Using Daily Gauge Analysis to Adjust Biases in the Multi-sensor (radar+gauges) Analysis

The radar-and-gauged based hourly/hourly precipitation analysis typically contain some systematic biases. When used as primary driver of soil moisture (either directly via land data assimilation or indirectly via precipitation data assimilation in atmospheric forecasting models), small, systematic bias can build up over time to a large soil moisture bias.

In an attempt to correct for this bias, we seek to compare the hourly analysis used in Eta/EDAS precipitation assimilation (the NCEP Stage II analysis) against the daily gauge analysis, since the daily gauges are usually more accurate than either radar precipitation estimates or hourly gauge reports.

The daily gauge analyses (12Z-12Z) we use here are the quality-controled product based on 6,000-7,000 gauges, analyzed to 1/8 degree, by NCEP/CPC (we will refer to it as the 'CPCRFC' analysis).

We sum up the hourly analyses to 24h totals, then interpolate both the 24h Stage IV and the CPCRFC analyses to the 12-km Eta grid.

We look at each daily analysis and binned the non-zero (actually non-zero is taken to mean 'at least 0.1mm/day' here) amounts into 26 bins, for daily amounts (mm) of between

      [25.0, largest daily value)

To make the comparison, we first interpolated both the 24h accumulation of the hourly analysis and the CPCRFC daily analysis to the 12km Eta grid, then calculated the 'cumulative precipitation frequency', i.e. for a given x-axis value of n (mm/day), the corresponding y-axis value is the percentage of non-zero rainfall grid points that has rainfall value between 0.1 an n mm/day. Because the biases in multi-sensor precipitation analysis have quite different characteristics in the mountainous, data-sparse West, we make separate calculations for Western and Eastern U.S., with 100W as the dividing longitude (100W is where Oklahoma Panhandle starts, and it bisets the Dakotas and Nebraska).

Our hypothesis is, the systematic biases in a given day's multi-sensor precipitation analysis can, to a large degree, be corrected by this kind of bias statistics compiled for the days preceding it. We plan to test this hypothesis by using a running 21-day bias statistic to correct the current day's hourly analysis. A two-month (Feb-Mar 2002) test run can be found in

The daily statistics were compiled starting 1 Feb, and modifications to the hourly precipitation analysis started on 23 Feb 2002.

The initial test runs show a positive, though likely insufficient impact in addressing the need to correct the low bias in hourly precipitation analysis. Work is underway to improve the algorithm for greater positive impact.