Sebastian Gutierrez from Aarhus University and David Robinson from The Centre for Ecology and Hydrology (CEH), explore the new ways of understanding the complexity of soils across continent’s like Europe.
Why do we need a new way to map soils?
Soils are fundamental to food production, climate regulation, and biodiversity—but they remain surprisingly difficult to map in detail. Traditional soil monitoring relies on field sampling and laboratory analysis, which are accurate but can be slow to analyse and sparse in coverage. This creates a fundamental problem: how do we understand soil variation across entire landscapes when we only measure a tiny fraction of it?
New European ambitions, such as the Soil Monitoring Law, require consistent and scalable methods to assess soil health across countries. However, conventional approaches often struggle to capture the complexity of soils, which vary not only in their chemical properties (such as carbon or pH), but also in their mineral composition, hydrology, and history of formation.
A key limitation is that soils are typically mapped property-by-property, even though these properties are tightly interconnected. What is missing is a way to capture soil as an integrated system, rather than a collection of separate measurements.
A new approach: compressing soil “fingerprints” into pedo‑spectral units
Recent advances in soil spectroscopy and machine learning provide a way forward. Visible and near-infrared (vis–NIR) spectroscopy measures how soils interact with light, producing a spectral “fingerprint” that reflects organic matter, minerals, moisture, and chemistry in a single signal. These spectra are information-rich but complex.
In earlier studies, thousands of soil spectra were compressed into a small number of latent variables—mathematical representations that summarise the dominant patterns in soil composition. These latent variables were then mapped across Denmark at 10 m resolution using environmental data such as climate, topography, and parent material. The result is a set of spatial layers that capture soil variation as an integrated signal, rather than as individual properties. For further details, the reader is referred to this blogpost.
In the present work, we clustered the latent variable layers to define spatial zones with similar spectral characteristics. We propose calling these zones pedo‑spectral units—areas of land that share a common soil “signature” derived from spectral data. Unlike traditional soil classes, pedo‑spectral units do not rely solely on predefined categories but emerge directly from the data, reflecting real patterns of soil formation and landscape processes.
Figure 1 provides a conceptual workflow showing how the continuous latent variables are translated into discrete spatial zones with similar spectral characteristics (pedo-spectral units) through hierarchical clustering techniques.
What are the benefits—and the challenges?
This approach offers several important advantages. First, it provides high-resolution, high throughput mapping variables, leveraging spectroscopy to augment limited laboratory data. Second, it captures multi-dimensional soil information in a compact form, linking mineralogy, organic matter, and environmental drivers. Third, it aligns naturally with digital soil mapping and machine learning frameworks, enabling consistent mapping across large regions.
Importantly, pedo‑spectral units can serve as an integrative basis for soil health assessment. Instead of monitoring many individual variables everywhere, monitoring could be stratified across these units, improving efficiency while retaining representativeness.
However, there are also challenges. The latent variables generated through techniques such as autoencoders are not always easy to interpret—they can act as “black boxes”. There is a trade-off between predictive power and explainability. Soil change is not easily attributed, so measurements are best seen currently as a supporting covariate. In addition, the approach depends on the availability of high-quality spectral data and environmental covariates. Standardisation across countries and datasets will be essential if this approach is to support policy.
From research to practice: implications for monitoring and AI4SoilHealth
Pedo‑spectral units offer a practical pathway for next-generation soil monitoring programmes. They could be used to define Soil Monitoring Units under the EU Soil Monitoring Law, providing a scientifically grounded and data-driven framework for stratification and reporting. By integrating spectral data with climate and landscape information, they support consistent, scalable, and policy-relevant soil assessments.
Within AI4SoilHealth, this work directly contributes to the development of new indicators and data layers for soil health assessment. The latent-variable maps represent a novel environmental layer that can be used to predict physical, chemical, and even biological soil properties, including carbon storage, nutrient cycling, and microbial activity. This aligns with the project’s goal of combining advanced sensing, data integration, and AI to support decision-making.
Looking ahead, the potential is significant. Pedo‑spectral units could underpin digital soil twins, support precision agriculture, and enable real-time monitoring using field sensors or drones. By linking spectroscopy with other technologies—such as mid-infrared sensing, X-ray fluorescence or geophysical sensors—even richer soil information could be captured in future systems.
Ultimately, this approach shifts how we think about soils: from static maps of individual properties to dynamic, data-driven representations of soil systems. Pedo‑spectral units may provide the bridge between scientific complexity and practical soil management—helping us monitor, understand, and protect soils at the scale that society now demands.
