Artificial intelligence and soil carbon modeling demystified: power, potentials, and perils

INTRODUCTION

The global soil carbon (C) pool has been estimated to exceed the amount of carbon stored in the atmosphere and vegetation^[1-3], though uncertainties to quantify below-ground carbon accurately still exist. According to Scharlemann et al.^[2] (2014), the total estimated global SOC stock mean is about 1500 Pg C with highest soil organic carbon (SOC) stored in boreal moist (~350 Pg C), cool temperate moist (~210 Pg C), and tropical moist ecosystems (~150 Pg C). However, the uncertainty of studies that assessed soil carbon stocks (0-1 m) is considerable ranging from 504 to 3000 Pg C among 27 different global assessments^[2]. These uncertainties are due to the large variability in SOC, changing climate and environmental conditions that impact soil C dynamics, different methods and modeling approaches to estimate SOC, as well as data limitations and overuse of legacy data^[4]. A substantial proportion of SOC has been measured only in the topsoil (< 30 cm) with sparser observations in the subsoils that have been estimated to store about half of the global soil carbon^[5-7].

The spatial and temporal variability of soil carbon storage, soil carbon sequestration (SCseq), and carbon fluxes are critically important to address soil health, soil security, food security, regenerative agriculture, and climate-smart soil conservation management. Soil carbon provides an ecosystem service implicated in numerous soil functions, such as nutrient regulation and mitigation of greenhouse gas (GHG) emissions that are pivotal to emergent carbon economies and markets. The significance of soil carbon in global biogeochemical cycles is profound. To sustain multiple soil functions and preserve soil health and soil security several quantification methods, among them artificial intelligence (AI), have been utilized at escalating spatial scales. The purpose of this review paper is to explore the emergence, applications, and potential of AI - machine leaning (ML) and deep learning (DL) algorithms - to model soil carbon storage and soil respiration (R_s) at regional and global scales. A critical discussion of the power, potentials, and perils of AI soil carbon modeling is provided.

AI MODELING OF SOIL CARBON STORAGE AND DYNAMICS

In the discipline of pedometrics, the adoption of AI algorithms in digital soil mapping emerged in the early 2000s. From 2015 onward advanced AI soil models were developed in the domains of proximal soil sensing and soil carbon modeling using large environmental data hypercubes^[28]. A comprehensive review of AI-ML algorithms applied in digital soil mapping, including soil carbon modeling, was provided by Khaledian and Miller^[29] (2020), a review of DL for digital soil mapping was provided by Padarian et al.^[30] (2019), and a review of DL in agriculture was provided by Kamilaris and Prenafeta-Boldú^[31] (2018). Recently, a comprehensive review of ML and remote sensing methods to estimate various soil indicators was presented by Diaz-Gonzalez et al.^[32] (2022). In this section I present a brief overview of some of the most prominent AI methods that have been employed in soil carbon modeling which informs a critical discussion of the power, potentials, and perils of these methods.

Soil carbon AI models are build using hypercubes of environmental covariates as inputs [Figure 1 and 2A]. These environmental covariates represent the domains of soils (S), topography (T), ecology (E), parent material or lithology (P), atmosphere or climate (A), water or hydrology (W), biota with vegetation and organisms (B), and human activities/management (H)^[4,33-35] similar to the conceptual framework of SCORPAN (McBratney et al.^[36] 2003). The STEP-factors are relatively stable across the human lifetime, while the AWBH-factors are dynamic in space and across time. Each of these factors can be quantified through a set of variables. For example, the S factor can be characterized by soil data such as soil texture, pH, soil taxonomic class, cation exchange capacity, etc. derived from legacy soil maps or databases, proximal soil sensing (e.g., visible-near infrared spectroscopy, VNIR; mid-infrared spectroscopy), gamma ray sensing, and remotely sensed soil moisture data. The factor A may be populated by climatic data such as long-term average of mean annual precipitation, seasonal variation of minimum and maximum temperature, and long-term solar radiation, while B can be populated by satellite-derived land use/land cover maps, vegetation indices like the Normalized Difference Vegetation Index (NDVI) or Enhanced Vegetation Index (EVI) derived from satellite data, biodiversity, and habitat data. The factor H may be populated by variables from the social, cultural, economic, and political domains (e.g., greenhouse gas emission data, land management data such as tillage operations, and fertilization amount and type). The aim is to populate the STEP-AWBH factors with environmental geodata that influence the carbon cycle. Xiong et al.^[37] (2014) exemplified the STEP-AWBH model and multiple AI-ML methods in Florida, United States, to develop prediction models for SOC stocks.

