Atmospheric carbon dioxide (CO2) contributes to climate change through its impact on the Earth’s radiation budget1. Atmospheric CO2 concentration has increased at an unprecedented rate from 280 ppm at the start of the industrial revolution to over 400 ppm in recent years, arising primarily from the burning of fossil fuels2. Not all CO2 emitted from anthropogenic sources remains in the atmosphere; about one-fourth of the emissions are currently absorbed by the terrestrial biosphere3, which has significantly slowed down global warming4. Unlike the anthropogenic emissions, the terrestrial carbon (C) sink is driven by two large opposing ecosystem fluxes, i.e., photosynthesis and respiration, and is often indirectly estimated from the residuals of global C budgets. This results in large uncertainties regarding the magnitude and distribution of this terrestrial C sink across spatial and temporal scales5.

While it is challenging to precisely quantify the terrestrial C sink strength, trends and variations in atmospheric CO2 concentration provide a unique lens through which to probe the dynamics of the terrestrial C cycle as well as its interactions with climate. There is wide recognition that the amplitude of the seasonal oscillation of atmospheric CO2 (i.e., annual peak-to-trough difference in atmospheric CO2 concentration) in the northern hemisphere has increased by ~50% since the 1960s, tracking the pace of the contemporary climate warming6,7. A feature of this increase in the CO2 seasonal amplitude is a progressively earlier and larger drawdown of atmospheric CO2 concentration in northern spring and summer, indicating warming-driven lengthening and intensification of photosynthetic activity in northern terrestrial ecosystems5,8, in line with a widespread “greening” trend during the early 1980s to late 1990s9,10. While this positive warming impact on terrestrial C uptake can be traced back to the 1980s, there is emerging evidence that the interannual correlations between growing season temperatures and CO2 drawdown anomalies have been substantially attenuated or even reversed in direction at Mauna Loa11 and northern high latitudes12,13,14 in recent decades. Although the underlying causal mechanisms remain unclear, these changing relationships appear to signal an emergent shift in both phase and magnitude of the terrestrial C sink and underscore a pressing need to better understand how C exchange dynamics is responding to ongoing climate change across a diverse range of northern terrestrial ecosystems15.

Though much research has focused on the C source and sink activities of tropical and boreal forest ecosystems, less attention has been paid to the role of terrestrial ecosystems at northern temperate latitudes (30° to 50° N) in the context of the global CO2 seasonal cycle16. Previous studies showed that recent warming has resulted in altered phenology and increased net primary productivity (NPP) in spring and autumn in temperate forest ecosystems, suggesting a current and possible future enhancement of C sequestration in these ecosystems17,18. Importantly, at northern mid-latitudes, terrestrial ecosystems are spatially heterogeneous and include a substantial area of croplands19. We know from a network of ecosystem-scale CO2 exchange measurements20,21 and satellite observations22 that densely vegetated croplands have shorter but more intense C uptake periods than natural ecosystems and are one of the most productive systems on planet earth. Based on top-down and bottom-up models, Gray et al.23 and Zeng et al.24 argued that the intensification of agriculture at northern temperate latitudes was a major, yet largely overlooked, driver of changes in the CO2 seasonal cycle of the northern hemisphere during the past five decades, accounting for 17–45% of the enhanced C exchange needed to explain the increasing CO2 seasonal amplitude. Corn alone constitutes about two-thirds of this agricultural forcing, owing mostly to increasingly concentrated corn production in the Midwestern United States (i.e., the U.S. Corn Belt) and northern China23,24. However, due to the scarcity and limited time period of direct observations, considerable uncertainties remain with respect to the overall strength of this agricultural forcing and the extent to which heterogeneous terrestrial ecosystems at northern mid-latitudes will respond to future climate warming11.

Here we present a decadal record (2007–2019) of direct boundary layer CO2 measurements from a very tall tower in southern Minnesota (the University of Minnesota tall tower Trace Gas Observatory (KCMP)) (Fig. S1a) – a heterogeneous agricultural region that typifies the Corn Belt25,26,27. We compared this decadal record with other long-term time series (2007–2018) of atmospheric CO2 data within the U.S. Midwest (i.e., Park Falls, Wisconsin (LEF) and West Branch, Iowa (WBI) from NOAA’s Global Greenhouse Gas Reference Network)28 to examine the imprint of crop production on the regional CO2 seasonal cycle. Through a statistical examination of these long-term CO2 records, together with inversion products of net ecosystem exchange (NEE; CarbonTracker 2019)29, we quantified the sensitivity of net CO2 exchange to interannual temperature variations and attribute this sensitivity to CO2 exchange dynamics of croplands and natural ecosystems, respectively. The quantified sensitivity was then used to evaluate how the CO2 seasonal cycle and net CO2 uptake in the Corn Belt will respond to future climate warming by year 2050.

Results and discussion

Agricultural imprint on the CO2 seasonal cycle

The decadal CO2 records measured at KCMP, LEF, and WBI were de-spiked, gap-filled, and digitally filtered to extract the long-term CO2 growth rate and the detrended CO2 seasonal cycle (Fig. S1) (see Methods). To probe the link between the CO2 seasonal cycle characteristics and crop production within the region, we defined a cropland fraction (fCS), calculated as the ratio of land area of corn and soybean to total area of land ecosystems (i.e., croplands plus natural ecosystems) (see Methods). Over 2008 to 2018, fCS was 0.43 ± 0.01 (1 σ), 0.12 ± 0.01, 0.56 ± 0.01 within the intense concentration footprints (i.e., 300 km radius to each tower; see Methods) of KCMP, LEF, and WBI, respectively (Figs. S2 and S3), forming a unique gradient for examining the regional impact of corn and soybean on the CO2 seasonal cycle. In addition, we take advantage of the heterogeneous land use within the intense concentration footprint of KCMP (Fig. S2) by sampling the hourly CO2 concentrations based on wind direction. Two monthly CO2 datasets were built using the northwesterly CO2 observations (i.e., 270°–360°; hereafter, KCMPNW) and the southern and southeastern sector observations (i.e., 120°–210°; hereafter, KCMPSSE) along the dominant wind directions (Fig. 1; see Methods). Notably, the south and southeast sector had a significantly higher fCS (0.52 ± 0.01) compared to the northwest sector (0.23 ± 0.01).

