Lei Zhang

All too often, it is unclear whether digital soil mapping (DSM) models can support causal interpretation. A common practice in DSM studies is to interpret the importance of covariates for prediction. This carries an implicit causal assumption that is rarely stated and even more rarely justified. Because DSM relies entirely on observational data, it is widely assumed that causal inference is not possible. But is it? Here, we discuss the conditions under which causal inference with observational data is possible and two views of causality. We show that while under each of the views causal inference may be possible, a so-called generative view is the one most capable of satisfying the conditions for causal inference in DSM. Generative causality treats causation as the system of processes that produce observed associations, rather than relying on associations themselves, as is common in current DSM studies. Realizing this perspective requires DSM to shift towards models in which soil-forming factors influence soil properties through explicitly modelled processes, which some would call process-informed DSM. Since these processes are ‘fully determined’ by the modeller’s specification, they offer a structured means to control confounding and open the door to applying existing causal inference frameworks. While generative DSM is formally possible, we should ultimately ask whether causal inference ought to be a primary goal, since the primary strength of DSM lies not in establishing causality but in delivering accurate predictions and highlighting patterns that warrant further investigation.

Tidal wetlands are important blue carbon ecosystems, yet global estimates of soil organic carbon (SOC) stocks and dynamics across mangroves, salt marshes, and tidal flats remain unclear. Using 42,154 records and 63 environmental covariates, we generate global tidal wetland SOC maps for two periods (2009-2014 and 2015-2020) at three depths. Here we show that SOC stocks decrease from \\{8.80}_{2.61}^{20.94}\\Pg (mean, 5th-95th percentile prediction interval) to \\{8.27}_{2.27}^{20.61}\\Pg, representing a 6.11% decline, with greater losses in deeper layers. Low-latitude regions, particularly South Asia and Africa, are loss hotspots, and lower-middle and upper-middle income countries experience pronounced losses. Mangroves suffer the largest losses primarily due to SOC density decline, whereas salt marshes gain in both area and SOC density. Climate, elevation, and total suspended matter drive mangrove SOC changes; oceanographic factors influence salt marshes and dominate tidal flats. This work reveals the magnitude of blue carbon spatio-temporal change, underscoring the need for targeted conservation.

Soil organic carbon (SOC) underpins the global carbon cycle and represents a central lever for climate change mitigation and food security. Yet its accurate spatiotemporal quantification remains a challenge, owing to the complex interactions among biological, chemical and physical processes operating across scales. This review critically evaluates the current state of SOC spatiotemporal modelling frameworks and their limitations, and future directions. Process-based models provide mechanistic insight into carbon dynamics but are constrained by parameterisation, structural assumptions and computational demands. In contrast, machine-learning (ML) approaches excel at capturing spatial patterns from large datasets but often struggle to represent temporal kinetics, enforce physical consistency or generalise across scales and environmental contexts. We argue that the next generation of SOC modelling will emerge from the convergence of process-based understanding and data-driven inference through knowledge-guided ML and hybrid modelling strategies. We synthesise four principal integration pathways: (1) meta-modelling to accelerate computationally intensive simulations; (2) sequential hybridisation to embed mechanistic trends as dynamic covariates; (3) ensemble frameworks to reduce structural uncertainty; and (4) data assimilation to constrain model trajectories with observations. We further highlight emerging frontiers, including residual and parameter learning, physics-informed neural networks, foundation models for earth observations, and omics-informed frameworks that link microbial functional potential to carbon turnover processes. We conclude that progress toward causally interpretable, uncertainty-aware monitoring frameworks will require tighter integration of mechanistic theory, interpretable artificial intelligence and cross-scale data synthesis.