Figure 1. Environmental covariates from the domains of soils (S), topography (T), ecology (E), parent material (P), atmosphere (A), water (W), biota (B), and human (H). Each of these domains is represented by a wide variety of variables that facilitate AI-soil carbon modeling (image is courtesy of S. Grunwald).

Figure 2. (A) Functional relations between environmental covariates (STEPAWBH factors with variables i = 1, 2, 3, ……, N). SC denotes a variable of the carbon cycle (target output), for example, soil organic carbon (SOC) stock, SOC density, total soil carbon, soil respiration (R_s), soil carbon sequestration (SCseq), soil carbon pools, soil carbon fractions, etc. x is spatial location (with xy coordinates; or latitude/longitude); z is soil depth with z = 1, 2, 3, ……, Z; and t is time with t1, t2, t3, …., T. (B) AI model predicting soil carbon (SC) from environmental covariates. Simplified representation of a machine learning ensemble tree method (e.g., Classification and Regression Trees, CART, or Cubist) with tree branches and data splits.

Commonly applied ai algorithms to model soil carbon

Classification and Regression Trees (CART) were introduced by Breiman^[38] (1984) and have served as foundational approach onto which other ML have built on [Figure 2B]. According to Breiman^[38] (1984), CART involves constructing a set of decision trees on the predictor variables. The trees are grown by repeatedly stratifying the dataset into successively smaller subsets (child node) with binary splits based on a single categorical or continuous predictor variable. The splitting procedure is applied until the best split is chosen based on the one that maximizes the response into two homogenous groups (i.e., minimizing variability within each child node)^[39].

A variant of CART is Bagged Regression Trees (BaRT) which is an ensemble decision tree method that involves the averaging of several individual trees to acquire a final prediction. Individual regression trees have shown somewhat erratic modeling results where small changes in input variables produces large differences in output trees^[40,41]. This limitation of individual CART is overcome by bagging (i.e., bootstrap aggregation) in BaRT. Bagging is an ensemble learning method that is commonly used to reduce variance within noisy datasets. In bagging, a random sample of data in a training (or calibration) set is selected with replacement, which means that the individual data points can be chosen more than once. Thus, the procedure grows a regression tree from each bootstrap sample. To obtain the overall final prediction for a target variable the results of each individual tree are averaged^[42].

Boosted Regression Trees (BoRT) belong to the Gradient Boosting Modelling family, which is one among many methods to predict the function F that maps the values of a set of predictor variables x = {x1,..,xp} into the values of the output variable y, by minimizing a specified loss function L. In BoRT the prediction is performed using boosting^[43]. In general, boosting methods are applied to significantly improve the performance of a given estimation method, by generating instances of the method iteratively from a training data set and additively combining them in a forward “stagewise” procedure. BRT uses a specialized form (for regression trees) of the Stochastic Gradient Boosting^[44]. The gradient boosting machine algorithm was described in detail by Friedman^[44] (2001). The regression tree algorithm developed by Breiman^[38] (1984) served as the foundation of BoRT, which has shown to boost accuracy compared with simple regression trees, mainly due to its stochastic gradient boosting procedure aiming at minimizing the risk of overfitting and improving its predictive power^[45]. According to Hastie et al.^[41] (2009), in BoRT trees are grown sequentially with each tree grown using the information from previously grown trees. The BoRT algorithm facilitates fitting the model to the data in an iterative process. At each iteration, individual regression trees, are fitted on a fraction (namely the bag fraction) of the dataset sampled without replacement. The main parameters for fitting BRT are the tree size and the learning rate.

