X-ray fingerprints turn river sediment into maps of erosion

Rivers carry an enormous amount of material downstream, from silt ground beneath glaciers to sand stripped from mountain slopes. That sediment can clog reservoirs, damage hydropower infrastructure, smelt riverbed habitats and ferry nutrients and contaminants far from where they originated. Yet for all its importance, one deceptively simple question has often been impossible to answer: where exactly does the sediment come from? A new study published in Earth Surface Dynamics by Fien De Doncker, Frédéric Herman, Bruno Belotti and Thierry Adatte offers a fast and inexpensive answer, using ordinary X-ray diffraction measurements to turn the mineralogy of suspended sediment into spatially explicit maps of erosion rates, even in landscapes hidden beneath ice.

The technique builds on a long tradition of sediment fingerprinting, the idea that every source area imprints a characteristic signature on the material it sheds. Researchers have traced sediment origins using geochemistry, radionuclides, isotopes, magnetic susceptibility, organic carbon, pollen, environmental DNA and even colour. The trouble is that many of these tracers are costly and slow to measure, which limits how widely they can be applied. X-ray diffraction, by contrast, is a workhorse technique: powdered samples are bombarded with X-rays, and constructive interference occurs only at specific angles determined by the spacing of atomic planes in the crystals present. The resulting diffraction pattern is a mineralogical fingerprint that can be collected quickly and cheaply for large numbers of samples.

The team’s workflow transforms raw diffractograms into usable fingerprints through a series of processing steps. Baseline drift is corrected and background noise removed with a peak-finding algorithm, and the patterns are aligned to a pure quartz reference to eliminate instrumental misalignments. Diagnostic two-theta windows for each target mineral are selected from reference spectra in the RRUFF database, and the total peak area within each window is averaged and normalised, yielding relative mineral concentrations on a scale from zero to one. Crucially, averaging peak areas across a window rather than relying on individual peaks mitigates the overlap that occurs when different minerals diffract at similar angles. The researchers focused on 21 primary rock-forming minerals, from quartz and feldspars to pyroxenes, amphiboles, micas and garnet, deliberately excluding the clay minerals produced by chemical weathering.

Turning fingerprints into erosion rates requires an inversion. The forward model is conceptually straightforward: each source area, typically a geological unit, carries its own mineral signature, erodes at some rate, and contributes sediment in proportion to that rate. The mineral concentrations measured in suspended sediment downstream are therefore a weighted average of the source signatures, with the weights being the erosion rates. Because erosion rates must be positive, the team solved the problem in log-space, making the mixing model non-linear. The inverse problem is underdetermined, since the number of tracers is far smaller than the number of pixels for which erosion rates are desired, so many different maps could explain the same data.

To tame this ambiguity, the authors adopted a Bayesian maximum a posteriori approach. They sought the most probable erosion map given the observed detrital data and prior knowledge encoded in a model covariance that favours spatially smooth solutions, with the smoothing distance and prior variance acting as tunable hyperparameters. Two iterative optimisation schemes were tested: steepest descent and a quasi-Newton method that uses curvature information to accelerate convergence. In synthetic forward-inverse tests, the quasi-Newton approach reconstructed a known erosion pattern with a localised hotspot far more accurately, while the steepest descent scheme proved unstable unless its gradient was normalised, which smoothed out sharp features and required many more iterations. The quasi-Newton method was therefore adopted for all subsequent experiments.

The sensitivity tests revealed which ingredients matter most. When the researchers progressively blended the source fingerprints together, making different lithologies mineralogically similar, the inversion deteriorated sharply: once a similarity metric called the average Jensen-Shannon distance fell below about 0.55, both errors and posterior uncertainty rose steeply, and in the most blended cases the inversion failed to converge at all. Uncertainty in the geological map, relevant in glaciated terrain where bedrock cannot be directly observed, was the second most influential factor. Reducing the number of tracer minerals had comparatively little effect, provided the most informative ones were retained: seven tracers produced essentially the same posterior solution as the full set of 21. Parameter sweeps showed the method is robust across a wide range of smoothing distances, standard deviations and step sizes, with convergence essentially complete after roughly 115 iterations.

The real-world validation took place in the Gornergletscher catchment in the Swiss Alps, an ideal testing ground because it combines multi-year suspended sediment records with strongly heterogeneous bedrock. The catchment drains lithologies from both oceanic and continental domains of the Pennine Alps, including serpentinites and eclogites of the Zermatt-Saas Fee ophiolites, calcareous mica schists and quartzites of the Penninic Mesozoic sediments, and the granites and garnet-mica schists of the Monte Rosa nappe. The XRD fingerprints derived from bedrock samples aligned well with established geological knowledge of these units, confirming that the cleaning and binning procedure preserves meaningful mineralogical information.

Two natural experiments put the method through its paces. In the first, the team exploited a natural analogue of a known-mixture test: sediment collected at the Gorner-Grenzgletscher confluence, where a subcatchment drains only part of the full study area. When the inversion was run using the geology of the entire catchment, it correctly attributed erosion only to lithologies actually present within the subcatchment, predicting rates close to zero for the Zermatt-Saas Fee sediment unit, which has no outcrops there. In the second, the XRD approach was benchmarked against zircon U-Pb age fingerprinting on the same detrital sample. The zircon-only inversion was highly unstable, producing implausible erosion rates above 8000 millimetres per year and high uncertainty, partly because two lithological units lacked zircon data altogether. The XRD-based solution was markedly more stable, and concatenating both datasets improved the distinguishability of source fingerprints and stabilised the posterior further, suggesting the two tracers are complementary rather than competing.

The authors are careful about the method’s limits. Grain-size effects, mineral fractionation during transport and post-depositional alteration can all bias fingerprinting, but several lines of evidence suggest these are modest in the Gorner setting: the suspended load is dominated by silt-sized glacial flour produced by physical comminution, only primary rock-forming minerals were used as tracers, and pump samples and depth-integrated samples yielded nearly identical mineralogies. The inversion also assumes that sediment is exported promptly rather than stored, so catchments with large sediment reservoirs are poor candidates. With those caveats, the recipe is appealingly simple: binned XRD data from sediments and source rocks, a map of potential source areas, and an estimate of total annual suspended sediment export. Because XRD is fast and affordable, the approach could open up provenance analysis to far more catchments, and the framework is flexible enough to accept other tracers, from zircon ages to geochemistry, or source maps based on land use rather than geology. For glacierised basins, where ice conceals the very landscapes doing the eroding, a cheap X-ray scan of river mud may now be enough to reveal where the ground is giving way.

Subject of Research: X-ray diffraction sediment fingerprinting and non-linear inversion to map erosion rates in glacierised catchments

Article Title: From XRD signal to erosion rate maps

Article References: From XRD signal to erosion rate maps. (n.d.). https://doi.org/10.5194/esurf-14-443-2026

Image Credits: AI Generated

DOI: 10.5194/esurf-14-443-2026

Keywords: sediment fingerprinting, X-ray diffraction, erosion rates, provenance analysis, Bayesian inversion, glaciers, Gornergletscher, suspended sediment, mineralogy, Swiss Alps, Earth Surface Dynamics, inverse problems