The research team, led by Ayoub Alayoub and Samir Mustapha of the Department of Mechanical Engineering, together with Reinde Alhouri Alhomsi and Darine A. Salam, set out to test whether hyperspectral imaging could serve as a fast screening tool for soil contamination across Lebanon’s diverse agricultural landscapes. Hyperspectral imaging works by capturing reflected light from a sample across hundreds of narrow, contiguous wavelength bands, spanning the visible, near-infrared, and shortwave infrared regions of the electromagnetic spectrum. Where the human eye registers only a single color of soil, the hyperspectral sensor records a full spectral fingerprint, a curve of reflectance values that encodes information about the minerals, organic matter, moisture, and particle surfaces that light has interacted with on its way back to the detector.

The logic behind using such spectra to detect heavy metals is subtle, because nickel and cadmium themselves are not strong absorbers of light in these wavelength ranges. Instead, the metals influence the soil’s spectral signature indirectly, through their associations with clay minerals, iron oxides, organic matter, and other matrix components that do absorb light in characteristic ways. This indirect relationship is precisely why machine learning is required: algorithms such as partial least squares regression, support vector regression, neural networks, and convolutional neural networks can detect statistical correlations between subtle spectral features and metal concentrations that no simple linear rule could capture. The technique has precedent, with earlier studies using reflectance spectroscopy to estimate contamination after mining accidents and to map heavy metals from airborne platforms, but the Lebanese study adds a rigorous comparison of model families across contrasting soil textures.

The researchers worked with three agricultural soils from Lebanon, classified as clayey, silty, and sandy, deliberately chosen to represent the range of physical matrices that a field-deployed sensing system would encounter. Before any modeling, the team contaminated the soils with known levels of nickel and cadmium and recorded their hyperspectral images. Even at this early stage, the three soil groups told different optical stories. They exhibited distinct absorption features, different overall reflectance intensities, and different spectral shapes, a reminder that soil is not a generic brown medium but a complex optical material whose signature depends on texture, mineralogy, and surface chemistry.

Four modeling approaches were then trained to translate spectra into contamination predictions. Partial least squares regression served as the classical chemometric baseline, projecting the high-dimensional spectral data onto a small number of latent variables that maximize covariance with the metal concentrations. Support vector regression, a kernel-based method that fits a function within a tolerance margin while ignoring small errors, represented the machine learning tier. On the deep learning side, the team built artificial neural networks and convolutional neural networks, the latter designed to automatically extract local spectral patterns from the raw reflectance curves much as CNNs extract edges and textures from photographs. Across the board, support vector regression and convolutional neural networks emerged as the strongest performers.

The numbers reveal both the promise and the texture-dependence of the approach. For nickel contamination, support vector regression achieved coefficients of determination of 0.850 in clayey soil, 0.830 in silty soil, and 0.790 in sandy soil. For cadmium, the same model performed even better, reaching 0.930, 0.931, and 0.945 in the clayey, silty, and sandy soils respectively. The convolutional neural network matched or exceeded these figures in several cases, with values of 0.899 for nickel and 0.944 for cadmium in silty soil, and 0.872 and 0.938 in sandy soil. Notably, the clayey soil yielded its best nickel and cadmium predictions from the neural network model, which reached 0.930 and 0.950 respectively. Cadmium was consistently predicted more reliably than nickel across all soil groups and models, a pattern the authors highlight as one of the study’s key findings.

Crucially, the team did not stop at internal cross-validation. They validated the support vector regression and convolutional neural network models against newly contaminated soil samples that the models had never seen, and the models held up, maintaining coefficients of determination above 0.790 across all tested soils. This external validation step matters enormously in spectroscopic modeling, where overfit models can look spectacular on training data and fail embarrassingly in the field. The result suggests that, within a given soil type, the spectral-to-metal relationship is stable enough to support practical screening applications.

But the study also delivers a sobering caveat. When the trained models were transferred across soil types, asked to predict contamination in a soil texture different from the one they were calibrated on, performance collapsed. The support vector regression models dropped to coefficients of determination ranging from 0.107 to 0.242, and the convolutional neural network models fared little better, ranging from 0.110 to 0.260. In other words, a model calibrated on clayey soil was essentially useless on sandy soil, and vice versa. The authors attribute this to the distinct physicochemical properties of each soil matrix, which reshape the spectral signature in ways the models cannot generalize across, and they conclude that soil-specific or matrix-informed calibration models are needed. They also acknowledge a limitation: because only three soils were investigated and detailed mineralogical characterization was not performed, the specific contribution of individual soil characteristics to the transferability failure could not be isolated.