Fig. 1: CO2 concentration measured at KCMP.
figure 1

a Hourly CO2 concentration measurements made from 2008 to 2018. b Wind rose showing distributions of wind direction and speed during 2010 to 2018. c Enhancement of CO2 concentration relative to the northwestern sector (270°–360°) when wind speed was greater than 3 m s−1.

The long-term growth rate of CO2 (2.12–2.37 ppm yr−1) was similar across the three tall tower sites and between the three sites and a continental background site without significant agricultural influences (2.26 ppm yr−1; Niwot Ridge, Colorado; see Methods) (Fig. S1). The detrended CO2 seasonal cycle at all three sites reached annual minimum values in late July/early August after a period of rapid CO2 drawdown and then increased gradually until December (Fig. 2b). Compared to NWR, all three sites had earlier and larger CO2 drawdowns and more elevated CO2 concentrations during the dormant season (October–April next year) (Fig. 2b), indicating vigorous C source and sink activities within this highly productive region. Among the three tower sites, WBI had the largest average annual CO2 drawdown (−19 ppm), followed by KCMP (−18 ppm), and LEF (−15 ppm) (Fig. 2e). Correspondingly, the average CO2 seasonal amplitude was largest at WBI (28.1 ppm), followed by KCMP (27.9 ppm) and LEF (23.9 ppm) (Fig. 2e). These spatial gradients are in concordance with the large differences in fCS and the growing season (May–September) NEE among the three sites (Fig. 2c). The strong sensitivity of the CO2 seasonal cycle to corn and soybean production is also highlighted by the larger CO2 drawdown and seasonal amplitude of KCMPSSE compared to KCMPNW (Fig. 2a, d). To further examine the convolution between atmospheric transport, land use, and the CO2 seasonal cycle, we subtracted the northwestern sector (270°–360°) measurements binned by month from the monthly mean of the CO2 concentration in 12 evenly spaced directional sectors around KCMP. From this wind sector analysis, a clear depletion of CO2 is evident in the peak growing season (i.e., July and August) when winds were coming from the heart of the Corn Belt (i.e., south and southeast) (Fig. 1c). Collectively, these results underscore a large spatial gradient in the atmospheric CO2 concentrations within the study domain and corroborate the strong imprint of crops on regional CO2 uptake26,30.

Fig. 2: The annual CO2 exchange metrics at the three tall tower sites.
figure 2

The detrended CO2 seasonal cycle (a, b) and the CarbonTracker NEE (c) averaged over the entire observation period. The CO2 seasonal amplitude (d, e) and the NEE amplitude f at the three tall tower sites. Trends in the annual CO2 exchange metrics were estimated using the nonparametric Theil–Sen estimator and shown in the legend. Statistically significant trends (the Mann-Kendall trend test; P < 0.1) are indicated by an asterisk symbol in the legend.

Over the 11-year analysis, total production and yields of corn and soybean increased significantly (the Mann-Kendall test; P < 0.1) within the intense concentration footprints of KCMP and WBI (Fig. S5). The increasing yield trends are consistent with the continued growth in crop productivity over the entire Corn Belt and can be attributed to several possible mechanisms related to advances in breeding and genetic technology (e.g., longer maturity cultivars that can adapt to higher sowing density), agronomic practices (e.g., improved herbicide and weed management), and favorable growing conditions (e.g., enhanced water use efficiency under rising atmospheric CO2)31. Over the study period, the CO2 seasonal amplitude increased at a rate of 0.18 ppm yr−1 (the nonparametric Theil–Sen estimator) and 0.39 ppm yr−1 at KCMP and WBI, respectively (Fig. 2c), although the trends are not statistically significant (the Mann-Kendall test; P = 0.12 for KCMP and P = 0.14 for WBI). In contrast, the CO2 seasonal amplitude observed at LEF exhibited much smaller interannual variability than those at KCMP and WBI (Fig. 2e), and a significant increasing trend of the CO2 seasonal amplitude (0.19 ppm yr−1; P < 0.1) was evident (Fig. 2e). To compare with the seasonal amplitude of CO2, we have also calculated the annual amplitude of NEE using the NEE inversion products within the intense concentration footprints of the three tall tower sites (see Methods). While the mean magnitude of the annual NEE amplitude varied across the three sites, consistent with the CO2 seasonal amplitude, significant trend was absent at any site (Fig. 2f). To further examine causal relationships between crop yields and the CO2 exchange dynamics, a correlation analysis (Pearson correlation coefficient) was applied after a linear detrending of all the variables. No significant correlation (P > 0.1) was detected between crop yield anomalies and the CO2 concentration- or NEE-based annual metrics at any of the three sites.

