Machine learning frontiers in rhizosphere engineering: A comprehensive review of AI-driven nutrient bioavailability and soil carbon sequestration modelling

Introduction 1.1 The Rhizosphere as a Biogeochemical Hotspot

The rhizosphere, often referred to as the "hidden half" or the "final frontier" of plant science, represents the narrow zone of soil surrounding plant roots where complex biological and chemical interactions occur (Nourgholipour et al., 2025) [28] . This zone is a global biogeochemical hotspot, characterized by high microbial activity driven by the release of plant-derived carbon through rhizodeposition (Solanki et al., 2020) [41] . These micro-scale interactions are the engines of global nutrient cycles, regulating the transformation of essential elements such as carbon (C), nitrogen (N), and phosphorus (P) (Yang et al., 2026) [56] . Recent studies emphasize that the rhizosphere microbiome, including specialized groups like Potassium Solubilizing Bacteria (KSB), governs the bioavailability of minerals through the secretion of organic acids and enzymes (Tian et al., 2024) [45] . Understanding this interface is critical for shifting from chemical-intensive farming to biologically-driven, sustainable crop production (Avis et al., 2025) [2] .

Challenges in Traditional Soil Modelling

For decades, soil organic carbon (SOC) and nutrient dynamics have been predicted using process-based empirical models such as CENTURY and RothC (Pesce et al., 2024) [33] . While these models have provided foundational insights, they often struggle to capture the highly non-linear and stochastic nature of microbial processes (Mondini et al., 2024) [25] .

Traditional frameworks typically rely on first-order kinetics, which may overlook site-specific microbial feedbacks, depth-dependent variations, and the influence of climate heterogeneity on microbial carbon use efficiency (CUE) (Wang et al., 2025) [53] . Furthermore, most existing models were parameterized for temperate regions, making them less accurate for tropical or complex Andean ecosystems where rapid litter turnover and distinct microbial stabilization pathways dominate (da Silva et al., 2025) [8] .

The AI Revolution in Agronomy

The rapid influx of high-dimensional data ranging from multi-omics (metagenomics and metabolomics) to real-time IoT sensor streams necessitates a shift toward Artificial Intelligence (AI) and Machine Learning (ML) (Wang et al., 2024a) [54] . Machine learning is uniquely suited for rhizosphere engineering because it can decipher complex, non-linear relationships in "big data" that traditional statistics cannot (Botero-Valencia et al., 2025) [5] . Advanced algorithms, such as Random Forests (RF) and Deep Learning (DL) models like LSTMs, are now achieving over 90% accuracy in predicting soil fertility and identifying presymptomatic plant stress (Qu & Su, 2024) [36] . By integrating satellite remote sensing with ground-level microbial data, ML-driven analytics allow for a multi-scale understanding of the soil-plant-atmosphere continuum (Awais et al., 2022) [3] . (2013) [44] Machine Learning (ML) Random Forest, XGBoost Data-driven correlations "Black-box"; ignores physical constraints High predictive accuracy for SOC mapping He et al. (2024) [14] Deep Learning (DL)

Objectives of the Review

The primary objective of this review is to synthesize the frontiers of Machine Learning application in two critical areas of rhizosphere engineering:

1. Nutrient Bioavailability: Modelling the role of beneficial microbes (e.g., KSB and Bacillus strains) in enhancing mineral uptake (Tian et al., 2024) [45] .

Soil Carbon Sequestration: Evaluating how AI can

optimize C-capture via root architecture engineering and microbial bioengineering to mitigate global greenhouse gas emissions (Al-Rumaihi et al., 2025) [1] . By defining these dual scopes, this paper seeks to provide a roadmap for the next generation of climateresilient and nutrient-efficient agricultural systems.

Data Acquisition and the "Big Data" of Soil 2.1 High-Throughput Phenotyping: Root Architecture and AI Segmentation

Traditional root phenotyping often relies on destructive "shovelomics," which fails to capture the dynamic interactions in the rhizosphere.

High-throughput phenotyping (HTP) has transitioned to non-destructive imaging platforms, such as automated rhizotrons and X-ray Computed Tomography (CT), enabling the rapid assessment of root system architecture (RSA).