Accurate modelling and mapping soil organic carbon are crucial for supporting soil health restoration and climate change mitigation at both regional and global scales. However, regional soil predictions often suffer from data scarcity and high prediction uncertainty. Utilizing a pre-trained global-to-regional soil carbon predictive model can be a potential solution to address this challenge. Despite its promise, how to construct and apply the global-scale model to enhance regional-scale soil carbon mapping remains largely unexplored. Here, we propose the Global Soil Carbon Pre-trained Model (GSoilCPM), a deep-learning-based domain adaptative model, to enhance regional-scale soil carbon predictions. Based on large amount of environmental covariate data and 106,167 soil samples across the globe, we verify our hypothesis of the effectiveness of this ’global-to-regional’ modelling strategy. The pre-trained model can be then transferred and fine-tuned to bridge the regional- and global-scale soil–environment relationships. We applied and validated this modelling strategy in four regional-scale study areas, three in the Northern Hemisphere and one in the Southern Hemisphere, each with distinct environmental background. Compared to traditional modelling approaches as a baseline, four case studies all demonstrated significant improvement in prediction accuracy across diverse environments and varying data availabilities. The average percentage improvement across all regions is 10.93% (absolute values decreased by 1.20 g kg−1 averagely) in MAE and 29.04% (absolute values increased by 0.10 averagely) in CCC. The applicability and future horizons of using GSoilCPM were further discussed. We further reveal that regions with fewer soil samples or lower baseline accuracy benefit more from the pre-trained global model. Our findings highlight the advantages of leveraging the generalized knowledge from global models to enhance specifically localized soil modelling, positioning a potential paradigm shift in digital soil mapping, and far-reaching implications for soil monitoring and land management.

The turnover time (τ) of global soil organic carbon is central to the functioning of terrestrial ecosystems. Yet our spatially-explicit understanding of depth-dependent variations and environmental controls of τ at a global scale remain incomplete. In this study, we combine multiple state-of-the-art observation-based datasets, including over ninety thousand geo-referenced soil profiles, the latest root observations distributed globally, and large amounts of satellite-derived environmental variables, to generate global maps of apparent τ in topsoil (0–0.3 m) and subsoil (0.3–1 m) layers with a spatial resolution of 30 arcsec ( 1 km at the Equator). We show that subsoil τ (385 [20, 3485] years [mean with a variation range from 2.5th to 97.5th percentile]) is over eight times longer than topsoil τ (15 [11, 137] years). The cross-validation shows that the fitted machine learning models effectively captured the variabilities in τ, with R2 values of 0.87 and 0.70 for topsoil and subsoil τ mapping, respectively. The prediction uncertainties of the τ maps were quantified for better user applications. The environmental controls on top- and subsoil τ were investigated at global, biome, and local scales. Our analyses illustrate that how temperature, water availability, physio-chemical properties and depth exert jointly impacts on τ. The data-driven approaches allow us to identify their interactions, thereby enriching our comprehension of mechanisms driving nonlinear τ–environment relationships from global to local scales. The distributions of dominating factors of τ at local scales were mapped for identifying context-dependent controls on τ across different regions. We further reveal that the current Earth system models may underestimate τ by comparing model-derived maps with our observation-derived τ maps. The resulting maps with new insights demonstrated in this study facilitate the future modelling efforts of carbon cycle–climate feedbacks and supporting effective carbon management. The dataset is archived and freely available at https://doi.org/10.5281/zenodo.14560239 (Zhang, 2025).

