Biodiversity-production feedback effects lead to intensification traps in agricultural landscapes

Global food demand is rising rapidly, historically met through cropland expansion and conventional intensification (high external inputs in monocultures). These strategies have greatly reduced local biodiversity and associated ecosystem functions, creating negative feedbacks on yields, especially at high management intensities where yield responses saturate and biodiversity losses can cross tipping points. Such conditions can produce ‘intensification traps’—lose-lose scenarios where both biodiversity and production decline due to over-intensification and difficult-to-reverse legacy effects (for example, soil biodiversity and fertility loss, slow recovery of above-ground biodiversity). The mechanisms and contexts that generate these traps are not well resolved because crop, soil and biotic conditions vary widely across landscapes, shaping biodiversity–production relationships. This study aims to identify biophysical mechanisms driving intensification traps and evaluate how their emergence varies with crop, soil and biotic characteristics. The authors introduce an analytical framework that integrates biodiversity as both predictor and response in agricultural planning via five key relationships linking land management, biodiversity, and yields.

The authors conducted systematic literature reviews for each model constant representing effect sizes and shapes (concave/linear/convex) of five key relationships. A restricted meta-analysis complemented by a snowball search was used to capture the natural variability across agricultural landscapes, crops, and regions. For each constant, 11–26 datasets were assembled to estimate means and standard deviations, which parameterized normal distributions from which model parameters were drawn in stochastic simulations. Literature also informed species pool characteristics and habitat requirements influencing biodiversity responses.

The framework considers two land-management dimensions: (1) the proportion of land used for agriculture (working land, WL) and (2) the level of conventional intensification (I, representing inputs such as fertilizers and pesticides). Total production Pt = WL × Y. Yield Y depends on: (A) the extent of working land (WL) through the distribution of yield potential across the landscape (higher-potential areas are converted first; mean and variance determine average realized yield as WL expands), (B) conventional intensification I (saturating positive effects), and (C) biodiversity Bt (positive effects via ecosystem functions like pollination, pest control, nutrient cycling). Biodiversity Bt, in turn, responds to management via (D) conventional intensification I (negative effects varying in effect size and curve shape depending on community sensitivity) and (E) the extent of agricultural land WL through species’ habitat requirements and regional species pool composition (species categorized by ability to persist in natural/semi-natural habitat, working lands, or both, and by minimum habitat threshold for persistence). Effect sizes (0–1) and relationship shapes (slope scaled −1 concave to +1 convex, 0 linear) parameterize functions f1–f5 linking Y and Bt to WL and I. Spatial elements are included where WL is a predictor (A and E). For each landscape, the model evaluates 10,201 management options (WL and I from 0 to 1 in 0.01 steps). Outputs (biodiversity and total production) are range-transformed to 0–1 within each landscape. Analyses included: (1) stochastic generation of 10,000 artificial landscapes by drawing each model constant from literature-informed normal distributions; (2) three archetypal case studies (US wheat belt; Southeast Asian rice; sub-Saharan small-holder diversified systems) parameterized from literature; and (3) systematic sensitivity analysis, varying one model constant across an extreme but realistic range while holding others at mean literature values, with 100 landscapes per constant. Modeling was implemented in R 4.1.0. The framework also derives opportunity-cost curves indicating maximal biodiversity attainable at each production level.