Therefore, although the agriculture intensification at northern mid-latitudes is believed to be an important driver of the increasing CO2 seasonal amplitude in the northern hemisphere across the decadal to multi-decadal scales23,24, our results, based on direct observations over an intensively agricultural region, suggest a decoupling between crop yields and CO2 exchange intensity at the interannual scale. Because the CO2 seasonal amplitude is an integrated measure of annual CO2 exchange, this decoupling may be due to compensating responses of photosynthesis and ecosystem respiration to variations in climatic forcings at sub-annual scales5. Moreover, the large differences in magnitude and seasonal dynamics between KCMPNW and KCMPSSE (Fig. 1) imply that changes in atmospheric transport and circulation may have also played a role in weakening the interannual association between crop yields and atmospheric CO211,16. On the other hand, although grain yield constitutes a large fraction of crop NPP, the extent to which assimilated C in crops is partitioned to grain depends on whether and when the temperature threshold has been exceeded31,32, implying an intricate climate modulation on the yield–CO2 relationship. Notably, the past decade exhibited numerous extreme temperature and precipitation variations in the U.S. Midwest. The long-lasting and pervasive heat wave and drought in the spring and summer of 2012 damaged a substantial proportion of crop commodities33 and adversely affected the functionality of natural ecosystems in this region34; notwithstanding, four of the wettest years in the last 100 years (ranks 1–4) were also recorded in the past decade35. These large interannual climate variations provide a natural experiment to observe the behavior of terrestrial ecosystems under anomalous climate conditions, making it possible to unravel the dynamic link among climate, crop production, and CO2 exchange at scales finer than the annual scale.

Temperature sensitivity of the CO2 exchange anomalies

To examine how changes in the CO2 seasonal cycle were linked to temperature variations at the monthly scale, we calculated the first-time derivative of the CO2 time series, ΔCO2 (Fig. S6), which has been demonstrated to be a better proxy of net land-atmosphere CO2 fluxes than the original CO2 seasonal cycle at northern latitudes36. The sensitivity of ΔCO2 (or NEE) anomaly to temperature variations (βT) within the intense concentration footprint of each tall tower site was then estimated as the slope of the regression on temperature in a multiple linear regression (MLR) of ΔCO2 (or NEE) against temperature, water availability (i.e., 3-month cumulative precipitation including the current month; see Methods), and radiation (all variables detrended). Climate variations explained 14–65% and 19–81% of variances in the ΔCO2 and NEE anomalies of individual months, respectively (Fig. S7 and S8). The climate anomalies explained a much larger fraction (i.e., >60%) of ΔCO2 and NEE variances in the growing season months than in the dormant season (Figs. S7 and S8), signifying the important role of climate–vegetation interactions in controlling the regional CO2 exchange variability.

There were pronounced seasonal patterns in βT of ΔCO2 and NEE at all three tall tower sites (Fig. 3; Supplementary Data 1 and 2). Notably, βT of both ΔCO2 and NEE were significantly positive (i.e., an increase in monthly mean temperature leads to reduced net CO2 uptake) in July and August, whereas only βT of NEE was significantly negative in June at KCMP and WBI (Fig. 3). A correlation analysis (Pearson correlation coefficient) between the monthly ΔCO2 and NEE anomalies shows that the ΔCO2 and NEE anomalies were significantly (P < 0.05) and positively correlated in summer (July and August) and early spring (April and/or May) at all three sites, whereas no significant correlation emerged in June at any of the sites (Fig. S9). Because the NEE inversion products were derived taking into account atmospheric transport and circulation, the absence of significant correlation between ΔCO2 and NEE in June suggests an important role of air mass transport and mixing in determining the temperature–ΔCO2 relationship at the regional scale37. Despite large interannual variations in dormant season air temperature (data not shown), βT of ΔCO2 and NEE were mostly small and not significantly different from zero in the dormant season months (Fig. 3). It is noteworthy that incorporating uncertainties in defining the concentration footprints did not qualitatively change the estimated βT of ΔCO2 and NEE at any of the three tall tower sites (see Methods; Figs. S10 and S11).

Fig. 3: Temperature sensitivity of ΔCO2 and NEE within the intense concentration footprints (300 km radius) of KCMP, LEF, and WBI.
figure 3

Panels a and b, c and d, and e and f are for KCMP, LEF, and WBI, respectively. Gray shaded area denotes the 90% confidence interval of estimated sensitivity derived through resampling. Sensitivities significant at the 90% confidence level are denoted by solid squares.

A panel data model that combines the climate and NEE anomalies of the three tall tower sites was used to derive βT of NEE specific to croplands (i.e., corn and soybean) and natural terrestrial ecosystems (see Methods). This approach leverages the contrasting fCS across the three sites and assumes that the differences in climate sensitivities between croplands and natural ecosystems were coherent within the domain of the three tower sites (see Supplementary Discussion for an extended discussion). The estimated biome-specific βT can therefore be viewed as an area- and CO2 flux-weighted net temperature sensitivity that encompasses the entire range of ecosystem processes by which temperature impacts ecosystem CO2 exchange within each biome category. The estimated biome-specific βT illuminates the dominant role of croplands in driving the negative βT of NEE in June, while NEE of both croplands and natural ecosystems responded positively to temperature variations in July and August (Fig. 4; Table S1). These results are robust for a range of tower footprint radii (150–450 km) used in the panel analysis (Fig. S13) and are not sensitive to changes in the definition of the two biome categories (Fig. S14).

Fig. 4: Biome-specific temperature sensitivity of NEE for croplands and natural ecosystems.
figure 4

Panels a and b are for croplands and natural ecosystems, respectively. Gray shaded area denotes the 90% confidence interval of estimated sensitivity derived through resampling. Sensitivities significant at the 90% confidence level are denoted by solid squares.

