Bias correction of daily GCM rainfall for crop simulation studies

• Published in: Agricultural and forest meteorology 2006-08, Vol.138 (1), p.44-53
• Main Authors: Ines, Amor V.M., Hansen, James W.
• Format: Article
• Language: English
• Subjects: Agricultural and forest climatology and meteorology. Irrigation. Drainage; Agricultural and forest meteorology; Agronomy. Soil science and plant productions; autocorrelation; bias; Biological and medical sciences; calibration; corn; Corn ( Zea mays); crop models; Crop simulation model; crop yield; crops; equations; errors; Fundamental and applied biological sciences. Psychology; General agronomy. Plant production; General circulation model (GCM); General Circulation Models; Generalities. Techniques. Climatology. Meteorology. Climatic models of plant production; mathematical models; model validation; plant growth; plant-water relations; Precipitation; rain; Seasonal climate prediction; soil-plant-atmosphere interactions; water availability; Zea mays
• ISSN: 0168-1923; 1873-2240
• DOI: 10.1016/j.agrformet.2006.03.009

General circulation models (GCMs), used to predict rainfall at a seasonal lead-time, tend to simulate too many rainfall events of too low intensity relative to individual stations within a GCM grid cell. Even if bias in total rainfall is corrected relative to a target location, this distortion of frequency and intensity is expected to adversely affect simulations of crop growth and yield. We present a procedure that calibrates both the frequency and the intensity distribution of daily GCM rainfall relative to a target station, and demonstrate its application to maize yield simulation at a location in semi-arid Kenya. If GCM rainfall frequency is greater than observed frequency for a given month, averaged across years, GCM rainfall frequency is corrected by discarding rainfall events below a calibrated threshold. To correct the intensity distribution, each GCM rainfall amount above the calibrated threshold is mapped from the GCM intensity distribution onto the observed distribution. We used a gamma distribution for observed rainfall intensity, and considered both gamma and empirical distributions for GCM rainfall intensity. At the study location, the proposed correction procedure corrected both the mean and variance of monthly and seasonal GCM rainfall total, frequency and mean intensity. The empirical (GCM)-gamma (observed) transformation overestimated mean intensity slightly. A simple multiplicative shift did a better job of correcting monthly and seasonal rainfall totals, but left substantial frequency and intensity bias. All of the bias correction procedures improved maize yield simulations, but resulted in substantial negative mean bias. This bias appears to be associated with a tendency for the GCM rainfall to be more strongly autocorrelated than observed rainfall, resulting in unrealistically long dry spells during the growing season. Nonlinearity of crop response to the variability of water availability across GCM realizations may also contribute. Averaging simulated yields each year across multiple GCM realizations improved yield predictions. The proposed correction procedure provides an option for using the daily output of dynamic climate prediction models for impact studies in a manner that preserves any useful predictive information about the timing of rainfall within the season. However, its practical utility for yield forecasting at a long lead-time may be limited by the ability of GCMs to simulate rainfall with a realistic time structure.