• Across 10,000 artificial landscapes, applying the highest management intensities produced intensification traps in 73% of cases, reflecting real-world variability in the five key relationships.
• Risk of traps and their maximal production loss increase with the effect size of biodiversity on yield, although there is considerable scatter due to multiple interacting drivers.
• Landscapes with very high risk (>80% of possible land uses) were rare (<1%), occurring when biodiversity peaks at intermediate to high fractions of working land.
• Small-loss large-gain trade-offs are prevalent: reducing maximum production by 5–10% often yields disproportionally large biodiversity gains; in the rice case study, a 5% production reduction doubled biodiversity.
• Archetypal case studies: (i) US wheat belt and (ii) Southeast Asian rice did not exhibit intensification traps even at high intensities (despite positive biodiversity effects), whereas (iii) sub-Saharan small-holder diversified systems showed substantial production reductions at high intensities due to stronger biodiversity dependence (pollination, pest control) and higher yield heterogeneity.
• Sensitivity analysis identified mechanisms: risk of traps increases with larger biodiversity-to-yield effect size (C), stronger and more immediate biodiversity losses with intensification (higher effect size in D), more concave biodiversity–yield responses (benefits realized at low biodiversity but easier to fall into traps), and stronger reliance of beneficial species on natural habitats (E). Risk decreases with stronger and more linear direct intensification benefits on yield (B), higher average yield potential and lower variance across the landscape (A), and convex, resistant biodiversity responses to intensification (D).
• Management intensities that maximize production often occur near either a high-intensification optimum or a high-biodiversity optimum, producing bimodal patterns in optimal strategies; weak ecological minimum requirements risk pushing systems into low-yield states between these optima.
• Optimal land management (highest production) shows a positive relationship between conventional intensification and agricultural land used (type 2 regression R² = 0.36, p < 0.001, Extended Data Fig. 5).

Findings clarify when and why intensification traps occur: when biodiversity-mediated services (pollination, pest control, nutrient cycling) are crucial to yields and biodiversity declines rapidly with conventional intensification, indirect negative feedbacks can outweigh direct production gains, yielding lose-lose outcomes. The framework demonstrates strong context dependence driven by crop type, landscape yield heterogeneity, community sensitivity to inputs, and species pool habitat requirements. For globally dominant cereal systems with high direct responsiveness to inputs and lower biodiversity dependence, trap risk is lower. However, diversified systems, especially those dependent on animal-mediated services, are more vulnerable. The strong nonlinearity of opportunity-cost curves implies that modest production sacrifices can deliver large biodiversity benefits, even where traps do not occur. Policy and management implications include designing precautionary safety margins below estimated intensification optima to hedge against uncertainty and avoid costly double losses; considering ecological intensification and habitat conservation to sustain ecosystem services; and avoiding policies that enforce weak ecological minima that can strand systems between production optima. Incorporating regional species pool traits and landscape structure into planning can better align biodiversity recovery with production goals.

The study introduces an analytical framework that integrates biodiversity feedbacks into agricultural production planning via five key relationships linking management, biodiversity, and yield. It shows that intensification traps—simultaneous biodiversity and production losses—are common across landscape types under high input use, particularly where yields strongly depend on biodiversity and communities are sensitive to intensification. Conversely, many landscapes exhibit small-loss large-gain opportunities: 5–10% production reductions can yield substantial biodiversity gains. The authors recommend precautionary safety margins in intensification to avoid traps and emphasize that reconciling production and conservation is feasible by strategically managing habitat and inputs. Future work should expand the framework to other degradation drivers (for example, soil degradation, salinization), integrate economic and labor dimensions (ecological versus conventional intensification), and explicitly represent landscape structure and regional species pool dynamics to tailor context-specific management and policy interventions.

• The framework simplifies multidimensional aspects of conventional intensification and biodiversity into tractable parameters (effect sizes and shapes), potentially omitting important interactions among practices and taxa.
• Spatial structure is reduced to proportions of land-use types; explicit landscape configuration, connectivity, and spatial spillovers are not modeled, though they can strongly influence biodiversity and service delivery.
• The approach assumes no correlation between biodiversity potential and yield potential across space, which may not hold in all contexts and can affect trade-off outcomes.
• Farm-level application is constrained by data needs for precise process-based parameterization; thus, recommended safety margins reflect precaution under uncertainty rather than exact optima.
• Case studies and artificial landscapes draw on literature-derived parameter distributions and do not represent the prevalence of global production systems; generalizability requires local calibration.
• No original empirical data were collected; results depend on the quality and representativeness of reviewed studies.