The negative temperature impact on cropland NEE in June (i.e., higher temperature favors enhanced CO2 uptake) is consistent with the rapid phenological development of corn and soybean during this critical transition period when cumulative thermal energy typically meets the threshold of crop leaf emergence in the Corn Belt (Fig. S15a)38. Field evidence shows that during this early vegetative stage, the positive feedback between crop canopy development and photosynthetic capacity amplifies the response of crop photosynthesis to temperature variations, leading to accelerated crop growth under warmer temperatures32,39. Furthermore, the negative βT in June might have been an indirect result of human responses to spring climate variations40. In the Corn Belt, the timing of crop planting is largely determined by temperature and precipitation in early spring41. Earlier crop planting in years with warm and dry springs can therefore hasten crop growth in early to mid-spring, giving rise to a stronger photosynthetic response to temperature in late spring to early summer. Indeed, there was a significant linear relationship between the anomaly of June NEE and anomaly of simulated corn leaf emergence date (CLED) within the intense concentration footprints of KCMP and WBI (Fig. S15b; see Methods). This positive correlation remains robust at both sites after removing the control of June temperature variations (Partial correlation analysis; P < 0.1), underpinning a crop phenology-mediated legacy effect of early spring temperature variations on CO2 uptake in June. On the other hand, in contrast to previous studies in temperate deciduous and evergreen forests17,18, natural ecosystems within the study domain did not show a strong response to temperature variations during spring to early summer (Fig. 4b). Importantly, with the tower footprints on the order of several hundreds of kilometers (Fig. S4), the derived βT inherently represents a net sensitivity across a diversity of ecosystem types (e.g., grasslands, forests, and wetlands). As a result, the lack of coherent temperature response likely reflects asynchronous temperature–phenology regimes among different natural ecosystems that compensates temperature-driven NEE anomalies with increasing levels of spatial aggregation42,43.

While most previous long-term C cycle studies in northern ecosystems have concentrated on the temperature–phenology–CO2 interactions in spring and autumn12,18, less attention has been given to the interannual relationship between temperature and CO2 anomalies at the height of the growing season14. Our analyses, based on either ΔCO2 or NEE anomalies, unequivocally identified positive βT of net CO2 exchange (i.e., higher temperature reduces net CO2 uptake) in both croplands and natural ecosystems during the peak growing season (Figs. 3 and 4), where temperature is highest within a year and plants reach their peak photosynthetic capacity in this region38,42. High temperatures affect NEE of crop and natural ecosystems through a variety of direct and indirect pathways. At the ecosystem scale, a well-accepted conceptual model is that photosynthesis responds to temperature variations following a quadratic function, defined by a maximum photosynthetic rate at optimal temperature44, whereas ecosystem respiration increases with temperature through stimulated metabolic rates in an exponential fashion45. Therefore, a positive summer βT can be an indicator of ecosystems operating beyond their thermal optima of photosynthesis. In addition, high temperatures can also suppress photosynthesis and ecosystem productivity by imposing water stress on plants46,47. High summer temperatures not only limit soil water supply by sustained evapotranspiration but also increase atmospheric water demand by increasing the vapor pressure deficit (VPD) of the atmosphere. In response to increased VPD coupled with limited soil moisture, plants close their stomata to prevent excessive water loss, at the cost of reduced CO2 uptake48,49. Importantly, this temperature-induced negative impact on plant CO2 uptake is exacerbated by insufficient summer precipitation, resulting in heat and drought stresses on ecosystem productivity50.

To gain insight into the importance of this potential interaction between temperature and water availability (i.e., 3-month cumulative precipitation including the current month) in driving the summer CO2 exchange anomalies, we estimated the summer temperature sensitivity under different water availability conditions. Specifically, we transformed the climate variables and CO2 exchange of July and August to z-score anomalies using their monthly means and standard deviations, pooled the z-score anomalies of the three tower sites, and grouped this combined dataset into four bins: dry (z-score less than −1), moderate dry (z-score between −1 and 0), moderate wet (z-score between 0 and 1), and wet (z-score greater than 1) summers. An MLR was then applied to estimate βT for each bin. As shown in Fig. 5, βT of ΔCO2 and NEE were significantly greater in dry summers than in other bins of summer water availability, indicating that dry conditions indeed increased summer βT by imposing plant water stress that can also lead to lowered temperature optimum for photosynthesis44,50. However, even in summers with above average water availability (i.e., moderate wet and wet summers), βT of ΔCO2 and NEE were significantly positive (Fig. 5), suggesting that both croplands and natural ecosystems have adapted to current summer temperature and were operating at their thermal optima of CO2 uptake. Thus, high summer temperatures caused reduced net CO2 uptake in this mesic, seasonally cold region both during drought and modestly dry periods that regularly occur in the peak growing season51.

Fig. 5: Temperature sensitivity of ΔCO2 and NEE for different bins of summer water availability.
figure 5

Error bar denotes the 90% confidence interval of estimated sensitivity derived through resampling (i.e., deemed significant at the 90% confidence level if the error bar does not contain zero). Sample size for each bin is also shown.