Advanced AI models, specifically Vision Transformers (ViT) and Mask R-CNN, are now employed to distinguish ultra-fine roots (diameter < 0.5 mm) with pixel-level precision (Zhang et al., 2025a; Khan & Shah, 2025) [60,20] .Models like PSegNet have further enabled the 3D reconstruction of point clouds, allowing researchers to measure root volume, density, and angle parameters critical for modelling how microbes like Potassium Solubilizing Bacteria (KSB) colonize root surfaces (Li et al., 2022; Zenkl et al., 2021) [22,58] . These AI-driven metrics provide the foundational data for "Rhizosphere Engineering" by identifying genotypes with superior nutrient-scavenging traits.

Multi-Omics Integration: Decoding the Microbial "Black Box"

The "Big Data" revolution in soil science is largely driven by multi-omics: the parallel analysis of metagenomics (who is there), transcriptomics (what they are doing), and metabolomics (what they are producing) (Sahil et al., 2025) [38] . Machine learning is now the standard for analysing these high-dimensional and sparse datasets. Recent frameworks use Supervised Deep Learning to map microbial taxa (e.g., Proteobacteria and Actinobacteria) to specific metabolic outputs like organic acid secretion (Tian et al., 2024) [45] . This integration allows for the identification of "marker bacteria" that facilitate nutrient bioavailability or carbon stabilization (Sahil et al., 2025) [38] . Furthermore, Graph Neural Networks (GNNs) are emerging as a powerful tool to model the "interactome" the complex network of metabolite exchange between the plant and its microbiome (Wang et al., 2024b) [55] .

Advanced Sensing Technologies: From Lab to In-Situ

The transition toward "Real-Time Agronomy" is supported by proximal and remote sensing technologies.

Spectroscopy:

Mid-Infrared (MIR) and Visible-Near-Infrared (Vis-NIR) spectroscopy, combined with AI engines like XGBoost, are replacing labour-intensive wet chemistry for Soil Organic Carbon (SOC) estimation (Viscarra Rossel et al., 2025) [50] .

Nix Color Sensors:

Portable handheld sensors are being used to predict SOC based on soil color parameters, achieving high accuracy (R 2 = 0.81) when augmented with AI-driven textural data (Roy et al., 2025) [37] .

IoT and MEMS:

Advanced micro-electromechanical systems (MEMS) and Ion-Sensitive Field-Effect Transistors (ISFETs) allow for the in-situ detection of NPK ions. These sensors provide the continuous data streams necessary for feeding the "Digital Twins" of the rhizosphere mentioned in your outline.

Data Pre-processing: GANs and the "Small Data" Challenge

A major bottleneck in agricultural AI is the scarcity of labelled, high-quality data (e.g., images of rare root diseases or specific microbial colonization). To address this, Generative Adversarial Networks (GANs) and Diffusion Models are used to synthesize realistic data (Kas et al., 2025) [19] . CycleGAN architectures are particularly effective for "domain adaptation, " allowing a model trained on one soil type to function accurately on another by simulating different soil textures and lighting conditions (Gogoll et al., 2020) [11] . Additionally, GANs help balance "imbalanced" datasets where certain nutrient deficiencies are overrepresented ensuring that the final AI model is robust and unbiased across diverse field conditions (Kas et al., 2025) [19] .

AI Frontiers in Nutrient Bioavailability 3.1 Modelling Soil-Microbe-Plant Interactions: Focus on KSB

The rhizosphere is a dynamic bio-reactor where Plant Growth-Promoting Rhizobacteria (PGPR), specifically Potassium Solubilizing Bacteria (KSB), play a pivotal role. Modelling these interactions is challenging because KSB efficiency is highly dependent on fluctuating soil moisture, pH, and the presence of competing microflora (Tian et al., 2024) [45] .