Understanding spatio-temporal patterns of soil organic carbon (SOC) is critical for global climate change mitigation and sustainable soil management. However, information on long term dynamics of SOC over large area is lacking. Supported by soil samples collected over years and environmental covariates, space and time digital soil mapping (ST-DSM) has become an important and effective method to reveal the spatio-temporal changes of SOC. The contiguous United States (CONUS) has abundant and well-documented soil samples with time labels, which lays the groundwork for us to estimate the long-term SOC dynamics in multiple soil layers over that region with high resolution. Specifically, we propose leveraging time-series soil data from World Soil Information Service (WoSIS) and International Soil Carbon Network (ISCN) to build ST-DSM models at different soil depths based on matching environmental covariates and machine learning techniques (random forest framework). Then, multi-depth ST-DSM models are employed to generate spatial prediction of SOC in different layers (0–15 cm, 15–30 cm, 30–60 cm and 60–100 cm) from 1955 to 2014 at 250 m resolution and 5-year intervals (1955–1959, 1960–1964,…., 2010–2014). Meanwhile, predictive uncertainties are quantified via Quantile Regression Forest (QRF). Furthermore, we analyze the dynamic trends in SOC stocks across various depths and land uses. The results indicate that over 60 years, overall SOC stocks in 0–100 cm demonstrate a multi-stage change of “rising-fluctuating”. SOC stocks in 0–100 cm rose from 68.40 Pg (1 Pg = 10^15 g) to 70.33 Pg with an increase rate of 32.25 Tg (1 Tg = 10^12 g) per year. SOC stocks in 0–15 cm layer declined slightly before 1970 and increased thereafter; the 15–100 cm layer fluctuated with an overall rising trend; notably, the 60–100 cm layer exhibited a steady upward trend with minimal fluctuations since 1980s, suggesting a greater SOC accumulation potential in soil beneath 15 cm. Across different land uses, the topsoil of cropland may be a source of carbon, while forests are important carbon sinks. SOC stocks in surface soil (0–15 cm) in cropland areas notably decreased, and subsurface and deep soil in cropland (15–30 cm, 30–60 cm and 60–100 cm) showed a fluctuating rising trend. SOC stocks in forest show an overall fluctuating rising trend. Pastureland, unmanaged natural grasslands and sparse/no vegetation areas once showed some carbon sequestration capacity in the past, but they have transitioned to slower accumulation. This research enhances our understanding of soil carbon dynamics at a national scale and provides references for the development of effective soil management and climate mitigation strategies.

Coastal wetlands are important blue carbon ecosystems that play a significant role in the global carbon cycle. However, there is insufficient understanding of the variations in soil organic carbon (SOC) stocks and the mechanisms driving these ecosystems. Here we analyze a comprehensive multi-source dataset of SOC in topsoil (0–20 cm) and subsoil (20–100 cm) across 31 coastal wetlands in China to identify the factors influencing their distribution. Structural equation models (SEMs) reveal that hydrology has the greatest overall effect on SOC in both soil layers, followed by vegetation, soil properties, and climate. Notably, the mechanisms driving SOC density differ between the two layers. In topsoil, vegetation type and productivity directly impact carbon density as primary sources of carbon input, while hydrology, primarily through seawater salinity, exerts the largest indirect influence. Conversely, in subsoil, hydrology has the strongest direct effect on SOC, with seawater salinity also influencing SOC indirectly through soil and vegetation mediation. Soil properties, particularly pH, negatively affect carbon accumulation, while climate influences SOC indirectly via its effects on vegetation and soil, with a diminishing impact at greater depths. Using Random Forest, we generate high-resolution maps (90 m × 90 m) of topsoil and subsoil carbon density (R2 of 0.53 and 0.62, respectively), providing the most detailed spatial distribution of SOC in Chinese coastal wetlands to date. Based on these maps, we estimate that SOC storage to a depth of 1 m in Chinese coastal wetlands totals 74.58 ± 3.85 Tg C, with subsoil carbon storage being 2.5 times greater than that in topsoil. These findings provide important insights into mechanism on driving spatial pattern of blue carbon and effective ways to assess carbon status on a national scale, thus contributing to the advancement of global blue carbon monitoring and management.