The finding that the summer CO2 uptake of croplands has already reached its thermal optimum implies a strong temperature control on crop production in the Corn Belt. In light of the revealed βT of ΔCO2 and NEE, we used a panel data model with mean spring temperature (i.e., May and June), summer temperature (i.e., July and August), and growing season precipitation as the explanatory variables to probe the linkage between climate and crop yield variations within the footprints of KCMP and WBI (see Methods). The model results show that temperature and precipitation variations together with an increasing yield trend explain about 80% of variances in corn yields at the two sites (Fig. S16a, S16b; Table S2) and that higher summer temperature reduced corn yield at a rate of −0.36 t ha−1 °C−1 (90% CI: −0.54 to −0.19 t ha−1 °C−1 with a base summer temperature of 23 °C), or about 3.0% °C−1, in strong agreement with previous findings (e.g., 2.5% reduction for every 0.8° rise above 23°)52. Evidence from agronomic research shows that July and August correspond to the critical reproductive stage of corn (i.e., grain filling) in the Corn Belt and that temperature stress during this stage of corn growth contributes directly to yield reduction by shortening the grain filling period and reducing the translocation of photosynthate into reproductive biomass (i.e., lower harvest index)31. On the other hand, although higher spring temperature has favored net CO2 uptake (Fig. 4a) and thus vegetative development of corn, the effect of spring temperature on the final corn yield was not significant (Table S2). This finding is in line with results from recent heating experiments conducted in the central Corn Belt that warmer leaf temperatures during the vegetative stage do not exert lasting effects on corn reproductive growth, possibly due to the high optimal temperature of photosynthesis relative to the background spring temperature during this period32,53.

Interestingly, while growing season precipitation had a significant positive impact on soybean yields within the footprints of KCMP and WBI, neither spring nor summer temperature was significantly correlated with soybean yield variations over the 11-year analysis (Fig. S16c, S16d; Table S3). The lack of significant temperature effect on soybean yields has been previously reported by Lobell et al.46 using historical (1995–2012) soybean yield records in the central Corn Belt. Given that atmospheric CO2 concentrations increased by ~30 ppm over the 11-year span at KCMP and WBI, temperature effects of soybean yields may have been confounded by the CO2 fertilization effect, which has been shown to alleviate temperature stress on soybean reproductive development by enhancing soybean water-use efficiency and photosynthesis54, but have no significant impact on corn yields53. These results imply a dominant control of corn in mediating the link between temperature variations and cropland CO2 exchange within the study domain. Indeed, there was a significantly negative relationship between the detrended corn yield anomaly and the anomaly of July NEE within the intense footprints of KCMP and WBI (ordinary linear regression; P < 0.05) (Fig. S17). This direct connection between corn yield and July NEE was persistent even without including 2012 – a historical bad year for corn yields in this region (Fig. S17). Therefore, despite the decoupling between crop yields and CO2 exchange intensity at the annual scale, this study highlights a dynamic modulation of temperature on cropland CO2 exchange and crop yields, which provides important implications for the impact and adaptation of crop production systems to future climate warming.

Implications for carbon cycle impacts of future climate warming

Projected climate data were retrieved from 10 general circulation models that have contributed to the Coupled Model Intercomparison Project Phase 5 (CMIP5) and run under the RCP4.5 and RCP8.5 scenarios (see Methods). Ensemble mean projections of average air temperature change by 2050 in the Corn Belt were roughly 2 °C for most months under RCP8.5 and between 0 °C and 2 °C under RCP4.5 (Fig. S18a). In contrast to the unanimous warming, models were mixed in the direction of projected precipitation and radiation changes under both the RCP4.5 and RCP8.5 scenarios, resulting in small overall monthly changes (e.g., <±10%) relative to inter-model variability in both cases (Fig. S18b, S18c). Because the land use characteristics (Fig. S3), crop yields (Fig. S5), and CO2 exchange dynamics (Fig. S19) of KCMP are representative of the broader Corn Belt (see Supplementary Discussion for an extended discussion), we applied the projected mean temperature changes to the estimated βT of KCMP and its uncertainty to predict how future climate warming may impact the CO2 seasonal cycle and net CO2 uptake in the Corn Belt. Here, we define the Corn Belt by those states in the U.S. Midwest with significant corn and soybean land use (Fig. S19)55,56. The total area of land ecosystems within the Corn Belt is estimated at 148 million ha55. It is important to note that all the projected mean air temperature changes in the Corn Belt are within the range of historical observations at KCMP (2010–2019; Fig. S18a), which improve the plausibility of extrapolating to future warming scenarios.

Assuming a stasis of seasonal changes in atmospheric transport and circulation, warming in the next decades could alter the trajectory of the CO2 seasonal cycle (Fig. 6a). Higher summer temperature will limit CO2 drawdowns and consequently attenuate the CO2 seasonal amplitude from the current level by 1.5 ppm (~5%) to 3 ppm (~10%) under the two warming scenarios (Fig. 6a). This prediction is in line with the emerging negative impact of warming on summer CO2 drawdown in boreal ecosystems (−2.06 ppm °C−1)14 and suggests that the loss of stimulating effects of warming on the CO2 seasonal amplitude, as recently discovered at the northern high latitudes (>50°N)13, may have a larger spatial extent than previously thought. Extrapolating to the land ecosystems of the entire Corn Belt, the negative warming impact can reduce net CO2 uptake during the peak growing season by 30 Tg °C (90% CI: 10 to 60 Tg °C) under RCP4.5 and 60 Tg °C (90% CI: 20 to 117 Tg °C) under RCP8.5, equivalent to approximately 10 to 20% of the annual net CO2 sequestration (i.e., 292 Tg °C; Fig. S19) of this highly productive region (Fig. 6b). This negative warming impact, however, can be partially offset by the positive impact in June (12 to 29 Tg °C under the two warming scenarios) and, to a lesser extent, May (5 to 10 Tg °C) (Fig. 6b), as a result of crop phenological development. Integrated over the entire growing season, warming by 2050 is projected to reduce the net CO2 uptake by 9 Tg °C (90% CI: reduction by 54 to enhancement by 36 Tg °C) to 13 Tg °C (90% CI: reduction by 92 to enhancement by 55 Tg °C) under the two warming scenarios, although this negative impact is not significant at the 90% confidence level under either scenario due to the compensatory temperature effects on spring and summer CO2 uptake. Combining phenology observations with ecosystem-scale NEE measurements, Keenan et al.18 showed that increased spring and fall temperature has lengthened the growing season of temperate forests over the eastern U.S. (total land area = 38 million ha), leading to enhanced CO2 uptake at a rate of 16 g C m−2 per 1 °C increase in spring or fall. Applying this increasing rate of CO2 uptake to the future warming scenarios suggests an annual gain of CO2 sequestration ranging from 9 to 23 Tg °C in these systems. While it is unclear how these systems are currently responding to temperature variations in summer, this projected increase in net CO2 uptake is of similar magnitude to the net reduction of growing season CO2 uptake in the Corn Belt. Collectively, these results highlight that overall magnitude and timing of future climate warming could be equally critical in determining the C sink strength of terrestrial ecosystems at northern temperate latitudes.