For Lebanon, a country whose agricultural soils face pressure from industrial activity, unregulated waste practices, and decades of environmental stress, the implications are tangible. Conventional heavy metal analysis relies on techniques such as atomic absorption spectroscopy and inductively coupled plasma mass spectrometry, which require sample collection, acid digestion, and laboratory time. A hyperspectral workflow, by contrast, could in principle screen many samples rapidly once calibrated for the local soil type, allowing regulators to prioritize the handful of samples that truly need confirmatory wet chemistry. The study was supported by the US-Middle East Partnership Initiative Tomorrow’s Leadership Graduate program at the American University of Beirut, and all of the analyzed data are included in the published article and its supplementary files, with additional data available on reasonable request.

The broader lesson extends well beyond Lebanon’s borders. Around the world, soil monitoring programs struggle with the cost of chemical analysis, and the dream of cheap, fast, spectroscopy-based contamination screening has been building for two decades. This study moves that dream forward by showing that deep learning can extract cadmium and nickel signals from hyperspectral images with accuracy rivaling traditional methods, while simultaneously delivering a warning that the field must take seriously: soil is heterogeneous, and a model trained on one matrix cannot be blindly deployed on another. The path forward, the authors suggest, lies in building calibration frameworks that explicitly account for soil matrix information, whether by calibrating per soil class or by incorporating physicochemical descriptors into the models themselves. Until then, the combination of hyperspectral imaging and artificial intelligence stands as a powerful but soil-aware tool, one that could transform how the world’s farmland is screened for the heavy metals that threaten both crops and the people who eat them.

Subject of Research: Prediction of nickel and cadmium contamination in Lebanese agricultural soils using hyperspectral imaging and machine learning models

Article Title: Prediction of Ni and Cd contamination in agricultural soils in Lebanon using hyperspectral imaging and artificial intelligence

Article References: Prediction of Ni and Cd contamination in agricultural soils in Lebanon using hyperspectral imaging and artificial intelligence. (n.d.). https://doi.org/10.1007/s11356-026-38271-0

Image Credits: AI Generated

DOI: 10.1007/s11356-026-38271-0

Keywords: hyperspectral imaging, heavy metal contamination, nickel, cadmium, agricultural soils, machine learning, support vector regression, convolutional neural networks, soil spectroscopy, Lebanon, soil calibration, environmental monitoring

The physical basis of the technique rests on two fundamental processes. Molecular vibrations, the bending and stretching of chemical bonds when they absorb radiation matching their vibrational frequency, produce characteristic absorption bands that dominate the shortwave-infrared and mid-infrared regions. Electronic transitions between atomic orbital energy levels generate absorptions concentrated in the visible range. The key soil chromophores are well understood: organic matter darkens reflectance progressively as carbon content rises, iron-bearing minerals such as hematite and goethite leave distinctive fingerprints through charge transfers and crystal-field transitions, and clay minerals produce O-H stretching and Si-O-Al bending signatures at defined wavelengths. Because these features overlap and interact, no single wavelength maps cleanly onto a single soil property; instead, chemometric and machine-learning models must untangle the composite signal.

The review, led by Tushar Kumar of Maharana Pratap University of Agriculture and Technology in Udaipur, traces the field from the X-ray and microscopical studies of soil colloids in the 1930s through Condit’s foundational 1970 spectral library of American soils to the digital revolution of the 1980s and 1990s, when chart recorders gave way to instruments capturing thousands of data points per spectrum. Today the instrumental ecosystem spans laboratory benchtop analysers, portable handheld sensors, drone-mounted hyperspectral cameras, and satellite-borne spectrometers. Vis-NIR instruments operating from 350 to 2500 nanometres have been miniaturized into smartphone-compatible units, while the most sensitive mid-infrared systems still require liquid nitrogen cooling, confining them largely to well-resourced laboratories.

On the central question of which spectral region performs better, the evidence is nuanced. A comprehensive synthesis by Soriano-Disla and colleagues concluded that mid-infrared spectroscopy outperforms Vis-NIR for pH, soil organic carbon, cation exchange capacity, particle-size fractions, phosphorus, and potassium, because MIR wavelengths probe directly the functional groups that govern these properties. Yet biological properties such as microbial biomass carbon are more reliably predicted in the visible-near infrared, whose chromophores respond to the organic matrix in which biological activity is embedded. Practical considerations sharpen the trade-off: portable NIR units are dramatically cheaper and more field-deployable, making them the default for large surveys, while MIR’s theoretical advantage justifies its cost only for properties where the accuracy gain is significant.

