Geochemical-integrated machine learning approach predicts the distribution of cadmium speciation in European and Chinese topsoils

ML-based framework

Our geochemical-integrated ML framework utilized regional soil surveys, crop uptake collection and MSM modeling to predict the Cd associated with soil interfaces and the dissolved fraction of nonindustrial soil Cd at the continental scale (Fig. 1). The overall structure of the work can be broadly divided into five sections (details on the ML framework and validation procedures are described in “Methods”): (1) To predict the amorphous ferrihydrite content (represented by hydrous ferric oxides and abbreviated as HFO) and Cd speciation distribution (abbreviated as dist) in the EU and China, four datasets (EU_HFO, EU_dist, CN_HFO and CN_dist) were compiled; (2) ML algorithms were developed on the EU_HFO and CN_HFO to predict the HFO content at each site in the EU_dist and CN_dist (Methods and Texts S1–S2); (3) The Cd_dist dataset was formed by the Latin hypercube sampling technique and was combined with MSM output variables to train ML models (Methods and Text S3). The best-performing ML algorithm was used to predict the solid/liquid distribution at each site in EU_dist and CN_dist, and a comparative analysis of the differences between maps of the total and dissolved fractions of Cd was conducted, and (4) Knowledge transfer (KT) models were established within the ML framework to estimate the accumulation of Cd in wheat grains and roots.

The pipeline integrates data collection, database construction, model development, and spatial visualization.

Relationships between soil properties and Cd distribution patterns

Based on previous research on the retention of Cd in soils, the bioavailable Cd content in soils is predominantly controlled by adsorption onto four major reactive soil components: soil organic matter (SOM), dissolved organic matter (DOM), clay, and HFO. Cd adsorption to SOM and DOM was described with different ratios of humic acid (HA) and fulvic acid (FA). HA-Cd, FA-Cd, Clay-Cd, and HFO-Cd represent the amount of Cd adsorbed on these surfaces (Methods and Text S7). The reliability of the MSMs in both the EU and China was demonstrated by validating the correlation between the MSM-calculated equilibrium dissolved Cd (MSMs-Cd) and empirical data on plant uptake and extracted Cd. The robust linear correlation demonstrated that MSM-calculated dissolved Cd effectively indicates soil Cd bioavailability (Text S4) and was used to evaluate the Cd bioavailability. Within the geochemical-integrated ML framework, the gradient boosting regression tree (GBRT) model trained on the Cd_dist dataset achieved satisfactory R² values of 0.998 and 0.989 on the internal and external test sets, respectively, and thus was chosen as the final algorithm to predict the solid‒liquid distribution of Cd at each site (Text S3). The dissolved Cd concentration predicted by GBRT was termed as ML-Cd.

Taking the prediction of the ML-Cd as an example, three different feature analysis techniques were utilized to evaluate the importance of different features (Figs. S1 and S2). The total Cd content, pH, and SOC (accounting for 58% of OM²⁷) content were identified as the three most important features for the prediction of the ML-Cd, among which SOC was the dominant sorbent in the soil. As shown in Fig. S1a, SHapley Additive explanation (SHAP) analysis (a feature importance method, detailed in Methods) for predicting ML-Cd revealed that pH, SOC, clay, and HFO were negatively correlated with the content of ML-Cd. Conversely, higher concentrations of total Cd were associated with greater predictions of ML-Cd. A distinct pattern of total Cd content was observed where a dense cluster of high Cd concentration instances (red points) with small and positive SHAP values, while instances of lower Cd concentration (blue points) extended further toward the left, suggesting that low total Cd concentration had a stronger negative impact on ML-Cd. More details about the interaction effects can be found in the dependence plot, interaction plot, and heatmap, and the details are given in Text S5 and Fig. S3.