Fig. 6: Changes in the detrended CO2 seasonal cycle and net CO2 uptake in the Corn Belt under future warming scenarios by 2050.
figure 6

Black squares and line in a denote the observed average seasonal cycle of CO2 during 2008–2018. Shaded area in a and error bars in b denote 90% confidence intervals derived through resampling. It is important to note that in b, a negative warming impact denotes reduced net CO2 uptake under the future warming scenarios.

It is important to note that the projected warming impacts, based on the average monthly temperatures, do not account for substantial reduction in CO2 sink strength by extreme heat events34,46, which are expected to continue increasing in frequency and severity in the future57. Besides the direct and indirect physiological impacts of warming discussed above, the regional CO2 seasonal cycle and CO2 sink strength are also modulated by a myriad of slow-evolving and climate-sensitive processes (e.g., CO2 fertilization effect, soil C turnover, and nutrient cycling)58, which may not vary linearly with the projected future warming at multi-decadal scales. Furthermore, the projected future warming impacts can be countered by adaptation measures taken by farmers, such as changes in planting dates or use of longer-maturing cultivars40. For example, earlier planting may be enabled by warmer spring temperatures in the future. Shifts in development timing will therefore modulate the weather experienced by crops and may alleviate the adverse effects of higher summer temperatures. Because our projection does not account for farmer adaptations, the projected warming impacts on the CO2 uptake can be viewed as the expectation in the absence of explicit recognition of, and adaptation to, temperature trends from present to 2050. Thus, a key question that remains to be answered is whether the revealed negative warming impacts on net CO2 uptake in northern terrestrial ecosystems indicate a future climatic tipping point for CO2 sequestration and plant productivity in these dynamical systems. Regardless, this study challenges the paradigm that warming will continue to benefit CO2 sequestration in terrestrial ecosystems at northern mid-latitudes and emphasizes the need to robustly represent the temperature sensitivity of cropland CO2 exchange for current climate in C cycle models in order to improve the predictability of future carbon-climate feedbacks.


Atmospheric CO2 concentration

The tall tower CO2 observations reported here were measured from April 2007 to December 2019 at the University of Minnesota tall tower Trace Gas Observatory (KCMP tall tower; 44.6888°N, 93.0728°W) (Fig. S2a). Air was pulled continuously from 100 m above ground to the base of the tower, where it was dried, subsampled, and measured for CO2 concentration at 10 Hz using a tunable diode laser spectrometer (TGA100A, Campbell Scientific Inc., Logan, Utah, USA). The calibrated 10 Hz data were then block averaged into hourly values, with a long-term precision of 0.2 ppm. Wind speed and direction at 100 m were measured using a sonic anemometer (CSAT3, Campbell Scientific, Logan, Utah, USA). Further details regarding the tall tower sampling and calibration scheme can be found in Griffis et al.59. CO2 concentration data collected from 2007 to 2010 have already been reported by Zhang et al.27 and Hu et al.37 in assessments of regional-scale CO2 fluxes. Here we use all available data and focus our analyses on the CO2 seasonal cycle at the interannual timescale.

Two additional long-term tall tower sites from NOAA’s Global Greenhouse Gas Reference Network, Park Falls, Wisconsin (LEF; 45.9451°N, 90.2732°W) and West Branch, Iowa (WBI; 41.7248°N, 91.3529°W), are located within the region and were used in this study (Fig. S2a). Hourly CO2 concentration data measured at LEF and WBI from 2007 to 2018 at 99–122 m above ground were obtained from NOAA’s ObsPack data products28.

The hourly CO2 time series of the three tall tower sites were de-spiked, gap-filled, and block averaged into daily values (see Supplementary Methods for more details). Following the method of Barlow et al.36, a wavelet transform was used to spectrally decompose the daily CO2 time series. The detrended seasonal cycle and long-term growth of CO2 were then isolated by summing frequencies at periods of 3–18 months and >18 months, respectively36 (Fig. S1; see Supplementary Methods for more details). The CO2 seasonal amplitude was obtained as the peak-to-trough difference of the detrended seasonal cycle.

To examine how the CO2 seasonal cycle measured at KCMP was mediated by the convolution of atmospheric transport and ecosystem CO2 exchange, we sampled the KCMP CO2 time series based on wind direction for the period 2010–2018, where we have complete wind data at 100 m height. We only considered the hourly CO2 data with wind speed greater than 3 m s−1 to reduce local source effects60,61. Two CO2 datasets (KCMPNW and KCMPSSE) were built for the dominant wind directions 270°–360° (northwest) and 120°–210° (south and southeast), respectively (Fig. 1a, b). The CO2 seasonal cycle characteristics of KCMPNW and KCMPSSE were extracted using the same wavelet method described above.