Raw spectra, the review stresses, are rarely fit for modelling. Variable sample packing, particle-size differences, moisture, and instrument drift superimpose additive and multiplicative distortions on the true signal. Pre-processing is therefore integral, not optional. Standard steps include trimming noisy spectral extremes, converting reflectance to absorbance via the Beer-Lambert relationship, smoothing with Savitzky-Golay filters that preserve peak shape better than simple averaging, and scatter correction through the standard normal variate transformation, which centres each spectrum to zero mean and unit variance. Derivative transformations sharpen overlapping bands. Crucially, the review cites evidence that no universally optimal pre-processing sequence exists; the best combination depends on soil type, instrument, property, and model, and reported optima shift with spectral resolution rather than converging on a single setting.

Among prediction models, partial least squares regression remains the dependable baseline. Unlike principal component regression, which selects components by spectral variance rather than predictive relevance, PLSR extracts latent variables that maximize covariance with the target property, making it robust against the severe collinearity of spectral data. Where the spectral-property relationship is genuinely non-linear, alternatives earn their keep. Support vector regression achieved R² of 0.912 for total nitrogen in one study, and ensemble tree methods such as random forests outperformed PLSR for available sulphur. Deep-learning architectures including convolutional neural networks have produced strong results at global scale, but the review counsels caution: CNNs are data-hungry, overfit readily on the small local datasets typical of a single regional survey, and their reported gains over well-tuned classical baselines are frequently modest and inconsistent. The most defensible use of deep learning in data-scarce settings is transfer learning, pre-training on large external libraries and fine-tuning locally, an approach whose advantage for heterogeneous soils remains an open empirical question.

The property-by-property verdict falls into three reliability tiers. Clay and texture, soil organic carbon, moisture, cation exchange capacity, calcium, magnesium, and total nitrogen sit in a robust tier, predicted well wherever calibration is representative, with texture predictions across studies reaching R² values of 0.67 to 0.87 and organic carbon reaching R² of 0.89 in some Vis-NIR calibrations. A conditional tier, including pH, potassium, sorbed phosphorus, bulk density, and aggregate properties, is predictable only through proxies such as clay mineralogy and organic matter, so performance travels poorly between soils. A weak tier remains unreliable regardless of model: electrical conductivity in non-saline soils, extractable plant-available phosphorus, and DTPA-extractable micronutrients, whose trace concentrations are swamped by the spectra of dominant constituents. Potassium illustrates the contradictions sharply, with one study reporting R² of 0.41 and another 0.84, the difference hinging on whether soils contain abundant, spectrally active potassium-bearing clays such as illite.

The review is unsparing about the field’s untested assumptions. Laboratory reference measurements are treated as error-free ground truth despite carrying their own uncertainties; the chromophore framework, which assumes spectral features arise from a stable, approximately additive set of constituents, has never been subjected to systematic perturbation experiments; and model transferability is assumed whenever a global library calibrates predictions for a new region, even though Indian studies consistently report weaker performance than global benchmarks promise. Five frontier questions are identified, chief among them whether spectroscopy can deliver transferable predictions across the extreme variability of landscapes like Udaipur in Rajasthan, where soils range from red gneissic types at 1300 metres to deep black Vertisols on the plains.

For India, whose Soil Health Card programme has issued 24.74 crore cards through 8,272 laboratories, the realistic near-term role for spectroscopy is complementary rather than substitutive: rapid spectral screening for well-supported properties, reserving wet chemistry for extractable phosphorus, potassium in low-illite soils, and micronutrients. Global-to-local transfer learning offers a tested shortcut, with one study reporting lower error in 91 percent of comparisons after transfer, but local reference samples and independent validation remain indispensable. The review also flags data governance concerns, noting that digital soil mapping ties fertility records to identifiable landholdings and calling for consent, de-identification, and access controls as spectral monitoring scales.

The ultimate message is that diffuse reflectance spectroscopy has matured from a mineral identification tool into a credible multi-property platform underpinning digital soil mapping and precision agriculture, yet the field’s next advance will come not from another incremental model comparison but from rigorously testing its own foundations. Whether the technology can finally deliver affordable, trustworthy soil surveillance to the farmers who depend on it will be decided by the answers.

Subject of Research: Advances in diffuse reflectance spectroscopy for rapid soil fertility assessment

Article Title: A critical review of advances in diffuse reflectance spectroscopy for rapid soil fertility assessment

Article References: Kumar, T., Yadav, K. K., Meena, R. H., Lakhawat, S. S., & Jat, G. (2026). A critical review of advances in diffuse reflectance spectroscopy for rapid soil fertility assessment. Discover Soil, 3(1), Article 160. https://doi.org/10.1007/s44378-026-00314-w

Image Credits: AI Generated

DOI: 10.1007/s44378-026-00314-w

Keywords: diffuse reflectance spectroscopy, Vis-NIR, MIR, soil fertility, soil organic carbon, PLSR, machine learning, digital soil mapping, spectral pre-processing, soil testing, transfer learning, India