HFO is identified as an important interface for Cd, since it can enhance Cd fixation during soil redox cycles²⁸. Furthermore, it plays a crucial role in certain localities, especially in areas with high pH and low SOC content²⁹. For example, the pH on the Ibérian Peninsula increases from north to south, accompanied by decreases in the contents of SOC and HFO (Fig. S4)^30,31. Three representative areas within the region were selected, the corresponding distributions of Cd species on various soil components were computed for each soil site within these areas, and the 25 sites with the highest proportions of HFO-Cd (the percentage ratio of HFO-Cd relative to total soil Cd) are presented in Fig. 2. The results indicated that the proportion of Cd adsorbed by HFO in these three areas increased from north to south, with average values of 1.5 ± 2.9%, 15.5 ± 10.5%, and 30.7 ± 4.0%, respectively. Similarly, along the east coast of Italy, where the soil exhibited high pH, low SOC content, and relatively high HFO levels, the adsorption capacity of HFO was more pronounced, with an average HFO-Cd proportion of 33.3 ± 6.0%. The fraction of Cd on HFO ranged from 0 to 53.1% of the total Cd in the four selected areas.

a pH interpolated map of the EU (Adapted from Ballabio et al.³⁰ licensed under CC BY 4.0.), with the study area shown as a black inset, and (b) four hotspots exemplifying the varying significance of the HFO-Cd fraction.

Comparison of the spatial distribution of the total and dissolved fractions of soil Cd between the EU and China

The Cd content in soil is influenced not only by human activities (e.g., industrial activities, mining processes, and fertilization inputs¹³) but also by the regional geological background and weathering-to-soil processes. Specifically, soils over carbonate bedrocks accumulate Cd through a self-regulating cycle: carbonate rocks have a high potential for releasing Cd, while OM and Fe/Mn oxides immobilize it via adsorption/complexation³². This equilibrium between mobilization and retention explains the consistent spatial overlap of high-Cd areas and carbonate-rich regions in both EU and China (Fig. 3a, b)³³. As shown in Tables S1 and S2, the total Cd content in the EU varied between 0.11 and 1.55 mg kg⁻¹, with an average of 0.37 mg kg⁻¹ and a standard deviation of 0.17 mg kg⁻¹, while the total Cd content in China ranged from 0.01 to 14.22 mg kg⁻¹, with an average of 0.41 and a standard deviation of 1.03. The total Cd content in Chinese topsoil was approximately 10.8% greater than that in the EU.

a European map of total Cd¹³, b Chinese map of total Cd; c European map of dissolved Cd, and d Chinese map of dissolved Cd.

The prediction of the Cd solid‒liquid distribution at each site was carried out using the geochemical-integrated ML framework (GBRT trained on the Cd_dist dataset), and the major statistics of the prediction results are summarized in Table S3. Figs. 3c, d and S5 present distribution maps illustrating the ML-Cd and adsorbed-Cd concentrations on different soil reactive components in the EU or China. In particular, the predicted range for the ML-Cd within nonindustrial topsoil in the EU was 0.3–971.2 μg L⁻¹, with a mean value of 96.9 μg L⁻¹ and a standard deviation of 116.5 μg L⁻¹. In China, the corresponding range was 0.06–4151.2 μg L⁻¹, with a mean of 113.2 μg L⁻¹ and a standard deviation of 337.3 μg L⁻¹. ML-Cd was on average 16.8% greater than that in the EU. Table S4 and Fig. S6 present the statistical results and distribution maps of the proportions of various Cd forms to the total Cd content in China and the EU, respectively. On average, the proportions of ML-Cd to the total Cd in the EU and China were 26.3 ± 22.9% and 24.6 ± 18.0%, respectively. Moreover, the percentages of SOM-bound Cd (HA-Cd + FA-Cd) were 61.1 ± 20.0% and 65.4 ± 19.3%, respectively, surpassing the percentages of Cd associated with the other interfaces (clay-Cd, constituting 3.5 ± 3.3% and 5.8 ± 4.6%, and HFO-Cd, constituting 9.1 ± 11.7% and 4.3 ± 6.8% of the total Cd in the EU and China, respectively). Based on Tables S1 and S2, the relatively higher total Cd and lower contents of OC and HFO in Chinese topsoil were thought to be the main soil components contributing to the greater bioavailability. With nearly comparable mean pH values (6.33 in the EU and 6.64 in China), the mean OC content in China (14.76 g kg⁻¹) was markedly lower than that in the EU (36.53 g kg⁻¹), and a similar pattern was also observed for the mean HFO content (2.6 g kg⁻¹ in the EU and 1.6 g kg⁻¹ in the CN). Consequently, as shown in Table S3, the mean adsorption amounts of Cd on HA and HFO in China are less than those in the EU (148.6 μg L⁻¹ and 7.2 μg L⁻¹ in China, respectively, compared to 194.3 μg L⁻¹ and 28.1 μg L⁻¹ in the EU).