Monitoring and modelling soil organic carbon (SOC) in space and time can help us to better understand soil carbon dynamics and is of key importance to support climate change research and policy. Although machine learning (ML) has attracted a lot of attention in the digital soil mapping (DSM) community for its powerful ability to learn from data and predict soil properties, such as SOC, it is better at capturing soil spatial variation than soil temporal dynamics. By contrast, process-oriented (PO) models benefit from mechanistic knowledge to express physiochemical and biological processes that govern SOC temporal changes. Therefore, integrating PO and ML models seems a promising means to represent physically plausible SOC dynamics while retaining the spatial prediction accuracy of ML models. In this study, a hybrid modelling framework was developed and tested for predicting topsoil SOC stock in space and time for a regional cropland area located in eastern China. In essence, the hybrid model uses predictions of the PO model in unsampled years as additional training data of the ML model, with a weighting parameter assigned to balance the importance of SOC values from the PO model and real measurements. The results indicated that temporal trends of SOC stock modelled by PO and ML models were largely different, while they were notably similar between the PO and hybrid models. Cross-validation showed that the hybrid model had the best performance (RMSE = 0.29 kg m−2), with a 19 % improvement compared with the ML model. We conclude that the proposed hybrid framework not only enhances space-time soil carbon mapping in terms of prediction accuracy and physical plausibility, it also provides insights for soil management and policy decisions in the face of future climate change and intensified human activities.

Sampling design can significantly reduce the uncertainty in geospatial predictions. In this paper, we developed an adaptive uncertainty-guided stepwise sampling (AUGSS) method to select sampling locations to supplement existing legacy sample points whose representation should be improved. The proposed method selects supplemental samples in a stepwise manner as guided by an objective function with two weighted sub-objectives. One reduces the area with high prediction uncertainty, and the other minimizes the overall prediction uncertainty for the entire area. The method takes an adaptive approach to adjust weights for the two sub-objectives and to tune an uncertainty threshold controlling whether a location can be reliably predicted during the sampling procedure. A case study on soil property prediction shows that AUGSS outperforms the stratified random sampling (SRS) and the non-adaptive uncertainty guided sampling method (UGSS) in terms of RMSE and Lin’s concordance correlation coefficient with different sample sizes. This study shows that the AUGSS method offers a potential for effectively adding supplemental samples to existing samples which are insufficient for spatial prediction. The adaptive strategy guided by predicted uncertainty provides an efficient support to improve the spatial pattern of samples, which plays a key role in the result accuracy of geospatial predictive mapping.

Urban environments, regarded as “harbingers” of future global change, may exert positive or negative impacts on urban vegetation growth. Because of limited ground-based experiments, the responses of vegetation to urbanization and its associated controlling factors at the global scale remain poorly understood. Here, we use satellite observations from 2001 to 2018 to quantify direct and indirect impacts of urbanization on vegetation growth in 672 worldwide cities. After controlling for the negative direct impact of urbanization on vegetation growth, we find a widespread positive indirect effect that has been increasing over time. These indirect effects depend on urban development intensity, population density, and background climate, with more pronounced positive effects in cities with cold and arid environments. We further show that vegetation responses to urbanization are modulated by a cities’ developmental status. Our findings have important implications for understanding urbanization-induced impacts on vegetation and future sustainable urban development. Positive indirect effects of global urbanization on vegetation growth partially offset the negative direct impact.

