Short-term Precipitation Forecasting with Radar and Gauge Data: Going Beyond the Z-R Relation
This work develops statistical models which incorporate radar reflectivity and rain gauge data for 30-min ahead precipitation forecasting across Singapore. A probabilistic forecast model is developed which predicts both the occurrence and amount of precipitation. The models developed here utilize: the past dBZ values of various layers (1 to 5 km); past ground observations in a autoregressive structure; future extrapolation of dBZ values using storm-tracking algorithms. We call this method ZP-RP where P stands for process in contrast to the Z-R relationship which inputs a single dBZ value and outputs a single precipitation amount. Since the users of the models are often interested in accuracy measures such as POD/FAR (probability of detection/false alarm rate), we develop a post-transformation method to optimize the output of the statistical models further for the end users.
The table below presents the performance of various models using test data (which were not used in the calibration of the models) for Singapore precipitation during the Inter-Monsoon season in 2011. Three categories of rain intensity are considered for calculating the POD and FAR: Low intensity (less than 20 mm/hr); Heavy intensity (between 20 to 50 mm/hr); Extreme Intensity (larger than 50 mm/hr). The first row of the table shows the accuracy obtained by applying the Marshal-Palmer (MP) relation to 6 observed 5-min dBZ values in 30 minutes to calculate rain values in the same 30 minutes (which is not forecasting but we include that as a baseline). The second row contains the values for Marshal-Palmer relation applied to the most recent available dBZ in the past to predict the rain intensity in the next 30 minutes. Both of these rows show unsatisfactory performance in the heavy and extreme categories. Rows three to five of the table contain the output of the models developed in this work with various data components. The third row only utilizes the past observed dBZ; the fourth also utilizes the past rain gauge data; the last row uses rain gauge data and the output of a storm-tracking algorithm (based on the image pattern matching techniques). We observe a very significant improvement in using the ZP-RP model even without the rain gauge data and the storm-tracking algorithm's output.
|Model||POD / FAR (Low, Heavy, Extreme)|
|MP (current dBZ)||100 22 0 / 5 36 100|
|MP (dBZ lag 1)||98 18 3 / 5 59 89|
|ZP-RP||86 62 44 / 1 80 73|
|ZP-RP (with gauge data)||83 64 45 / 1 83 71|
|ZP-RP (with gauge & storm-track.)||85 67 41 / 1 81 68|