-infrared region captures iron-oxide-related responses relevant to oxide-bound metals, while the shortwave-infrared region encodes clay mineral and organic matter features that proxy metals associated with these phases.
Can laboratory models be transferred directly to airborne or satellite imagery?
Only with caution. Differences in spatial resolution, soil moisture state, and atmospheric effects introduce scale-dependent bias. Successful transfer generally requires resampling spectra to sensor band configurations, harmonizing acquisition conditions, and recalibrating with image-synchronized ground-truth samples.