A robust DayCent model calibration to represent the impact of integrated soil fertility management on maize yields and soil carbon stocks in Kenya

Sustainable intensification practices, such as integrated soil fertility management (ISFM), form a strategy to close yield gaps while maintaining soil fertility and, typically, are locally tested in field trials. However, to estimate the potential impact of ISFM on a regional scale, field trials are insufficient and biogeochemical models are required. These models need to be calibrated and evaluated when applied to new environments. Here, we present a robust calibration of the DayCent agroecosystem model to simulate the impact of ISFM practices on maize productivity in Kenya, using a probabilistic Bayesian calibration technique with data from long-term field trials at four sites in central and western Kenya. We assessed the efficiency of DayCent in simulating: 1) maize grain yield, 2) changes in soil organic carbon (SOC), and 3) nutrient use efficiency of applied nitrogen (N) fertilizer under different ISFM treatments, which consisted of different organic resources combined with the addition or absence of mineral N fertilizer. After model calibration, both the simulations of maize yield (Nash Sutcliffe Efficiency, NSE 0.51) and change in SOC (NSE 0.54) improved significantly compared to runs using the standard DayCent parameters (NSE of 0.33 and -1.3 for yield and SOC change, respectively). A leave-one-site-out cross evaluation indicated the robustness of the approach for spatial extrapolation, i.e., the significant improvement of model simulations was achieved by calibrating the model with data from three sites and then evaluating it with data from the remaining site. The values of model parameters related to SOC decomposition were most altered  by the calibration, i.e., they were an order of magnitude higher compared to the default parameter values (derived for temperate climates). This suggests that the DayCent temperature function is not suitable to capture SOC decomposition across climates with a single set of parameter values. Further, similar maize yields were simulated for all treatments that received mineral N fertilizer and DayCent underestimated the yield increase observed in the field trials of the combined application of organic resources and mineral N compared sole mineral N application. In contrast, at low levels of nutrient inputs DayCent proved sufficiently sensitive to capture differences in maize yield levels. Finally, while mean yields by treatment were simulated well, year-to-year yield variation was not captured well by DayCent. In summary, our results indicate that DayCent is capable to estimate the mean impact that ISFM practices at typical rates of mineral fertilizer and organic resource applications have on yield and SOC, but may not be capable to estimate the differences in yield potential at very high inputs. While the cross evaluation indicated a robustness for upscaling, the suboptimal representation of year-to-year yield variabilities shows that future projections under a changing climate may be biased by the DayCent model. Consequently, improved model structures, such as improved soil moisture representation, are needed to reduce uncertainty.