As shown in Fig. 3a, the highest amount of total Cd in nonindustrial topsoil was found in Ireland, followed by northern Spain, northern Sweden, Finland, and Poland. Lithuania, Slovenia, central Romania, and the west coast of Greece also exhibited relatively high Cd contents. Nevertheless, Fig. 3c depicts a different picture in which northern Sweden, Finland, and Poland exhibited the highest bioavailability rather than Ireland; on the other hand, Italy showed higher total Cd levels but a very low Cd bioavailability. The total Cd and MSM-Cd in England, Estonia, Latvia, Hungary, and Bulgaria posed low environmental risks. By combining SHAP analysis and maps of pH³⁰, SOC³¹, and clay³¹, it was determined that pH and SOC content were the primary drivers behind the differences in the spatial distributions of the total and ML-Cd (Fig. S6). Typically, low soil pH was the major driver for regions with lower total Cd contents but higher Cd bioavailability, such as Sweden and Poland. This is consistent with the scenario depicted in Fig. S6, where regions with a high proportion of ML-Cd closely align with areas of lower pH. On the other hand, despite the high total Cd content in Ireland, the high SOC content in Ireland resulted in substantial Cd adsorption by HA and FA and therefore a comparatively low Cd bioavailability.

In contrast to those in the EU, the distributions of total Cd, ML-Cd and the proportion of ML-Cd in China show a similar pattern to that of total Cd. As shown in Fig. 3b, d, areas exhibiting higher levels of both total and ML-Cd were found in Taiwan Province, Guangxi Province, and the junction of Yunnan and Guizhou Provinces. This phenomenon can be attributed to a noticeable pattern in the distribution of soil properties across China. Specifically, soil pH tends to be greater in the north and lower in the south⁵, while SOC and clay content tend to be lower in the north and higher in the south^5,34.

Knowledge transfer for bioavailability prediction

Most regional or continental risk assessments of soil Cd solely relied on total content¹³, while bioavailable Cd measurements content remain problematic for risk evaluation because there are different types of bioavailable Cd contents as defined by different extraction procedures. Therefore, we integrated the above geochemical adsorption processes with the crop Cd uptake within the soil‒plant system using knowledge transfer (KT). The KT model was established by incorporating the predictive outcomes of the Cd speciation distribution obtained from the best-performing GBRT model (Table S5), and trained alongside a nonmechanistic data-driven (DD) model on Chinese wheat data. Although comparative performance metrics (performance score in Table S6; SHAP for DD shown in Fig. 4a, c, SHAP for KT shown in Fig. 4b, d) demonstrated similar accuracy between models, the KT model distinguished itself by encoding geochemical interaction governing Cd distribution. Feature importance analysis (Fig. 4b, d) confirms that ML-Cd and clay Cd exhibit predominant influence in the KT model, outperforming all other soil properties by substantial margins. It is physically sound that both the soluble and electrostatically clay-bound Cd can be attributed to the exchangeable fraction³⁵. These findings were also consistent with that Cd bioavailability is not solely determined by the ML-Cd fraction but also influenced by the presence of Cd at other interfaces during the dynamic Cd uptake process³⁶. Hence, the KT model effectively addresses this complexity by leveraging the knowledge transferred from a different but related task in an end-to-end paradigm. Consequently, our geochemical-integrated ML framework successfully bridged the knowledge gap by integrating knowledge from geochemical processes into crop uptake processes, leading to a better understanding of predicting Cd accumulation in both wheat roots and grains. Its potential applications in diverse species and complex scenarios await further validations.

Performance and SHAP values for predictive models of wheat Cd accumulation, a data-driven (DD) model in grain, b knowledge transfer (KT) model in grain, c KT model in root, d KT model in root.