Abstract

The quality of runoff/streamflow modelling is critically dependent on the quality of rainfall forcing data and a quantitative understanding of its uncertainty. More accurate rainfall data with clearly defined uncertainty has implications for water resource management and flood prediction. This is important for the Lake Eyre Basin (LEB) in central Australia where the landscape is very flat and hydrophobic soils are common. The LEB is also of interest in terms of climate given that it is predominantly semi-arid/arid with a large north-south rainfall gradient. The north is subject to intense tropical (summer-time) based rainfall events, with weaker predominantly winter rainfall occurring in the south where annual totals are approximately 3-4 times lower on average than in the north. The remoteness of this region means that the in- situ rainfall observation network is sparse compared to more densely populated regions across Australia's eastern seaboard, hence the potential of remotely sensed spatial rainfall data to provide useful information here needs to be examined. Three real-time satellite precipitation products, the TRMM Multi-satellite Precipitation Analysis 3B42RT (TRMM-RT) Version 7, CPC Morphed precipitation (CMORPH) Version 1, and Precipitation Estimation from Remote Sensing Information using Artificial Neural Network (PERSIANN), are assessed against in-situ rain gauge data over a nine-year study period to examine characteristics of their error in LEB. An up-scaled version of the Australian Water Availability Project (AWAP) spatial rainfall product (interpolated from gauge data) was also assessed against the same gauge data used for the satellite products, as a benchmark for comparing the satellite product assessments. All data were mapped to a 0.25⁰ resolution grid and aggregated to daily time step rainfall totals (9am to 9am), with errors calculated as differences in daily rainfall between the products and gauge data for the nine-years. Products were assessed by inter-comparing the distribution of their daily errors via boxplots, and through bias (average error) and root mean square differences (RMSD) over the nine-years. The ability of satellite products to detect rain gauge measured events was also examined via calculation of the Probability of Detection (POD) and False Alarm Ratio (FAR) metrics. Based on error distributions for each rainfall product over the full nine-years, TRMM-RT is the closest match to gauge observations of the three satellite products. Its mean bias is ~0.3 mm/day, approximately 3.5 and 4.5 times less than that of CMORPH and PERSIANN respectively, while the unbiased Root Mean Square Difference (RMSD) of ~4.3 mm/day for TRMM-RT is less than half that for both CMORPH and PERSIANN. Comparisons of product error distributions show consistently lower variation and less extreme outlier values occurring for TRMM-RT error. In addition, while the median TRMM-RT error displays a trend of increasing negative bias with increasing rainfall totals, the median CMORPH and PERSIANN errors show the opposite and steeper trend. Common amongst all of the products is an increase in absolute error ranges with increasing daily rainfall totals, in addition to absolute errors calculated over wet season months (mostly summer) having a generally larger spread than over dry season months (mostly winter) in the basin's tropical north. The results provide insights into how error for a product such as TRMM-RT may be best modelled. There are a number of extreme errors in all products, shown as outliers in boxplots, and mostly implying large over-estimates of rainfall. By defining extreme outlier error values as >20 mm/day and/or >100% of gauge values, a sample of these were found to occur at or near the edge of major rainfall systems as delineated by the TRMM-RT products, indicating they may be an artifact of spatial inaccuracy in resolving the edges of rainfall extents, warranting further study. POD and FAR metrics for TRMM-RT further demonstrated differences in performance across ranges of daily rainfall totals, and between tropical wet and dry seasons in the LEB. POD results show there is generally greater chance of detecting the presence of larger rainfall accumulations. From FAR statistics, false alarms are generally more prevalent amongst smaller TRMM-RT estimates and in dry seasons, where FAR values decrease in relation to increasing average TRMM-RT estimates.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call