Random Forest (RF) is a widely used ML method consisting of an ensemble of randomized classification and regression trees^[38,46]. The RF algorithm grows different trees by randomly and repeatedly selecting predictor variables and training cases to develop a random population of trees. The algorithm grows an ensemble of regression trees based on binary recursive partitioning, where the predictor space at each tree node is partitioned based on binary splits on a subset of randomly selected predictors^[47]. The output of RF is the average of individual tree predictions. It has been shown that the RF algorithm can be very efficient, especially when the number of descriptors is very large^[48]. The RF model is capable of simultaneously handling categorical and continuous variables, as well as complex high-order variable relationships such as nonlinearity and interaction effects. Conditional quantiles can be inferred with Quantile Regression Forests (QRF), a generalization of RF. Quantile regression forests is a non-parametric technique used to estimate the conditional quantiles of multidimensional predictor variables. The benefits of QRF is its ability to predict more accurate results for the conditional distribution of the response variable^[49].

The Support Vector Machines (SVM) applies a projection of the input data into a high-dimensional feature space using a valid kernel function and then it uses a simple linear regression within this enhanced space^[50]. This resulting linear regression function in the high-dimensional feature space corresponds to a non-linear regression in the original input space [Figure 3A]. In the new hyperspace, SVM aims to construct an optimal hyperplane that separates classes and creates the widest margin between their data (i.e., classification), or that fits data and predicts (i.e., SVR) with minimal empirical risk and complexity of the modelling function^[51,52]. Support Vector Regression (SVR) is a generalization of SVM and is used for nonlinear classification and regression^[53]. The ε-SVR uses a loss function to define the borders (hyperplane) of the regression function. Hence the regression function lies between ± ε (maximum error). Therefore, the loss is equal to 0 if the difference between the predicted and measured values is less than ε^[51].

Figure 3. AI models predicting soil carbon (SC) from environmental covariates. Idealized model representation of (A) support vector machines (SVM), and (B) partial least squares regression (PLSR).

The Partial Least Square Regression (PLSR) was developed by Wold^[54] (1975) in econometrics and has since been widely used in many disciplines, including soil science and pedometrics. The PLSR algorithm relates the response variable (e.g., SOC) and a large number of highly collinear predictor variables (e.g., environmental covariates) through a multivariate model to identify successive orthogonal principal components (latent variables) that maximize the covariance between the response and predictor variables (Garthwaite^[55], 1994). These latent factors are defined as linear combinations constructed between input variables (i.e., predictors) and response variables, such that the original multidimensionality is reduced to a lower number of orthogonal factors to detect the structure in the relationships between predictor variables and between these latent factors and the response variables [Figure 3B]. The extracted factors account for successively lower proportions of original variance^[56,57]. According to Carrascal et al.^[56] (2009), PLSR is especially suited to analyzing a large array of interrelated predictor variables (i.e., variables that are not truly independent). Soil carbon often covaries with other soil and environmental properties, and thus, PLSR is well suited to handle such multicollinearities.

The Cubist (Cub) algorithm is a decision tree model with piecewise linear models^[58]. Cubist partitions the response data into subsets within which their characteristics are similar with respect to the predictors. A series of if-else conditions define rule-based partitions which are then arranged in a hierarchy. The simplest partition is based on only one predictor, though often multiple predictors are used to form a partition which are expressed in form of regression equations making models transparent for users [Figure 2B].

