Traditional Earth system and land surface models (LSMs) excel at simulating natural ecosystems by grouping vegetation into broad categories known as Plant Functional Types (PFTs), such as "C3 grasses" or "deciduous forests." While this abstraction works well for pristine wilderness, it introduces a severe blind spot when applied to agricultural lands. Humans do not allow crops to grow naturally; instead, we actively manipulate them through irrigation, fertilization, tillage, crop rotation, and intensive harvesting.
When models omit these anthropogenic management practices, they fail to capture the true dynamics of the terrestrial carbon, water, and energy cycles. A growing body of research focuses on identifying why these omissions distort our understanding of environmental feedbacks and how explicitly modeling human interventions fixes these systematic biases.
Why Omitting Management Distorts Model Intuition
Leaving human management out of land surface and crop yield models creates three distinct mechanical failures in how models interpret data.
1. Confounding Climate Sensitivity with Human Intervention
If a model relies purely on satellite observations of greenness (such as Leaf Area Index) or climate variables to predict productivity, it remains blind to the inputs driving that growth. For example, if a crop thrives during a dry spell due to artificial irrigation, a management-blind model will incorrectly infer that the crop variety has a high natural tolerance for drought.
As demonstrated by Fudong Lin et al. in MMST-ViT: Climate Change-aware Crop Yield Prediction via Multi-Modal Spatial-Temporal Vision Transformer, public agricultural datasets frequently omit human-controlled factors like irrigation, fertilizer use, and seed variety. The authors point out that this omission prevents models from accurately distinguishing between natural environmental drivers and anthropogenic interventions, leading to highly uncertain climate-sensitivity estimates.
2. Breaking the Soil Carbon Equation
Soil organic carbon (SOC) dynamics depend heavily on the balance between carbon inputs (like crop residues left in the field) and carbon outputs (driven by microbial decomposition). Human management alters both sides of this equation. Tillage physically breaks up soil aggregates, accelerating microbial access and decomposition, while harvesting physically removes biomass that would otherwise return to the soil.
When these factors are ignored, models fundamentally miscalculate long-term carbon sequestration. Beth Drewniak et al. show in Modeling the impact of agricultural land use and management on US carbon budgets that omitting physical and management factors like soil erosion, irrigation, and specific tillage practices leads to incomplete representations of SOC dynamics. This mechanism is echoed by Mengjie Han et al. in Modeling biochar effects on soil organic carbon on croplands in a microbial decomposition model (MIMICS-BC_v1.0), who note that ignoring management practices like tillage, fertilization, or irrigation contributes directly to model underestimations of soil carbon changes.
3. Misattributing Global Carbon Feedbacks
The omission of management practices does not merely degrade local yield predictions; it distorts global climate projections. Historically, year-to-year fluctuations—or interannual variability (IAV)—in the global land carbon sink were attributed almost entirely to natural climate anomalies like El Niño.
However, explicitly accounting for human activities alters this narrative. Chao Yue et al. demonstrate in Contribution of land use to the interannual variability of the land carbon cycle that human land use and management drive roughly 30% to 45% of the land carbon cycle's interannual variability. When models treat agricultural land as static or purely climate-driven, they suffer from severe spatial mismatches in carbon flux simulations, a limitation highlighted by Zhihua Liu et al. in Precipitation thresholds regulate net carbon exchange at the continental scale.
The Path to Explicit Representation
To resolve these systematic errors, researchers are shifting from static ecological baselines to explicit, process-based modeling of human management. This transition is occurring across two primary modeling paradigms.
Explicit Process-Based Crop Modules
Instead of adjusting generic vegetation parameters to artificially mimic high agricultural productivity—a shortcut noted by Wei Li et al. in ORCHIDEE-MICT-BIOENERGY: an attempt to represent the production of lignocellulosic crops for bioenergy in a global vegetation model—modern land surface models embed dedicated crop modules.
Beth Drewniak et al. pioneered this approach in Modeling agriculture in the Community Land Model by introducing "CLM-Crop." By moving beyond generic grass representations, this framework explicitly simulates human interventions like fertilization, planting schedules, and the physical removal of biomass during harvest. Similarly, to capture the distinct environmental footprint of tropical agriculture, Yidi Xu et al. developed specialized structures in Oil palm modelling in the global land surface model ORCHIDEE-MICT to simulate the unique phenology, growth, and harvest cycles of perennial oil palm plantations.
Integrating Management Inputs into Machine Learning
In statistical and machine learning frameworks, researchers are breaking the reliance on climate-only predictors by treating human inputs as primary features. Traditional machine learning models frequently hit predictive ceilings or exhibit systematic biases during anomalous years because they lack access to management data.
Ben Parkes et al. demonstrate the necessity of this data in Modelling fertiliser significance in three major crops, showing that explicitly incorporating fertilizer application data alongside climate variables significantly improves a model's ability to capture yield variability. When these variables are omitted, forecasting models for major staples like maize and wheat suffer from high uncertainty and systematic errors, a limitation documented across diverse regions by Rui Yao 0002 et al. in Prediction of Maize Yield at the City Level in China Using Multi-Source Data and by Jichong Han et al. in Prediction of Winter Wheat Yield Based on Multi-Source Data and Machine Learning in China.
By treating human management not as an external disturbance, but as an intrinsic, process-driven component of the ecosystem, modern models are closing the gap between simulated carbon dynamics and real-world agricultural realities.
Key Foundational Work
Modelling LAI, surface water and carbon fluxes at high-resolution over France: comparison of ISBA-A-gs and ORCHIDEE (2012) — contributed 2 concept elements to this direction.
Application of the ORCHIDEE global vegetation model to evaluate biomass and soil carbon stocks of Qinghai‐Tibetan grasslands (2010) — contributed 1 concept elements to this direction.