To infer the CO2 source and sink strength within the Corn Belt, we compared the CO2 seasonal cycle measured at the three tower sites with continental background CO2 measured at Niwot Ridge (NWR; 40.0531°N, 105.5864°W; managed by NOAA’s Earth System Research Laboratory; Fig. S4). The NWR site sits approximately 27 km west of Boulder, Colorado, and 6 km east of the Continental Divide (Fig. S4). Although climate and biota of the site are characterized by alpine ecosystems, NWR, at an altitude of 3526 m (3523 m elevation; 3 m intake height), is well situated to measure CO2 concentrations in well-mixed continental boundary layer that are representative of large areas without significant influences from local anthropogenic emissions62,63 and agricultural activities55. Weekly CO2 concentrations measured at NWR from 2007 to 2018 were obtained from NOAA’s ObsPack data products28.

Concentration footprint

Detailed footprint analyses have been conducted by Hu et al.37 for KCMP using the Stochastic Time-Inverted Lagrangian Transport model64. From these analyses, 80% of the concentration signal (i.e., sensitivity of concentration variations to surface CO2 exchanges) originated from an area within 307, 255, 302, and 298 km radius of KCMP for the four seasons, respectively (Fig. S4)37. Therefore, we define a 300 km radius as the intense concentration footprint for all three tall towers by assuming that the area within this intense footprint has equally weighted influence on the CO2 observations. We also test a range of radii (150–450 km) to gauge how the definition of intense concentration footprint impacts our analyses and conclusions.

Net ecosystem exchange from atmospheric inversions

Monthly terrestrial biosphere net ecosystem CO2 exchange (NEE) within the tower footprints were obtained from the CarbonTracker assimilation system for 2007–2018 (CT2019)29. The CarbonTracker NEE is an inverse product (1° × 1°) derived from a priori NEE estimates from terrestrial biosphere models and optimized using simulated atmospheric transport and in situ atmospheric CO2 measurements (including LEF and WBI). Zhang et al.27 compared the CarbonTracker NEE to eddy covariance-based bottom-up estimates of NEE within a radius of 200–600 km to KCMP and found excellent agreement between the two methods (Nash–Sutcliffe efficiency (NSE) > 0.9), indicating that the CarbonTracker NEE is sensitive to the heterogeneous C exchange activities within the study domain. Using the NEE data within the intense concentration footprints of the three tower sites, we calculated the annual amplitude of NEE, defined as the difference in cumulative NEE between the dormant season (October–April next year) and the growing season (May–September).

Land use characteristics

High resolution (30 m) land cover data were obtained from the USDA’s National Agricultural Statistical Service National Cropland Data Layer (NCDL) for 2008–2018. We define an index, fCS, calculated as the ratio of land area of corn and soybean to total area of land ecosystems (i.e., croplands plus natural ecosystems), to quantify the fractional influence of corn and soybean within each tower concentration footprint (Fig. S3). Pasture, spring and winter wheat, oats, and perennial crops such as alfalfa hay, which were present to various degrees within the concentration footprints of the three tall tower sites (Fig. S2), were grouped into the category of natural ecosystems in this study because of the challenge of separating pastures from natural grasslands, as well as the long growing seasons of these crops relative to corn and soybean26. More description of the land use characteristics is provided in Supplementary Methods.

Crop data

County‐level corn and soybean statistics from 2008 to 2018 were retrieved from the USDA’s Quick Stats 2.0 database. Total grain production and harvested area for counties that fall within or intercept the intense concentration footprints (300 km radius) of KCMP and WBI were aggregated to calculate annual crop yields (in the unit of t ha−1) for the two sites.

Climate data

Gridded daily average air temperature (2 m), precipitation, and incoming shortwave radiation within the tower concentration footprints were obtained from the National Center for Environmental Prediction North American Regional Reanalysis (NCEP-NARR) for 2007–2019. Daily minimum and maximum air temperature data were obtained from PRISM climate data ( Projected climate data for 2006–2050 over the entire study domain were retrieved from 10 general circulation models that have contributed to the Coupled Model Intercomparison Project Phase 5 (CMIP5). We used projection data derived under two warming scenarios: RCP4.5 and RCP8.5. Importantly, RCP4.5 is a median scenario for future greenhouse gas emissions with modest climate mitigation, while RCP8.5 is a high emission scenario assuming no mitigation65,66. Following the method of Lobell et al.67, the projected climate date time series were downscaled to correct for biases in the coarse-scale outputs from the CMIP5 models. This downscaling ensures that the mean and variance of projected climate data match the observational record for the period 2008–2018, while preserving any simulated trends out to 2050. Changes by 2050 were then calculated as averages for 2041–2050 minus averages for 2010–2019.

Simulation of corn leaf emergence date

Corn leaf emergence date (CLED) within the intense concentration footprints of KCMP and WBI was simulated using growing degree time (GDT) with a base temperature of 8 °C (Fig. S15a)38. Corn leaf emergence occurs when GDT exceeds the threshold of 450 units, assuming that land managers have already planted their fields38. Following the method of De Wit et al.68, diel temperature cycles within the tower concentration footprints were approximated using daily minimum and maximum temperatures to enable accumulation of heat units on an hourly time step. Our previous research showed that CLED simulated using this method has close agreement with long-term observations made at multiple AmeriFlux sites throughout the Corn Belt (e.g., Minnesota, Iowa, Illinois, Nebraska)38.

Statistical analyses

Monotonic trends in the annual CO2 exchange metrics (i.e., the CO2 seasonal amplitude and the NEE amplitude) and crop yields were tested using the Mann-Kendall trend test and then estimated using the nonparametric Theil–Sen estimator, which is a robust method and insensitive to outliers.

