dc.description.abstract | Effective drought monitoring relies upon accurate long-term estimates of precipitation. Stage IV precipitation estimates provide a high spatial and temporal resolution that is effective for short-term applications, but in long-term periods useful for drought monitoring, compounding biases can reduce its accuracy. Errors in the dataset include beam blockage, mean field and range dependent, and two-dimensional biases. This study improves upon a three-step bias adjustment methodology that corrects for the Stage IV biases east of the Rocky Mountains, adds a fourth step to remove discontinuities caused by independent analyses, determines an optimal interpolation method for adjusting the data, and assesses the performance of each step within different regions and accumulation periods. Beam blockage is identified using image filtering and ridge detection software, then adjusted using unblocked data. Mean field and range dependent biases are adjusted using radar estimates, normal precipitation, and rain gauges. Discontinuities are adjusted with an inverse distance weighting (IDW) method that blends the data. Two-dimensional biases are adjusted using gauge-radar biases that are interpolated to the entire precipitation field. Extensive testing of the performance is done using a combination of data withholding and comparison of radar estimates to gauges.
The Stage IV bias adjustments generally result in lower root mean square error (RMSE), median absolute error (MAE) and median bias (MB) compared to gauges. The largest improvements are seen in gauge-based adjustments, which include the mean field and range dependent and two-dimensional steps, and in regions where there are fewer gauges and greater variations in seasons. The smallest improvements and cases of increased error arise in radar-based adjustments, which include the beam blockage and discontinuity adjustment steps. When testing interpolation methods, IDW is optimal versus inverse distance weighting squared (IDW2) and ordinary kriging (OK) for its lower error and computational expense. | |