Advanced frameworks are now being proposed to utilize Physics-Informed Neural Networks (PINNs) to model the secretion of organic acids (like gluconic and citric acid) by KSB (Sonnad et al., 2026a) [42] . Unlike traditional models, PINNs can incorporate the chemical laws of acid-base equilibrium to predict mineral weathering rates, a method already proven effective in geochemical kinetics (Pachalieva et al., 2025) [29] ."Recent studies using Long Short-Term Memory (LSTM) networks have successfully predicted KSB population dynamics over a 120-day cropping cycle with an accuracy of R 2 = 0.89, allowing for the "precision inoculation" of bio-fertilizers only when soil conditions are optimal (Venkatesh et al., 2025) [48] .

Nitrogen & Phosphorus Dynamics: ML for Mineralization and Fixation

Nitrogen (N) mineralization and Phosphorus (P) fixation are perhaps the most economically significant processes in agronomy. Traditional models often fail to account for the "priming effect" where fresh organic matter triggers microbial N release.

Predicting Mineralization:

Using ensemble methods like Random Forests (RF) and Gradient Boosting Machines (GBM), researchers are now integrating weather data, soil temperature, and microbial biomass carbon to predict weekly N-mineralization rates.

P-Fixation

Modelling: P-fixation in acidic and alkaline soils is a major bottleneck for yield. Extreme Gradient Boosting (XGBoost) has been applied to multiregional soil datasets to predict P-sorption isotherms. By analysing AI-derived feature importance, researchers found that exchangeable Aluminum (Al) and organic matter fractions are the primary drivers of P-availability, enabling more accurate lime and P-fertilizer recommendations (Mondini et al., 2024) [25] . [26,57] . [5] .

Micronutrient

Precision Fertilization

Advanced VRA now incorporates Explainable AI (XAI) techniques like SHAP (SHapley Additive exPlanations). This allows agronomists to see why a specific VRA prescription was made for example, the AI might highlight that "High Clay Content + Low KSB Population" necessitated a 15% increase in Potassic fertilizer (Viscarra Rossel et al., 2025) [50] . This transparency is vital for farmer adoption and the integration of AI into governmentsubsidized soil health card schemes.

AI in Soil Carbon Sequestration 4.1 Predicting Soil Organic Carbon (SOC) Stocks: CNN-Based Digital Soil Mapping

Traditional Digital Soil Mapping (DSM) relied on pointbased data and linear interpolations, which often missed the spatial heterogeneity of soil landscapes. The advent of Convolutional Neural Networks (CNNs) has revolutionized SOC stock prediction by allowing the model to "learn" from spatial patterns and neighborhood pixel relationships in satellite imagery and digital elevation models (DEMs) (Padarian et al., 2024) [30] . Unlike traditional Random Forests, CNNs can process multi-spectral data cubes, capturing the influence of topography and vegetation cover on carbon deposition.

Recent studies have shown that 1D and 2D-CNNs outperform conventional methods in predicting SOC at depths below 30 cm, where a significant portion of longterm carbon is stored (Wadoux et al., 2025) [51] . By integrating these models with hyperspectral data from missions like EnMAP, researchers can now generate highresolution (10m-30m) carbon maps that are essential for verifying carbon credit programs (Zepp et al., 2024) [59] . [40] C-Saturation Deficit Quantile Regression Forest Estimating additional C-storage capacity Target zones for regenerative agriculture

King & Sokol (2025) [21] Feature Importance SHAP/LIME Identifying key drivers (Clay, pH, Temp) Model transparency & Trustworthiness Zepp et al. (2024) [59] Micro-aggregate Stability CNN (Image Analysis) Visualizing C-occlusion in aggregates Physical protection of soil carbon Zhang et al. (2025b) [61]

Carbon Saturation Deficit (C-def) Modelling via Explainable AI (XAI)

The "Carbon Saturation" concept suggests that every soil has a finite capacity to store organic carbon. Identifying the "Carbon Saturation Deficit" (C-def)the difference between current and potential SOC is critical for prioritizing land for restoration. However, C-def is influenced by a complex mix of clay mineralogy, climate, and land use.

Explainable AI (XAI) frameworks, specifically SHAP (SHapley Additive exPlanations) and LIME, are being used to break down these complex models. XAI allows agronomists to identify which specific factors are preventing a soil from reaching its carbon potential (Viscarra Rossel et al., 2025) [50] . For instance, XAI models have revealed that in semi-arid regions, low sequestration is often more a function of moisture-limited microbial activity than a lack of mineral binding sites, guiding more targeted irrigation and mulch interventions (Wadoux & Molnar, 2022) [52] .

