Soil Carbon Mapping Goes Exponential as AI and Satellites Reshape a Research Field

Soils hold more carbon than the atmosphere and all vegetation combined, which makes soil organic carbon, or SOC, one of the most consequential variables in the global climate system. A new exploratory bibliometric analysis covering 1,561 peer-reviewed articles published between 2005 and 2025 shows that the science of predicting and mapping this carbon has undergone a dramatic transformation, growing at an annual rate of 28.68 percent and shifting decisively from traditional statistics toward artificial intelligence. The study, based on a structured Scopus search combining terms for soil organic carbon, prediction, geospatial technologies and machine learning, offers the most complete quantitative picture yet of how this research frontier has evolved, who drives it, and where it is geographically concentrated.

The numbers behind the growth are striking. The 1,561 documents in the final dataset were written by 7,107 authors, with only 23 single-authored papers in the entire corpus. The average document carries 6.09 co-authors, and 40.49 percent of publications involve international collaboration, pointing to a field built on teamwork rather than isolated effort. The average paper has accumulated 43.67 citations, and the average document age of just 4.05 years signals a young, fast-moving domain. Beneath the publications lie 225,682 cited references and 4,790 author keywords, a measure of the field’s thematic breadth and its deep roots in earlier soil science literature.

The temporal trajectory tells a story of three distinct phases. From 2005 to 2012, output remained very low, a foundational stage in which the topic was emerging but not yet established. Between 2013 and 2017, publications rose gradually as scholarly interest consolidated. Then came the inflection point: after 2018, production accelerated sharply, with a particularly strong surge between 2020 and 2024. Annual output peaked in 2024 at more than 300 articles, with 2025 showing only slight stabilization, a figure the authors caution should be read carefully because Scopus may index the final year’s publications retroactively. The authors attribute the post-2018 acceleration to concrete technological developments: the widespread availability of satellite observations, the expansion of open geospatial platforms and cloud-based processing, and advances in machine-learning and deep-learning algorithms capable of exploiting large, heterogeneous datasets.

Underlying the growth is a genuine scientific problem. SOC varies intensely across space and time through complex interactions among climate, topography, vegetation, parent material and land use. Traditional soil surveys, while reliable locally, are labor-intensive and cannot capture fine-scale heterogeneity across large regions. Digital soil mapping, which combines environmental covariates with predictive modeling, emerged as the response, and it has been supercharged by satellite platforms such as Landsat and Sentinel, whose multispectral imagery tracks vegetation dynamics and land surface characteristics tied to SOC variability. Vegetation indices like the Normalized Difference Vegetation Index and red-edge metrics have shown strong relationships with SOC by indirectly capturing biomass productivity and soil-vegetation interactions.

The algorithmic side has evolved just as rapidly. Random Forest, Support Vector Machines, XGBoost, Artificial Neural Networks and Deep Neural Networks now dominate the modeling landscape, valued for their ability to capture nonlinear, high-dimensional relationships among environmental variables. Comparative studies cited in the analysis indicate that ensemble methods, particularly Random Forest and XGBoost, frequently outperform traditional regression approaches in SOC prediction tasks. More recently, architectures that integrate spatial and temporal information, including Convolutional Neural Networks and hybrid CNN-LSTM models, have shown improved performance in complex vegetation and heterogeneous landscapes. The keyword analysis traces this shift precisely: between 2010 and 2014 research centered on regression trees and digital terrain analysis; from 2015 to 2018 the focus moved to data mining, partial least squares regression and artificial intelligence; and from 2019 onward machine learning, remote sensing and digital mapping surged, with peak term frequencies between 2021 and 2024.

Machine learning is the single most frequent keyword in the dataset, appearing 622 times, followed by remote sensing at 535 and organic carbon at 392. A co-occurrence network analysis reveals a dense, scale-free structure organized into three clusters: a red cluster centered on machine learning and remote sensing with terms like Random Forest, artificial neural network and satellite imagery; a green cluster around soils, organic carbon, soil surveys and digital soil mapping; and a blue cluster of environmental and contextual terms including environmental monitoring, land use and climate change. Bridging terms such as Random Forest, prediction and environmental monitoring connect the computational and soil-focused clusters, indicating a structurally integrated domain in which data-driven techniques are systematically applied to soil and environmental problems.