In general, an ANN is a massively parallel distributed processor made up of simple processing units, which has a natural propensity for storing experiential knowledge and making it available for use [Figure 4A]. The benefits of ANN are (1) ability to learn and therefore generalize; (2) solve complex problems (e.g., complex soil carbon-environmental relationships); (3) model linear, nonlinear, and high-order relations between inputs and outputs; and (4) provide input-output mapping (i.e., supervised learning)^[59]. The backpropagation ANN algorithm is a multi-layer perceptron neural networks (i.e., a MLP neural nets) [Figure 4B]. The architecture of the MLP neural nets consists of input, one or multiple hidden, and output layers, each with a set of interconnected nodes (neurons) working in parallel to fit input data and output values through adjusting weights and cost function^[60]. Hidden nodes represent abstract factors with no physical connection to ecosystems (i.e., the outside world). The purpose of hidden nodes is to transfer information from the input nodes to the output nodes. Backpropagation supervised learning is based on the error-correction learning rule. It consists of two passes through the different layers of the network: a forward pass and a backward pass. In the forward pass an activity pattern (input vector) is applied to the sensory nodes of the network and its effect propagates through the network layer by layer. Finally, a set of outputs is produced as the actual response of the network. During the forward pass the synaptic weights of the network are all fixed, while during the backward pass the synaptic weights are all adjusted in accordance with an error-correction rule. The actual response of the network is subtracted from a desired (target) response to produce an error signal. This error signal is then propagated backward through the network. During this backward pass the synaptic weights are adjusted to make the actual response of the network move closer to the desired response in a statistical sense^[59]. Various backpropagation-based implementation methods including structure-fixed training and structure-adaptive training methods as well as sparse representation and dictionary learning methods were described in Wythoff^[61] (1993). Recurrent neural networks (RNN) are algorithms for sequential data along a temporal sequence. These kind of algorithms remember its input due to an internal memory^[59]. For example, RNN are suitable for soil carbon dynamics modeled over many years (e.g., SCseq modeling after conversion from conventional to no-tillage or modeling of SCseq and global climate change).

Figure 4. AI models predicting soil carbon (SC) from environmental covariates. Idealized model representation of (A) feedforward artificial neural network (fANN) and (B) backward propagation artificial neural network (bANN).

Convolutional Neural Networks (CNN) is a DL AI method that was described in detail by^[27] for image, speech, and time series analysis. According to LeCun et al.^[26] (2015), DL discovers intricate structures in complex and large datasets by using a backpropagation algorithm. A DL architecture is a multilayer stack of simple modules, all (or most) of which are subject to learning, and many of which compute non-linear input-output mappings. Deep neural networks exploit the property that many natural signals are compositional hierarchies, in which higher-level features are obtained by composing lower-level ones. Importantly, CNNs use convolutional layers to detect local conjunctions of features from previous layers with the pooling layer merging semantically similar features into one. CNNs are suited for SOC predictive modeling from environmental covariates because they allow convolution filtering (e.g., a 3 × 3 window filter) and pooling of multiple layers. Each unit of the feature map is linked to local patches in the feature maps of the previous layer through a set of weights; and the local weight sum is emulated through a non-linear transfer function. CNN allow use of data augmentation to represent soils within a region, which can reduce overfitting and also improve prediction accuracy. Another benefit of CNN is to predict different soil depths simultaneously in a model inherently taking into account the depth correlation of soil attributes. This allows to improve the prediction of SOC or other soil properties in deeper layers, which has been a common problem in other soil modeling studies with ML algorithms^[30].

THE POWER, POTENTIALS, AND PERILS OF AI-BASED SOIL CARBON MODELING

There is no doubt that AI modeling provides advanced capabilities to predict SOC stocks and R_s. Though these AI models are still data limited in explaining the spatial variability of soil carbon storage within landscapes. Specifically, SOC measurements used at large region and global scale are derived from legacy databases with a smaller amount of data that represent current field conditions. The temporal mismatch between up-to-date environmental covariates and legacy SOC data may be another limiting factor. Proximal soil sensing (VNIR and MIR spectroscopy) has been pivotal to counter these trends and estimate SOC and other soil properties rapidly, cost-effectively, densely, and accurately through the application of AI. For example, SOC was estimated by AI from proximal sensing data at global scale by Viscarra Rossel et al.^[99] (2016), in Florida by Knox and Grunwald^[100] (2018), in regions in India by Clingensmith et al.^[101] (2019), in Brazil by Moura-Bueno et al.^[102] (2021), and in China by Shi et al.^[103] (2014). The incorporation of VNIR and MIR spectral data along with remote sensing data into AI models that upscale SOC storage to large regions is promising^[67].