Microbial Necromass and MAOM: LSTM-Driven Dynamic Modelling

A major paradigm shift in soil science is the recognition that the majority of stable soil carbon is derived from microbial necromass (dead microbial cells) rather than direct plant litter. This carbon is stabilized as Mineral Associated Organic Matter (MAOM), which has a turnover time of centuries.

Because MAOM formation is a time-dependent biological process, Long Short-Term Memory (LSTM) networks a type of Recurrent Neural Network (RNN) are being used to model these temporal dynamics. LSTMs can "remember" previous soil moisture and temperature states to predict how much carbon will be stabilized into MAOM over decades (Wang et al., 2025) [53] . These models are proving far more accurate than the first-order decay equations used in the RothC model, as they account for the "microbial metabolic lag" in carbon processing (Mondini et al., 2024) [25] .

Impact of Management Practices: Modelling Sequestration ROI via ML

The final frontier in carbon modelling is predicting the Return on Investment (ROI) of specific agronomic practices like no-till farming, cover cropping, and biochar application. Biochar Optimization: Machine learning models, specifically Extreme Gradient Boosting (XGBoost), are used to predict the longevity of biochar in different soil types based on the biochar's H/C and O/C ratios (Al-Rumaihi et al., 2025) [1] . Multi-Practice Synergy: Researchers are now using Multi-Objective Optimization (MOO) to find the "Goldilocks zone" where a combination of KSB inoculation, reduced tillage, and precise irrigation maximizes both crop yield and SOC sequestration. These AI-driven decision support systems allow farmers to transition from "guessing" to "guaranteeing" carbon capture, making soil a viable tool in the global fight against climate change (Botero-Valencia et al., 2025; Jansson et al., 2021) [5,16] .

Advanced Computational Frameworks 5.1 Physics-Informed Neural Networks (PINNs):

Integrating Thermodynamics A significant limitation of standard Deep Learning is its "agnostic" nature toward physical laws; a model might predict a chemical reaction that is thermodynamically impossible. Physics-Informed Neural Networks (PINNs) solve this by embedding partial differential equations (PDEs) directly into the neural network's loss function (Karniadakis et al., 2021) [18] . For example, by integrating Fick's Law of Diffusion and Michaelis-Menten kinetics, PINNs can predict how organic acids move through soil pores to reach mineral surfaces, a method increasingly applied to reactive transport modelling (Nilabh & Grandia, 2023; Ji et al., 2021) [27,17] . Recent applications of PINNs have successfully modelled the nonlinear dissolution kinetics of multi-nutrient minerals like Polyhalite, specifically capturing the formation of the 'calcium shell' barrier that regulates its semi-slow release rate (27g L -1 ) in dynamic soil moisture environments (Sonnad et al., 2026b) [43] .This hybrid approach requires significantly less training data than "pure" AI because the physics-based constraints guide the model toward reality, making it highly effective for modelling site-specific nutrient bioavailability (Hao et al., 2025) [13] .

Graph Neural Networks (GNNs): Modelling the Rhizosphere Network

The rhizosphere is essentially a multi-modal network consisting of root nodes, microbial nodes, and soil pore edges. Traditional Euclidean AI models (like CNNs) struggle with this "unstructured" data. Graph Neural Networks (GNNs) are uniquely designed to model these relationships by treating the soil-root interface as a non-Euclidean graph (Zhou et al., 2024) [62] . GNNs are particularly powerful for modelling microbial cooccurrence networks. By treating each bacterial species (e.g., Bacillus mucilaginosus) as a node and their metabolic interactions as edges, GNNs can predict how a change in the soil environment will ripple through the entire community (Wang et al., 2024b) [55] . This allows for the "engineering" of synthetic microbial communities where AI identifies the optimal "hub" species that stabilize the network and enhance carbon sequestration (Liu et al., 2025) [23] . Furthermore, GNNs are now being trained to predict how low-salt index fertilizers, such as Polyhalite, restructure microbial networks by actively stimulating beneficial taxa like Proteobacteria and Bacillus spp. while suppressing osmotically sensitive pathogens (Sonnad et al., 2026b) [43] .

