On the initialization of soil carbon models and its effects on model predictions for England and Wales

The initial distribution of carbon (C) between model pools can strongly influence predictions of soil C models. Models are often initialized by assuming that C stocks are near steady state, but in many cases this is unrealistic. We explored different ways of initializing the DAYCENT model over the range of soils and climate in England and Wales. We ran the model for the main soil types on arable land and managed grass, identified by taking the top five soil-land use combinations in each of 87 50-km grid squares, giving 376 ‘sites’ distributed across the two countries. We compared three initialization methods: (i) using the soil C contents and pool distributions (‘soil microbe C’, ‘slow C’ and ‘passive C’ in DAYCENT) predicted for steady state under the prevailing conditions; (ii) using the steady-state pool distributions but with the true, initial soil C content; and (iii) by fitting the initial pool distribution to the rates of change in soil C observed in the National Soil Inventory (NSI) of England and Wales during the 1980s and 1990s. The calculated mean net primary production (NPP), and hence C inputs to the soil, was realistic for arable land and permanent grass in England and Wales. The calculated rates of change in soil C were sensitive to the initialization method. Method 1 predicted little change with the actual climate over the NSI survey period and simulated management, but Methods 2 and 3 predicted losses of C. Methods 2 and 3 gave similar trends averaged across soil types and locations, but there were large differences for individual soils and locations. The relationships between losses and the mean soil C content were approximately linear with Methods 2 and 3, but the slopes differed. The predicted losses varied with different climate scenarios applied over the NSI survey period. The predicted differences between climate scenarios were less sensitive to initialization method than to C content. We discuss generic implications for modelling