Soil respiration observations have been integrated into global open-access databases^[87,97] to be shared and used by the scientific community. Regional R_s data can be spiked into these global databases which facilitates global research on AI-R_s. Soil respiration data represent carbon fluxes (i.e., temporal state of an ecosystem), while SOC storage infers on the spatially-explicit state of an ecosystem. AI models to predict SOC sequestration rates are still in its infancy due to data availability. One example, to model SCseq using CART in no-tillage systems compared to conventional tillage systems at global scale was provided by Sun et al.^[16] (2020). The open-access approach of global R_s data repositories differs from SOC data. The latter are limited by the access to data with due to different purpose: (1) regional AI-SOC research projects with up-to-date data that represent field conditions; (2) some national SOC data that have restricted access while others are public (e.g., U.S.); and (3) global public SOC data - for example, the World Soil Information Service (WoSIS) database - which includes more legacy data than up-to-date SOC data. In summary, some limitations for AI soil carbon modeling are due to limited data sharing and currency of data. Though the contribution of community SOC data into larger open-access global databases rests on fair data sharing policies that acknowledges the labor and costs involved in field and laboratory operations. Investments to collect new soil samples analyzed for SOC (topsoil and subsoil), consistent SOC monitoring at benchmark sites around the globe, and boosting of R_s measurements would greatly benefit future AI soil carbon modeling.

Furthermore, ethical concerns entail the amplified focus and reliance on AI technologies and machine-generated model outputs that lack knowledge discovery and human interpretation of soil carbon dynamics across large and complex soil-ecosystems^[33]. Wadoux et al.^[104] (2020) presented results that compared RF models created with real SOC observations and one from pseudo (“false”) variables for the same region. These AI models produced comparable results to predict SOC which raises major concerns about the possibility of AI generated “digital fake” versions of soil carbon storage. These concerns about AI models were echoed by Liao^[105] (2020) who pointed out that AI methods are prone to erratic behavior of model outputs due to outliers or misclassified pixels and are sensitive to pseudo (“false”) variables. As the collection of geospatial environmental datasets, specifically sensor-derived data, is steadily increasing the risk to incorporate spurious predictor variables into soil prediction models also increases^[33]. Data-driven AI modeling of soil carbon dynamics is prone to identify relations between massive and diverse datasets of input variables (environmental covariates) and outputs (SOC stock, R_s, or others) that may statistically exist, but from a physical or biogeochemical knowledge perspective make less sense.

Protection against such pitfalls is the application feature selection processing methods pre-AI modeling. Commonly used pre-processing methods are Recursive Feature Elimination (RFE) analysis to select the best performing subset of covariates which was described by Guyon et al.^[106] (2002). The RFE procedure starts with the maximum number of covariates and iteratively removes the weakest explanatory variable until a specified number of covariates is reached. Heuvelink et al.^[78] (2021) used RFE to identify covariates to model SOC stocks in Argentina using the QRF AI algorithm and Poggio et al.^[80] (2021) used RFE before running the QRF AI algorithm to model global SOC. The Boruta algorithm, often used in combination with RF, is another pre-processing method to strategically filter out the most important environmental covariates that relate most strongly to a target output^[107]. Boruta was applied successfully to build parsimonious RF-AI SOC prediction models that substantially reduced large environmental covariate sets^[37,69]. Xiong et al.^[37] (2014) compared various pre-processing algorithms that discerned all-relevant variables (i.e., strong and weakly relevant variables selected with Boruta and RF), minimal-optimal variables (four optimization algorithms were tested: greedy forward, greedy backward, hill climbing, and simulated annealing), and irrelevant environmental covariates to model SOC stock using four different AI methods (BaRT, BoRT, RF, and Cubist) in Florida, USA. The initial environmental covariate set comprised 210 variables, while the best performing parsimonious model identified with the all-relevant and minimal-optimal feature selection comprised just four covariates to predict SOC stock^[37]. This holistic AI environmental modeling framework was based on the consistent feature selections in ML approach developed earlier by Nilson et al.^[108] (2007). The advantage of automated feature selection compared to expert-based selection of covariates is that human bias is reduced that may, even unintentionally, impact SOC modeling and upscaling of results to region or global scale.