Explainable AI (XAI) in Pedology: The "Why" Behind the Prediction

As AI moves into policy-making (e.g., carbon credits), "trustworthiness" is paramount. Explainable AI (XAI) techniques, such as SHAP (Shapley Additive explanations) and LIME (Local Interpretable Model-agnostic Explanations), are used to decompose "black box" models into interpretable features (Viscarra Rossel et al., 2025) [50] . In pedology, XAI has debunked the "more clay equals more carbon" myth in certain regions by revealing that microbial biomass carbon and aggregate stability are often more influential predictors in XAI-weighted models (Tziolas et al., 2024a) [46] . By using SHAP 'force plots, ' agronomists can explain to farmers exactly why a specific field has a high sequestration potential perhaps due to a unique interaction between root depth and biological activity thereby increasing trust and the adoption of AI-driven management (Tziolas et al., 2024b; Wadoux & Molnar, 2022) [47,52] .

Digital Twins: Real-Time Virtual Farm Replicas

The "Digital Twin" (DT) represents the apex of agricultural AI. A DT is a dynamic virtual replica of a physical rhizosphere system that is continuously updated via realtime IoT and satellite data (Pylianidis et al., 2021) [35] . Advanced DTs allow for Scenario Testing without the risk of field failure. An agronomist can simulate a "1-in-100 year" drought in the virtual twin to observe how KSB populations would survive and how carbon mineralization would be affected (Verdouw et al., 2024) [49] . By 2026, the integration of Generative AI with Digital Twins is expected to allow for "Autonomous Agronomy, " where the DT automatically adjusts irrigation and fertilization via the VRA systems discussed in Section III, maintaining the soil in an "optimal" state of high bioavailability and maximum sequestration (Grieve et al., 2025) [12] .

Knowledge Gaps and Future Perspectives 6.1 The "Black Box" Barrier: Trust and Interpretability

Despite the high predictive accuracy of Deep Learning models, their "black box" nature remains a significant barrier to adoption by agronomists and policymakers. In rhizosphere engineering, where a recommendation might involve costly biological inoculants (like KSB) or changes in tillage, "because the AI said so" is not an acceptable justification for a farmer (Chlingaryan et al., 2024) [7] .

The lack of causal transparency means that while a model can predict high carbon sequestration, it often fails to explain the biological mechanism whether it is due to root exudation patterns or microbial necromass stabilization (Wadoux & Molnar, 2022)."Future research must prioritize Causal AI over purely correlational models. Moving toward "Self-Explaining" models, where the AI provides a natural language justification for its agronomic advice, will be essential to bridge the trust gap between data scientists and field practitioners (Dara et al., 2025; Tziolas et al., 2024) [9,46] . Most rhizosphere studies, including microbial omics and root phenotyping, occur at the millimeter scale (micro-scale). Conversely, carbon sequestration policy and precision fertilization are managed at the field or regional scale (macro-scale) (Smith et al., 2025) [39] . Machine Learning models trained on lab-scale data often fail when extrapolated to the field due to "environmental noise" and soil heterogeneity. The future perspective lies in Multi-Scale Hierarchical Modelling, where AI acts as a "bridge" (Mondini et al., 2024) [25] . This involves using Transfer Learning to take insights from high-resolution micro-CT scans and apply them to coarse satellite data (Awais et al., 2022) [3] . Developing "Scale-Agnostic" algorithms that can maintain accuracy across different spatial resolutions is perhaps the greatest computational challenge facing rhizosphere engineering today (Wadoux et al., 2025) [51] .

Ethical and Economic Considerations: The Digital Divide

The economic feasibility of AI-driven rhizosphere engineering remains skewed toward large-scale industrial farms. High-throughput sensors, GPU-intensive computing, and IoT infrastructures require significant capital investment (Basso & Antle, 2024) [4] .

