22 April 2020
This may be a question some of you have asked yourself. Perhaps not exactly like that, but working with data in forecasting systems like Delft FEWS to make hydrological predictions, you may have asked yourself a question very much like it. When mixing rainfall data from different datasets, some global, some local, you may be surprised how quickly questions arise; “on which day did this rain actually fall?” Or better still, “on who’s day did it fall?” Perhaps the question never came to mind, and if you really do not like rain, then perhaps it’s best to make a cup of tea and see if there is anything on the news. But if your curiosity has been kindled, here is a little story of why it may matter.
Finding a day in time The fun really starts when working with different time zones, which is where the question “who’s day is it anyway?” comes from. In this story we are working with daily rainfall data in Colombia in South America, which is five hours earlier than Greenwich Mean-Time (GMT-5). As the area we are doing the analysis has few consistent long-term rainfall datasets for hydrological modelling, we are using the CHIRPS satellite precipitation data. That has been shown to work well in Colombia, and the data is freely available from the University of California at Santa Barbara’s Climate Hazards Group. CHIRPS is daily gridded data and provides estimates of precipitation across the globe at a resolution of 0.05 degrees, but in our story, we will follow just one grid cell. Not that we selected a special grid cell; could have been any grid cell, though we did select one that gets the occasional heavy shower, which is not difficult with the tropical downpours this part of the world is used to.
Looking at the CHIRPS data in the NetCDF files we downloaded, we noticed that the time stamp for each full day of satellite rainfall was set at 00:00 GMT+0. So, selecting a nice wet day such as the 24th of November in 2017 for our chosen grid point, we found CHIRPS recorded 56.3 mm. Great! Easy to configure in FEWS using the standard NetCDF input, and there it is in the FEWS database and displays; 56.3 mm on the 24th of November at 00:00. However, in FEWS the convention this is the rainfall that fell starting 00:00 on the 23rd of November and ending 00:00 on the 24th, or one second before that. A quick chat with the researchers in Santa Barbara confirms that in CHIRPS it is, however, considered as the rainfall that fell from 00:00 on the 24th to 23:59 on the 24th. So, there we already have a shift of a whole day! Getting FEWS to import the data with that one-day shift is not easy, but with a bit of tolerance to snap times on import and a transformation it is done. So now the 56.3 mm also imports to the right day in FEWS. What a relief! All good and pretty standard FEWS stuff, I can hear the more experienced users think. But that is the day according to GMT, and not in the Colombian Time Zone. So, more transformations are needed.
Matching the day of the model In a next step, we mapped the daily CHIRPS data to the Colombian day, and we are using the meteorological day, as that is how observed daily rainfall in the country is recorded. The value recorded is the total rainfall that fell over 24 hours, which ends at 07:00 Colombian time (GMT-5), or 12:00 GMT. So exactly in the middle of the day for the CHIRPS data. We are running a model at a daily time step to estimate groundwater recharge, and that also runs to the same day as the observed rainfall. Time for some more transformations!
As we do not know from the daily data when the rainfall fell during the day, our only choice is to divide the CHRPS evenly across the 24th and 25th of November in the Colombian Time Zone. Another easy transformation, and now we are ready to run the model. 1473mm of recharge for the full year of 2017. This really is a wet place! But looking at the statistics of the transformed CHIRPS data we noticed that while the total rainfall for our grid cell over 2017 was the same, the variability was now of course much lower, with much lower peak intensities. And this story is repeated across all grid cells in our domain. The maps of rainfall on the 24th across our study area show that the intensities are much lower for the processed data. Some of the rainfall now spills over into the 25th of November, which in CHIRPS saw very little rainfall. So, it spreads out in space and time. The red dot in the middle of the map is where the grid cell of our story is.
Precipitation amounts for the 24th of November 2017 (GMT). Left the original CHIRPS data and right the data mapped to the 24th of November 2017 in the Colombian time zone (GMT-5). All amounts in mm.
Important? Well yes. It may be important to a model. As with all conceptual models, our recharge model uses an intensity threshold to partition precipitation between direct runoff and infiltration to the unsaturated soil column, where it is then available to evapotranspiration, depending on land use type and potential evaporation. Excess water in the unsaturated soil column then goes to recharge the groundwater. I would say this is more-or-less how most of the models used in forecasting systems work. In the particular model we are using here, the intensity threshold to partition precipitation is derived using the well-known curve number approach (as a function of soil texture, slope, and initial conditions). Schoolbook hydrology.
The difference a day makes But again, the question of on what day did the rainfall actually fall comes up, and with what intensity, and who’s day is that anyway? Considering that rainfall in this very wet and tropical place falls mainly in the (Colombian) afternoon, it would make sense that in the example above the 56.3 mm fell between noon on the 24th and the evening of the 24th, which means all the rainfall should come into the model on 25th at 07:00 GMT-5. So now we just need to snap the whole CHIRPS rainfall amount to the second of the two days. Back to the drawing board and some more transformations. We reran the model, and now our results showed 1699 mm of recharge for 2017. Actual evaporation and direct runoff are of course now lower. So, this place just got 15% wetter than it already was due to a simple change in just one transformation!
The message of the day When managing water resources or predicting floods and/or droughts this is obviously important. This is a wet place so maybe it is not of such great concern as there is way too much water. For arid regions this may be different. As rainfall in arid regions has higher spatial and temporal variability, the differences would be expected to be even greater. Intensities and how these relate to thresholds used in our conceptual models to partition fluxes need careful thought. When for example predicting floods in a flashy catchment in an arid region, the first approach of dividing the data across the two days may never generate a flash flood, while snapping all the rainfall to the single day would generate many more. So, the moral of this short story of a grid cell is that it really is important to think how data is processed and what a given choice in transforming data may mean when using that data in a hydrological model for example. There are I am sure many other and probably better ways to do this, but that is not why I thought it worth it to share the story of our grid cell.
FEWS a great help, but thinking hydrology worth the time of the day Delft FEWS provides an extensive library of transformations in time and space. Which is used and how it is used may have a profound influence on model outcomes and forecast quality. I will not bore you with more such stories, but selecting a transformation like simple Thiessen instead of an approach that takes the change of rainfall with elevation into account can make a real difference. It is also important to know how the model that one is using to create forecasts was calibrated. If the statistical characteristics of the rainfall used to drive the model in the FEWS environment are different than when calibrating, then so will the answers the model gives be different. Of course, it may be difficult to choose which approach to take, especially in the many areas of the world where there is little data. That is where we need to involve the hydrologists. There are other things we can look at, such as hydrological signatures and to make sure it all makes “perfect” hydrological sense. For the story of our grid cell, we made an estimate of the long-term average baseflow in the basin based on a downstream gauge with some years of data and compared that to the long-term recharge. That additional knowledge, as well as what we knew about the climate and that rainfall tends to fall in the afternoons, convinced us to choose the second approach. Turns out that also was easier to do with less FEWS transformations.