Another powerful feature selection methods is sparse Least Absolute Shrinkage and Selection Operator (LASSO)^[109]. For example, LASSO as well as Boruta feature selections were employed along with AI-ML and AI-DL algorithms to model SOC in two contrasting climatic regions^[110]. Wang et al.^[77] (2018) found that the genetic algorithm outperformed stepwise multivariate regression in strategically selecting predictor variables before applying RF and BoRT to model SOC stocks in rangeland in Eastern Australia. While the application of feature selections in AI-SOC modeling is prominent (see studies in Table 1), it seems relatively rare in global AI-R_s modeling studies. Interestingly, in global AI-R_s studies knowledge discovery is emphasized over technical information of AI modeling that often is only provided in small print in appendices and supplementary documents of publications. Instead, separate pre- or post-hoc analyses that complement AI modeling are commonly found in the global R_s literature^[111].

The dichotomy between data-driven and knowledge-driven soil modeling was discussed in detail by Wadoux et al.^[104] (2020). Authors cautioned about knowledge discovery purely from ML and pattern recognition processes. It was suggested that pedologically relevant environmental covariates should be selected through feature selection pre-processing. In similar vein, McBratney et al.^[28] (2019) raised concerns around over-parameterization and the generation of nonsensical soil predictions from AI models. In my view, an important vision going forward with AI soil carbon modeling is to keep balance between (1) data-driven feature selection and ML and DL modeling; and (2) knowledge-driven approaches pre-AI modeling and post-interpretation. Scientific interpretation enhances legitimacy and confidence in AI-generated digital soil C output. Ideally, an integral scientific approach involves consultation of multiple other sources (environmental datasets, literature, expert-knowledge) and comparative analysis derived from other methods rooted in a modeling paradigm different from AI (e.g., process-based simulation modeling, geostatistics, hybrid stochastic-deterministic methods, Bayesian methods, structural equation modeling, participatory action research). Ensemble modeling to aggregate soil carbon output from various, ideally contrasting model paradigms among them AI, may lower the risks of spurious AI model output. Meta-modeling, an approach rooted in integral ecology, was envisioned as a viable framework for integration of multiple models and data to address soil security issues, among them soil carbon sequestration^[112]. An integrative vision for soil carbon assessments would avoid the inflated hype about AI in carbon sciences. Such integrative strategy lowers the risk to “blindly” belief machine-generated soil carbon assessments; even validation of AI-generated results cannot fully inoculate from potential spurious modeling results that may be replicated in training and validation modes. Honoring the diversity of modeling approaches that provide partial knowledge of soil carbon dynamics rather than idealizing AI as superior to all other methods will further enhance scientific understanding of complex soil-ecosystems and carbon dynamics. It may also help to form resilient liaisons and partnerships among AI specialists and soil and environmental scientists.