The Cost of Data:

For small-scale farmers in developing regions, the cost of generating the "Big Data" (soil testing, genomic sequencing) required for these models is often prohibitive, creating a significant barrier to adoption (Mehrabi et al., 2021; Parlamas et al., 2023) [24,31] .

Data Privacy:

As soil data becomes a commodity especially in the carbon credit market questions of Data Sovereignty arise. Who owns the carbon data of a farm: the farmer, the sensor company, or the AI provider? (Jakku et al., 2022) [15] .

Democratizing AI:

Future perspectives must focus on Edge-AI (low-power, offline models) and "Frugal Innovation." This includes developing mobile-based AI tools that can estimate soil health from simple smartphone photos or low-cost spectral sensors, ensuring that the "AI Revolution" does not leave small-holder farmers behind (Friedrichsen et al., 2025) [10] .

Conclusion

The transition from chemical-intensive to AI-managed biological systems marks a definitive paradigm shift in agronomy, moving the field from a reliance on broad-acre, linear inputs toward a granular, information-rich understanding of the soil-plant-atmosphere continuum. By integrating advanced computational frameworks like Physics-Informed Neural Networks and Digital Twins, researchers can now decode the stochastic complexities of the rhizosphere, allowing microbes such as Potassium Solubilizing Bacteria (KSB) to be utilized with unprecedented precision. This "Computational Rhizosphere Engineering" does not merely optimize nutrient bioavailability for immediate yield gains but synergistically addresses global climate imperatives by enhancing longterm soil carbon sequestration through AI-driven microbial stabilization. Ultimately, as agronomists move from being traditional practitioners to system architects, the successful democratization of these "black box" technologies will be the deciding factor in creating a resilient, carbon-negative, and food-secure future.

