Hidden Arsenic Hotspots Mapped in Karst Farmland With AI and Geostatistics

Arsenic in the soil beneath our feet is invisible, odorless, and potentially dangerous, yet in the karst landscapes of southwestern China it follows patterns that scientists are only now beginning to decode. A new study of Zhijin County in Guizhou Province has combined classical geostatistics with modern machine learning to map where arsenic accumulates in agricultural topsoil and to identify which environmental factors best explain its distribution. The research, published in Environmental Monitoring and Assessment, analyzed soil from 144 sampling sites across a county where rugged carbonate terrain, mining activity, and intensive farming intersect in ways that make contamination risk unusually difficult to predict.

The stakes are high. Arsenic is a naturally occurring metalloid that becomes toxic to humans at relatively low exposures, and the World Health Organization recognizes it as a major public health concern. In Guizhou, the problem has a notorious history: chronic arsenic poisoning has previously been documented in villages where residents burned coal with exceptionally high arsenic content indoors. When arsenic sits in the topsoil of farmland, it can enter the food chain through crops, be inhaled as dust, or leach into the groundwater that threads through karst aquifers. Mapping its spatial distribution is therefore not an academic exercise but a prerequisite for protecting food safety in one of China’s most geologically distinctive agricultural regions.

The research team, led by Zhizhuo Liu and Lang Zhang with colleagues from Beijing Normal University, the Institute of Geophysical and Geochemical Exploration, and Tianjin Chengjian University, collected agricultural topsoil samples across Zhijin County and measured arsenic concentrations in the laboratory. The results revealed a striking range: values spanned from as low as 1.79 milligrams per kilogram of soil to more than 40 milligrams per kilogram, with a mean of 17.7 milligrams per kilogram. That spread matters, because it means some fields are relatively clean while others approach or exceed thresholds of concern, and the difference between them is not random. Understanding what drives that heterogeneity was the central question of the study.

Before any modeling could begin, the researchers had to confront a statistical challenge common in soil geochemistry: arsenic concentration data are typically skewed, with a long tail of high values that can distort conventional analyses. The team applied a Yeo-Johnson transformation, a flexible mathematical procedure that reshapes the distribution to reduce skewness. The transformation helped, but formal tests still rejected normality of the transformed data, with a p-value of 0.003. This detail is more than statistical housekeeping. It signals that arsenic in these soils is genuinely patchy and structured by underlying processes, rather than varying smoothly and randomly, which shaped the choice of methods that followed.

To characterize the spatial structure, the researchers turned to semivariograms, the workhorse tool of geostatistics. A semivariogram describes how similar soil values are as a function of the distance between sampling points, and fitting a mathematical model to it reveals the scale over which the variable behaves predictably. Among the candidate models, an exponential semivariogram fit best, achieving a coefficient of determination of 0.820. Two parameters stood out. The nugget-to-sill ratio of 0.627 indicated that a substantial fraction of the variation occurs at very short distances or within measurement error, a signature of strong local heterogeneity. The range of 37.02 kilometers showed that arsenic values remain spatially correlated over tens of kilometers, implying that broad regional forces, not just field-scale quirks, shape the pattern.

Using that fitted model, the team produced maps of arsenic across the county with ordinary kriging, a geostatistical interpolation technique that weights nearby observations according to the modeled spatial structure. After back-transforming the predictions to the original concentration scale, the maps revealed relatively high arsenic values concentrated mainly in the northern, northeastern, and central-eastern parts of Zhijin County. These hotspots provide exactly the kind of actionable intelligence that environmental agencies need: instead of monitoring uniformly, regulators can focus verification sampling and agricultural inspections on the zones where the geostatistical model suggests arsenic is most likely to be elevated.

Mapping where arsenic is high is only half the story; the other half is explaining why. For this, the researchers employed Geodetector, a statistical framework designed specifically to quantify how much of the spatial variation of a variable can be explained by a categorical environmental factor. The method computes a q statistic that measures explanatory power, and it can also test whether pairs of factors interact to explain more variation together than either does alone. In the Zhijin analysis, several factors showed statistically significant associations with arsenic patterns, with nominal p-values at or below 0.003: mercury concentration, soil organic carbon, distance to mining sites, elevation, distance to rivers, and lithology, the underlying rock type from which the soils developed.

The single most powerful factor was mercury, with a q statistic of 0.4241, the largest of any individual variable tested. The pairing of mercury and arsenic is geologically meaningful, because both elements are often enriched together by the same mineralization and coal-related geological processes that characterize parts of Guizhou. Even more intriguing was the interaction analysis: the combination of mercury and lithology produced the largest joint q statistic of the study, 0.5030, meaning that rock type and mercury together explained half of the spatial variation in soil arsenic. This suggests that arsenic accumulation in Zhijin is not driven by a single cause but by the interplay of geological substrate and geochemical processes that concentrate multiple potentially toxic elements simultaneously.

To push the explanatory analysis further, the team compared five candidate machine learning regressors for predicting arsenic concentrations from environmental covariates, including satellite-derived and terrain-based variables such as Sentinel-2 imagery, elevation from the Shuttle Radar Topography Mission, and river networks from OpenStreetMap. The winner was XGBRegressor, an implementation of extreme gradient boosting, a tree-based ensemble method that builds many sequential decision trees to capture nonlinear relationships. It achieved a pooled out-of-fold coefficient of determination of 0.4197, a root mean square error of 6.7793 milligrams per kilogram, and a mean absolute error of 4.5831 milligrams per kilogram, outperforming the other four models on every metric.

Crucially, the researchers did not treat the model as a black box. They applied SHAP, or SHapley Additive exPlanations, a technique borrowed from game theory that assigns each input variable a signed contribution to every individual prediction. At the global level, SHAP ranked mercury, elevation, and lithology as the three most influential predictors, and mercury and elevation remained among the top three across all five cross-validation folds, indicating that the ranking was stable rather than an artifact of a particular data split. The team then went a step further and produced sample-level SHAP maps, which visualize how each factor pushes arsenic predictions up or down at specific locations. These maps revealed heterogeneous signed contributions across the county, showing that the same factor can raise predicted arsenic in one area and lower it in another, a nuance that global averages completely obscure.

The authors are careful, and appropriately so, about what these findings do and do not prove. Predictive performance was moderate, with the best model explaining roughly 42 percent of the variance, and spatial transferability to other counties remains unverified. More fundamentally, the study distinguishes between spatial explanatory power, predictive contribution, and causality: the fact that mercury and lithology statistically explain arsenic patterns does not by itself establish a causal mechanism. What the results do support is targeted verification and zoned monitoring. For a region where karst hydrology can rapidly transport contaminants and where millions of people depend on local agriculture, that is a pragmatic and valuable outcome. The study also demonstrates a methodological template, pairing geostatistics, Geodetector, gradient boosting, and SHAP-based interpretability, that other regions with complex geology and legacy mining can adapt to trace the hidden geography of soil contamination.

Subject of Research: Spatial distribution and environmental drivers of arsenic in karst agricultural soils

Article Title: Spatial structure and multiscale environmental associations of arsenic in karst agricultural soils of Zhijin County, southwestern China

Article References: Spatial structure and multiscale environmental associations of arsenic in karst agricultural soils of Zhijin County, southwestern China. (n.d.). https://doi.org/10.1007/s10661-026-15960-4

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15960-4

Keywords: arsenic, karst soils, soil contamination, geostatistics, ordinary kriging, Geodetector, XGBoost, SHAP, heavy metals, Guizhou, agricultural soil, environmental monitoring