To examine how changes in the CO2 seasonal cycle were linked to interannual temperature variations, the daily values of the detrended CO2 seasonal cycle were block averaged into monthly values for calculation of the first time derivative of the CO2 concentration (i.e., change in CO2 concentrations from one month to the previous month), ΔCO2 (Fig. S6). The monthly time series of ΔCO2, NEE, and climate data were then linearly detrended (using the “detrend” function of MATLAB) for each month of the year to allow the following analyses to focus on the interannual relationship between temperature and CO2 exchange anomalies (i.e., ΔCO2 and NEE)69. The interannual sensitivity of ΔCO2 (or NEE) anomaly to temperature variations (βT) for a given month was estimated as the slope of the regression of temperature in a multiple linear regression (MLR) of ΔCO2 (or NEE) against temperature, water availability, and radiation, such that indirect effects arising from covariations between the climate anomalies are accounted for in deriving βT (i.e., equivalent to a partial correlation between temperature and CO2 exchange anomalies controlled for the effects of precipitation and radiation anomalies)50. We constructed detrended time series of precipitation anomaly cumulated for various lag time durations (2–6 months) to account for potential legacy effects of precipitation on ecosystem CO2 exchange70 and found that using a 3-month cumulative precipitation (P3m) anomaly in the MLR resulted in the best regression fits for the three tower sites in combination. Therefore, P3m was used as an index of water availability for all subsequent analyses. Uncertainty in the estimated βT due to finite historical observations was estimated using bootstrap resampling (1000 iterations).

To further contrast the temperature sensitivity between croplands (i.e., corn and soybeans) and natural ecosystems, we used a panel data model that combines the climate (ΔT, ΔP3m, and ΔR) and NEE anomalies of each month, year, and site (superscripts m, y, and s) and decomposes the site-specific temperature, precipitation, and radiation sensitivities (i.e., βT, βP(3m), and βR) into sensitivities specific to croplands and natural ecosystems (subscripts CS and NV). This is achieved by weighting the climate sensitivities using the land fractions of croplands (i.e., βT) and natural ecosystems (fNV; fNV = 1 − fCS) within the tower footprints:

$${\Delta}NEE^{m,y,s} = \, \left( {f_{{\mathrm{CS}}}^{y,s} \cdot \beta _{{\mathrm{T}},{\mathrm{CS}}}^m + f_{{\mathrm{NV}}}^{y,s} \cdot \beta _{{\mathrm{T}},{\mathrm{NV}}}^m} \right) \cdot {\Delta}T^{m,y,s} \\ + \left( {f_{{\mathrm{CS}}}^{y,s} \cdot \beta _{{\mathrm{P}}\left( {3{\mathrm{m}}} \right),{\mathrm{CS}}}^m + f_{{\mathrm{NV}}}^{y,s} \cdot \beta _{{\mathrm{P}}\left( {3{\mathrm{m}}} \right),{\mathrm{NV}}}^m} \right) \cdot {\Delta}P_{3{\mathrm{m}}}^{m,y,s} \\ + \left( {f_{{\mathrm{CS}}}^{y,s} \cdot \beta _{{\mathrm{R}},{\mathrm{CS}}}^m + f_{{\mathrm{NV}}}^{y,s} \cdot \beta _{{\mathrm{R}},{\mathrm{NV}}}^m} \right) \cdot {\Delta}R^{m,y,s} + \varepsilon ^{m,y,s}$$

where ɛm,y,s stands for the error term for site s in month m and year y. The model performance was evaluated using R2 for individual months. Confidence intervals of the biome-specific climate sensitivities were estimated using bootstrap resampling, assuming 10% uniform random error in fCS. The key assumption underlying Eq. 1 is that although fCS differed significantly, the biome-specific climate sensitivities were similar across the three tower sites. This assumption was evaluated by comparing the sensitivities reconstructed from the derived biome-specific sensitivities to the “true” sensitivities independently estimated from the MLR of each tower site. The results show that the site-specific climate sensitivities of NEE can be successfully reproduced by the biome-specific sensitivities and fCS at all three sites (NSE > 0.9; Fig. S12). This lends strong support for the use of the panel model that unifies the C exchange and climate anomalies across this heterogeneous region. Please see Supplementary Discussion for an extended discussion on the panel data analysis and its validation.

Following the method of Zhu et al.31, temperature sensitivity (γ) of corn yields (Y) within the intense concentration footprints of KCMP and WBI was estimated using a panel data model with mean spring temperature (i.e., May and June; TMJ), summer temperature (i.e., July and August; TJA), and growing season precipitation (P) as the explanatory variables:

$$Y^{y,s} = \gamma _1 \cdot t + \gamma _{{\mathrm{MJ}}} \cdot T_{{\mathrm{MJ}}}^{y,s} + \gamma _{{\mathrm{JA}}} \cdot T_{{\mathrm{JA}}}^{y,s} + \gamma _{\mathrm{P}} \cdot P^{y,s} + C^s + \varepsilon ^{y,s}$$

where t denotes each year and γ1·t captures the yield increasing trend observed within the footprints of KCMP and WBI. C corresponds to fixed effects of each site and accounts for time-invariant site differences, e.g., the soil quality. ɛy,s stands for the error term for site s in year y. We did not include a quadratic term of temperature in the model because of the limited number of observations (22 site-years) and the fact that growing season temperature spanned a relatively narrow range (e.g., 4.1° and 5.5° for mean July and August temperature at KCMP and WBI, respectively) during the study period.