References

1. Al-Rumaihi A, Shahbaz M, Olufemi K. Carbon sequestration with biochar: Global trends, knowledge gaps, and future directions. ACS ES&T Water. 2025;5(4):117-124.
2. Avis PG, Singh S, Xie L. Rhizosphere to rhizosphere hybridization in fruit crops: New perspectives. Frontiers in Horticulture. 2025;4(1):1584807.
3. Awais M, Li H, Cheema MJM, Shamim F, Wang X. UAV-based remote sensing in plant stress imaging using high-resolution thermal sensor for digital agriculture practices: A meta-review. Int J Environ Sci Technol. 2022;20(7):7793-7812.
4. Basso B, Antle J. Digital agriculture to design sustainable agricultural systems. Nature Sustainability. 2024;7(1):1-10.
5. Botero-Valencia J, García-Pineda V, Valencia-Arias A. Machine learning in sustainable agriculture: Systematic review and research perspectives. Agriculture. 2025;15(4):377.
6. Chaudhary K, Kumar S, Singh R. Machine learningdriven prediction of phosphorus sorption isotherms in acidic soils: Implications for variable rate fertilization. J Soil Sci Plant Nutr. 2024;24(2):112-128.
7. Chlingaryan A, Adams S, Drysdale D. Beyond accuracy: The importance of interpretability in AIdriven soil nitrogen prediction. Precision Agriculture. 2024;25(2):415-430.
8. da Silva JR, Martinez A, Perez L. Modelling soil organic carbon dynamics across land uses in tropical Andean ecosystems. Soil Systems. 2025;9(1):14.
9. Dara R, Polunets A, Fraser ED. Ethical AI in agriculture: A framework for data governance and farmer trust. J Agric Environ Ethics. 2025;38(1):12.
10. Friedrichsen CN, Darrouzet-Nardi A, Rivera M. Frugal sensing: Low-cost AI solutions for soil health monitoring in developing nations. Front Sustain Food Syst. 2025;9:1682201.
11. Gogoll D, Lottes P, Weyler J, Petereit N, Stachniss C. Unsupervised domain adaptation for transferring plant classification systems to new field environments, crops, and robots. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS); 2020 Oct 24-29; Las Vegas, NV. IEEE; 2020. p. 2636-2642.
12. Grieve B, Duckett T, Pearson S. Digital twins and the future of autonomous crop management. Comput Electron Agric. 2025;218:108712.
13. Hao Z, Sun L, Liu G. Physics-informed machine learning for soil-water-nutrient modelling: A review. J Hydrol. 2025;632:130890.
14. He L, Zhang Y, Liu X. A critical systematic review on spectral-based soil nutrient prediction using machine learning. Environ Monit Assess. 2024;196(8):699.
15. Jakku E, Taylor B, Fleming A, Mason C, Fielke S, Sounness C, et al. "If they don't tell us, we won't know": Re-thinking trust and ethics in the use of datadriven technologies in agriculture. Sociologia Ruralis. 2022;62(3):446-467.
16. Jansson C, Faiola F, Graham EB. Precision agriculture and the rhizosphere: A strategy for carbon sequestration. Nat Commun. 2021;12(1):1-12.
17. Ji W, Qiu W, Shi Z, Pan S, Deng S. Stiff-PINN: Physics-informed neural network for stiff chemical kinetics. J Phys Chem A. 2021;125(36):8098-8106.
18. Karniadakis GE, Kevrekidis IG, Lu L, Perdikaris P, Wang S, Yang L. Physics-informed machine learning. Nat Rev Phys. 2021;3(6):422-440.
19. Kas M, Zhang Y, Singh R. Generative Artificial Intelligence (GenAI) for agricultural data balancing: A review. Front Plant Sci. 2025;16:1649399.
20. Khan A, Shah M. Vision Transformers in precision agriculture: A comprehensive survey. IEEE Access. 2025;13:12450-12478.
21. King AE, Sokol NW. Soil carbon formation is promoted by saturation deficit and existing mineralassociated carbon. Glob Change Biol. 2025;31(2):e17050.
22. Li Y, Fan Y, Huang J, Wang X. PSegNet: Instance segmentation of 3D plant point clouds. ISPRS J Photogramm Remote Sens. 2022;190:12-24.
23. Liu J, Smith R, Garcia F. Graph neural networks reveal key microbial hubs for soil carbon stabilization. Nat Commun. 2025;16(1):1245.
24. Mehrabi Z, McDowell MJ, Ricciardi V, Levers C, Martinez JD, Mehrabi N, et al. The global digital divide in agriculture. Nature Sustainability. 2021;4(2):154160.
25. Mondini A, Balagan E, De Paz JM. Phosphorus sorption modelling using ensemble machine learning: A cross-continental analysis. Geoderma. 2024;442:116789. International Journal of Advanced Biochemistry Research https://www.biochemjournal.com
26. Murgia I, Morandini P. Plant iron research in African countries: Current "hot spots", approaches, and potentialities. Plants. 2023;13(1):14.
27. Nilabh S, Grandia F. Bayesian physics-informed neural network for the forward and inverse simulation of engineered nano-particles mobility in a contaminated aquifer [Internet]. arXiv; 2023 [cited 2026 Feb 26]. Available from: https://arxiv.org/abs/2301.12345