Sampling design plays a key role in digital soil mapping (DSM). Efficient sampling design for multiple soil properties is increasingly needed for multivariate soil survey and mapping. However, most of the present sampling methods are not developed for multiple soil properties. Different soil properties have different influential covariates, but usually only one set of covariates is used in designing samples for multiple soil properties which makes simultaneously mapping multiple soil properties accurately difficult. This paper proposed a multiple soil properties oriented representative sampling strategy (MPRS) by considering the influential environmental covariates for each soil property. The method first selects the most influential set of environmental covariates for each soil property, then uses fuzzy c-means (FCM) clustering to generate environmental clusters relating to spatial variation patterns for each soil property, and the selected samples are representative of as many typical locations of environmental clusters for multiple soil properties as possible. The proposed sampling method was applied for mapping soil sand content and soil organic matter content at surface (0–20 cm) and subsurface (20–40 cm) layers in a study area with 5900 km2 located in Anhui Province, China, and compared with two methods, the purposive sampling (PS) method and integrative hierarchical stepwise sampling (IHS) method. The results showed that the proposed sampling method achieved the most accurate prediction for most of the four soil properties over different sample sizes. The proposed sampling method also has an advantage to extract representative samples which can better cover multiple soil properties with a limit of a small sample size. On average, the improvement of prediction accuracy by using the MPRS method was 38.1% and 36.3% compared with PS and IHS in terms of R2, 4.8% and 4.6% in terms of RMSE, and 11.7% and 13.7% in terms of CCC, respectively. Our case study confirmed the necessity to consider the difference of the influential environmental variable combinations for the multiple soil properties oriented sampling design. We conclude that MPRS is a potential effective method for supporting DSM for multiple soil properties.

Cropland soil carbon not only serves food security but also contributes to the stability of the terrestrial ecosystem carbon pool due to the strong interconnection with atmospheric carbon dioxide. Therefore, the better monitoring of soil carbon in cropland is helpful for carbon sequestration and sustainable soil management. However, severe anthropogenic disturbance in cropland mainly in gentle terrain creates uncertainty in obtaining accurate soil information with limited sample data. Within the past 20 years, digital soil mapping has been recognized as a promising technology in mapping soil carbon. Herein, to advance existing knowledge and highlight new directions, the article reviews the research on mapping soil carbon in cropland from 2005 to 2021. There is a significant shift from linear statistical models to machine learning models because nonlinear models may be more efficient in explaining the complex soil-environment relationship. Climate covariates and parent material play an important role in soil carbon on the regional scale, while on a local scale, the variability of soil carbon often depends on topography, agricultural management, and soil properties. Recently, several kinds of agricultural covariates have been explored in mapping soil carbon based on survey or remote sensing technique, while, obtaining agricultural covariates with high resolution remains a challenge. Based on the review, we concluded several challenges in three categories: sampling, agricultural covariates, and representation of soil processes in models. We thus propose a conceptual framework with four future strategies: representative sampling strategies, establishing standardized monitoring and sharing system to acquire more efficient crop management information, exploring time-series sensing data, as well as integrating pedological knowledge into predictive models. It is intended that this review will support prospective researchers by providing knowledge clusters and gaps concerning the digital mapping of soil carbon in cropland.

Numerous machine learning models have been developed for constructing the relationship between soil classes or properties and its environmental covariates in digital soil mapping (DSM). Most machine learning models are trained with a supervised learning (SL) method based on training samples. However, the collected sample data is often limited in practice due to that field sampling is expensive and time-consuming. The insufficient samples may limit the learning ability of the model to a large extent. Semi-supervised machine learning, a new machine learning paradigm that makes use of both unsampled data and a small amount of sampled data in the learning process, can be a potential effective method for DSM. In this study, we present a self-training semi-supervised learning (SSL) method for DSM. Different with the SL method for machine learning models, the SSL method not only utilizes the sampled locations but also the abundant environmental covariate information at the unvisited locations. Its basic idea is to iteratively enlarge the training data set by adding the unsampled points with high prediction confidence from the unvisited locations until a stopping criterion reached. The proposed SSL method was applied in machine learning models for predicting soil classes in Heshan Farm of Nenjiang County in Hei­ longjiang Province, China. Three machine learning models, including multinomial logistic regression (MLR), knearest neighbor (KNN) and random forest (RF), were selected to evaluate the efficiency of the SSL method. The entropy threshold was an important parameter in the SSL method, and a sensitivity analysis on this parameter was conducted with using a series of entropy thresholds. The SSL method was compared with the SL method for the three machine learning models for soil prediction. A cross-validation was employed to evaluate the accuracy of the predicted soil class maps generated based on each method. The results showed that the prediction ac­ curacies (the proportion of the correctly predicted samples over the total number of validation samples) of the SSL method were higher than those of the SL method for MLR, KNN, and RF by 5.9%, 12.2%, and 6.0%, respectively. RF-SSL was the most accurate model in the study area, followed by KNN-SSL. Meanwhile, the selftraining SSL method for the KNN model had the largest improvement comparing with the other two models. Furthermore, the predicted soil maps using the SSL method showed a more reasonable spatial variation pattern of soil classes. In the study area, a suitable value of the entropy threshold was 0.8 ~ 1.0. We concluded that the SSL method improved the soil prediction accuracy compared with the SL method when applying machine learning models for DSM, and thus is a potential efficient method for DSM with limit sample data.