The geography of the field is starkly uneven. China leads in both productivity and citation impact, with 16,602 total citations, substantially exceeding all other countries, and it experienced exponential growth after 2018 to surpass the early leader, the United States, in annual output. The United States follows with 8,006 citations, ahead of Australia at 4,746, Germany at 4,286 and France at 3,873. Iran, Canada, the Netherlands, Italy and India also contribute significantly, while Brazil, Germany and India show marked recent increases. Collaboration maps reveal a hub-and-spoke global structure anchored by China and the United States, with intensive transcontinental links to Europe, Australia, Canada and Asia. At the author level, the network shows a scale-free co-authorship pattern in which a small core of highly connected researchers, led by Shi Zhou with 45 publications and Chen Songchao with 39, acts as a collaboration hub.

Journal output is equally concentrated. Geoderma dominates with 405 documents, followed by Science of the Total Environment with 315, Environmental Monitoring and Assessment with 174, and Remote Sensing of Environment with 170. Citation patterns reinforce the dual identity of the field: the most globally cited paper is a 2020 article in Remote Sensing of Environment with 1,550 citations, while the most locally cited work is Rossel’s 2010 Geoderma paper on soil spectroscopy and digital soil mapping, with 98 local citations. Most tellingly, Breiman’s 2001 work, the foundation of Random Forest, leads cited references with 211 local citations, evidence of how deeply machine learning methodology has penetrated the soil science canon.

The study’s authors are careful about what their indicators do and do not show. Bibliometric indicators describe publication and citation patterns, not scientific quality or practical impact, and field-normalized citation impact was not computed. The dataset relies on a single database and English-language journal articles, and statements about agricultural applications such as precision agriculture, carbon accounting and climate-smart land management are interpretations informed by the literature rather than outcomes measured by the analysis. The geographic imbalance, measured here as publication output by citations and corresponding-author country, does not by itself establish its causes; the authors propose, as hypotheses rather than findings, that unequal access to funding, field measurements, harmonized soil databases, computing resources and research networks may contribute.

Still, the implications reach well beyond the numbers. Because research output is concentrated in a few countries, the evidence base on SOC prediction models may reflect the soils, climates and data conditions of those places, and model transferability to underrepresented regions cannot be assumed without regional calibration and validation. The authors argue for open and interoperable soil and geospatial datasets, regional benchmark datasets, field-validation networks, capacity-building in GeoAI, and stronger South-South and North-South collaboration to improve representativeness across Africa, South Asia and Latin America. They also call for explainable artificial intelligence and rigorous field validation to turn methodological advances into dependable decision-support systems. What their analysis documents, in the end, is a field that has moved from niche to mainstream in two decades, converging on an integrated framework that extends digital soil mapping into a data-intensive, adaptive paradigm capable of supporting soil health restoration, carbon sequestration and evidence-based environmental decision-making, provided the benefits of that convergence are shared beyond the handful of countries currently driving it.

Subject of Research: Bibliometric analysis of soil organic carbon prediction research using geospatial technologies and artificial intelligence

Article Title: Trends in soil organic carbon prediction research using geospatial technologies and artificial intelligence: an exploratory bibliometric analysis (2005-2025)

Article References: Khaddi, E., Zaher, H., & Naim, H. (2026). Trends in soil organic carbon prediction research using geospatial technologies and artificial intelligence: an exploratory bibliometric analysis (2005-2025). Environmental and Sustainability Indicators, 32, Article 101585. https://doi.org/10.1016/j.indic.2026.101585

Image Credits: AI Generated

DOI: 10.1016/j.indic.2026.101585

Keywords: soil organic carbon, bibliometric analysis, machine learning, remote sensing, digital soil mapping, GeoAI, artificial intelligence, geospatial technologies, carbon mapping, climate mitigation, precision agriculture, Scopus