The shift from simple ML rooted in pattern recognition toward more complex DL models with multiple layers of nodes, processing strategies (e.g., convolution and pooling), and fitting strategies (e.g., latent factors or weights) makes these kinds of models (see Figures 2-5) more abstract losing more-and-more physical and pedological meaning. Theoretically, these fitting and learning strategies in ANN model variants if put to the extreme could achieve an ideal model fit (R² of 1), which has been approximated already in soil carbon modeling applications. For example, CNN and Cubist were used to model various soil properties, among them SOC, using VNIR and MIR spectral soil data using a large dataset (n = 14,594) from the U.S.^[113]. In this study, the two-channel 1D CNN model was best performing with R² between 0.95 and 0.98 for six different soil properties, the R² was 0.98 for both SOC and soil total carbon, and the RPIQ was 2.27 (SOC) and 3.01 (soil total carbon). These results are “near-perfect” though one may wonder about the many hyperparameters and processing layers in the CNN for tuning to achieve such superb model performance [Figure 5]. The many hyperparameters, latent factors, and fitting weights in AI models make sense to the machine, but are less meaningful for interpretation by human users or carbon scientists to infer on carbon cycle processes, soil functions, or ecosystem services. What pedological or biophysical insights in regard to SOC or ecosystem processes were derived in Ng et al.^[113] (2019) research study or similar AI-DL soil carbon models? The extraordinary capabilities of AI-DL algorithms to fit inputs and outputs have been hailed black-boxes and AI-ML algorithms gray boxes, respectively; in essence, AI models lack transparency^[105]. Black-boxes or gray-boxes mean, for example, that SOC storage or the ecosystem process of SCseq are encoded in AI-ANN models in form of multiple nodes, layers, and weighing factors replacing human understanding and striving for meaning-making about the soil-environment into machine code.

Figure 5. AI model predicting soil carbon (SC) from environmental covariates. Idealized model representation of a convolutional neural network (CNN).

From a philosophical perspective, the question is whether the physical environment or a simulated, virtual environment (digital worlds) is more real to us. Chalmers^[114] (2022) in his book Reality+ discerned between virtual simulated worlds and those we perceive as real and suggested that we can live meaningful lives in virtual reality. Applying Chalmer’s vision to carbon science this suggests that we would be able to live meaningful lives in machine generated worlds in which a soil carbon map or model is as real and satisfying as a soil in nature. Interestingly, in such a virtual/AI soil carbon world human knowledge and understanding of soil-environmental relations, mechanisms, ecosystem processes, carbon fluxes and cycling, and global climate change become irrelevant.

Given the rapid expansion of AI into carbon science, as well as many other sciences, poses urgency to think about ethical implications implicated in AI^[105,115]. Reality+ confronts us with the question what is “real and meaningful to us” - an observable soil in nature that we can touch, sense, and use (phenomenology), laboratory measurements of the soil carbon content (empiricism), proximal or remote sensors and AI providing inference on soil carbon, a digital image/map of soil and its carbon storage computed by AI (representation), or simulated worlds (simulacra) created with advanced AI and visualization techniques. These simulacra replace “environmental reality” with its representation according to Baudrillard^[116] (1994). He philosophized that society had become saturated with simulacra and that people live in a hyperreality in which meaninglessness prevails. While Baudrillard^[116]’s (1994) vision was somewhat dystopian, Chalmer^[114]’s (2022) Reality+ looks more optimistic. Whether the expansion of AI-DL and AI-ML soil carbon modeling is perceived as frustrating and frightening because only the machine knows leaving one confused, helpless, and meaningless or whether we get excited and enchanted by the beauty of machine-generated soil carbon data, maps, and models that we trust will have a profound impact on carbon science and how it is applied in carbon policies, carbon crediting, and carbon management. The risks involved are that AI-generated soil carbon hyperreality is prone to human manipulation (i.e., how the model is tuned and fitted) and misuse.

One key question is whether we are applying AI in data-driven or knowledge-driven ways to advance soil carbon science. The potentials and perils of AI in carbon science need to be carefully weighted to avoid pitfalls, and perhaps compute (surprising), spurious digital soil carbon predictions. The power of AI is the possibility of more accurate and precise soil carbon models that only machine algorithms can create. In the latter lies the profound potential of AI-ML and AI-DL to transform carbon science and modeling. Enhanced dialogue and awareness of AI model limitations may help to better understand soil carbon evolution and its ecosystem processes.