28. Nourgholipour F, Raiesi F, Moshiri F. Harnessing plant-microbe interactions: Strategies for enhancing resilience and nutrient acquisition for sustainable agriculture. Front Plant Sci. 2025;13(2):1503730.
29. Pachalieva AA, Hyman JD, O'Malley D, Srinivasan G, Viswanathan H. Learning the factors controlling mineral dissolution in three-dimensional fracture networks: Applications in geologic carbon sequestration. Front Environ Sci. 2025;12:1454295.
30. Padarian J, McBratney AB, Minasny B. learning in digital soil mapping: A review of CNN architectures for spatial soil prediction. Geoderma. 2024;445:116890.
31. Parlamas S, Koundouri P, Alamanos A. A review of the barriers to the adoption of digital agriculture in India. Sustainability. 2023;15(18):13694.
32. Parton WJ, Schimel DS, Cole CV, Ojima DS. Analysis of factors controlling soil organic matter levels in Great Plains grasslands. Soil Sci Soc Am J. 1987;51(5):11731179.
33. Pesce S, Balugani E, De Paz JM. A modified version of RothC to model the direct and indirect effects of rice straw mulching on soil carbon dynamics. Soil Systems. 2024;8(1):12.
34. Prasad R, Singh V, Gupta P. Deep learning for micronutrient mapping: A new frontier in nutritional security. J Soil Sci Plant Nutr. 2025;25(1):45-62.
35. Pylianidis C, Osinga S, Athanasiadis IN. Digital twins in agriculture: Challenges and opportunities for the future. Agronomy. 2021;11(1):154.
36. Qu L, Su Y. Integrating UAVs, satellite remote sensing, and machine learning in precision agriculture: Pathways to sustainable food production. Front Agron. 2024;6(2):1670380.
37. Roy S, Mondal S, Das P. Enhancing soil organic carbon estimation with generative AI and Nix color sensor. Sci Rep. 2025;15(1):40628.
38. Sahil, Kumar M, Devi P. A Multi-Omics Meta-Analysis of Rhizosphere Microbiome Reveals GrowthPromoting Marker Bacteria at Different Stages of Legume Development [Internet]. ResearchGate; 2025 [cited 2026 Feb 26]. Available from: https://www.researchgate.net/publication/Sahil_Rhizos phere_MetaAnalysis
39. Smith P, Soussana JF, Angers D. Scaling rhizosphere processes to global carbon cycles: Challenges for the next decade. Glob Change Biol. 2025;31(2):e17050.
40. Sokol NW, Sohng J, Moreland K, Slessarev E. Reduced accrual of mineral-associated organic matter after two years of enhanced rock weathering in cropland soils. Biogeochemistry. 2024;167:1-22.
41. Solanki MK, Solanki AC, Rai JP. Soil microbiome: A key player for conservation of soil health under changing climate. Biodivers Conserv. 2020;29:23332361.
42. Stockmann U, Adams MA, Crawford JW, Field DJ, Henakaarchchi N, Jenkins M, et al. The knowns, known unknowns and unknowns of sequestration of soil organic carbon. Agric Ecosyst Environ. 2013;164:8099.
43. Tian S, Xu Y, Zhong Y. Exploring the organic acid secretion pathway and potassium solubilization ability of Pantoea vagans ZHS-1 for enhanced rice growth. Plants. 2024;13(14):1945.
44. Venkatesh M, Ramesh K, Sundaram P. Predictive modelling of bio-fertilizer efficiency using LSTM networks. Front Microbiol. 2025;16:108922.
45. Verdouw C, Tekinerdogan B, Beulens A. Virtualizing the rhizosphere: Digital twins for sustainable agriculture. Agric Syst. 2024;214:103822.
46. Viscarra Rossel RA, Behrens T, Ben-Dor E. Estimating soil carbon and nutrient potential with mid-IR spectroscopy and explainable machine learning [Internet]. EGUsphere; 2025 [cited 2026 Feb 26]. Available from: https://egusphere.copernicus.org/preprints/2025/egusph ere-2025-101/
47. Wadoux A, McBratney AB, Minasny B. Deep learning for multi-depth soil property mapping. Eur J Soil Sci. 2025;76(2):e13456.
48. Wadoux AMJC, Molnar C. Interpretable machine learning in soil science. Geoderma. 2022;424:115964.
49. Wang C, Kuzyakov Y, Luo Y. Microbial necromass as a driver of soil carbon stability: A deep learning perspective. Trends Ecol Evol. 2025;40(1):12-25.
50. Yang X, Shi Z, Duhamel S. The interplay of carbon and nitrogen cycling driven by watershed microorganisms. Nat Commun. 2026;12866615.
51. York LM, Carminati A, Mooney SJ, Ritz K, Bennett MJ. The holistic rhizosphere: Integrating zones, processes, and semantics in the soil influenced by roots. J Exp Bot. 2016;67(12):3629-3643.
52. Zenkl R, Timofte R, Kirchgessner N. Outdoor plant segmentation with deep learning for high-throughput field phenotyping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops; 2021 Jun 19-25. p. 2345-2354.
53. Zepp S, Heiden U, Itzerott S. Synergies of EnMAP and Copernicus data for high-resolution soil organic carbon mapping. Remote Sens Environ. 2024;302:113964.
54. Zhou J, Wang G, Jia J. Graph neural networks in agriculture: A comprehensive survey. Information Processing in Agriculture. 2024.