Vegetation coverage in highly developed areas has been significantly altered in response to multiple disturbances over recent decades. However, the major driving factor of vegetation coverage change in these areas remains unclear, with climate change and anthropogenic factors playing interactive roles under different soil and terrain conditions. Comprehensively understanding the underlying drivers of vegetation change can provide references for regulating environmental management and prevention of vegetation degradation. In this paper, a structural equation modeling (SEM) method was employed to quantify the effects of fundamental natural environment (i.e. the relative stable variables including soil and topography), climate change and human activity change on vegetation coverage change in Jiangsu province, China from 2000 to 2015. Four variables including land use, population density, road impact and night lights were used to indicate human activities. The results showed that the increase of NDVI smaller than 0.10 covered 39.13% of the study area while the decrease of NDVI larger than 0.10 accounted for 20.23%. Areas with NDVI increase mainly distributed in croplands in northern Jiangsu. This could be explained by the increase of crop yield due to the development of modern agriculture. The decrease of NDVI was mainly observed in southern Jiangsu with higher urbanization level and city centers in northern Jiangsu, indicating the effect of rapid urbanization on vegetation degradation. The constructed SEM model suggested that the total effects (influential coefficients) of fundamental natural environment, climate change, and human activity change on NDVI change in Jiangsu were −0.24, 0.17, and −0.74, respectively. Although the fundamental natural environment didn’t have a direct effect on NDVI change, but it had an indirect effect through interactions with human activities. We also constructed SEM models for northern and southern Jiangsu separately, due to their different natural environment and changing patterns of climate change. The results indicated the different driving mechanisms of NDVI change in northern and southern Jiangsu. Furthermore, the results suggested night light as the best indicator of human activity change, followed by the road impact index. We concluded that our study offered a framework to better understand and explain the complex interrelationships behind the spatial temporal change of NDVI.

Obtaining the spatial distribution information of soil organic carbon (SOC) is significant to quantify the carbon budget and guide land management for migrating carbon emissions. Digital soil mapping of SOC at a regional scale is challenging due to the complex SOC-environment relationships. Vegetation phenology that directly indicates a long time vegetation growth characteristics can be potential environmental covariates for SOC prediction. Deep learning has been developed for soil mapping recently due to its ability of constructing high-level features from the raw data. However, only dozens of predictors were used in most of those studies. It is not clear that how deep learning with long term land surface phenology product performs for SOC prediction at a regional scale. This paper explored the effectiveness of ten-years MODIS MCD12Q2 phenology variables for SOC prediction with a convolutional neural network (CNN) model in Anhui province, China. Random forest (RF) was applied to compare with CNN using three groups of environmental variables. The results showed that adding the land surface phenology variables into the pool of the natural environmental variables improved the prediction accuracy of CNN by 5.57% of RMSE and 31.29% of R2. Adding phenology variables obtained a higher accuracy improvement than adding Normalized Differences Vegetation Indices. The CNN obtained a higher prediction accuracy than RF regardless of using which group of variables. This study proved that land surface phenology metrics were effective predictors and CNN was a promising method for soil mapping at a regional scale.