Does finer input data lead to a better model accuracy?

Question

Climate and chemical transport models all need input files including meteorology data, emission inventory, surface data (e.g. land use type) to initialize the simulation field.

Take the case of WRF for regional meteorology data. I always use the 1 degree x 1 degree NCEP FNL analysis data in grib2 format as the model-driven data supporting the boundary and initial information every 6 hours.

Now, I find that the ECMWF interim reanalysis data with spatial 0.25 degree x 0.25 degree can also be used as the WRF input.

So, I'm wondering does finer input data induce better model performance? How much can be improved if the answer is true?

I have thought out some pros of finer input file (in both spatial and temporal scale):

pros: (1) finer input field means the physical process of smaller scale(e.g turbulent transport of water) can be depicted better in some extent; (2) finer temporal variation of chemical species can emulate the real emission process better(e.g, emission related to traffic in a day).

Are there any cons that come with the performance improvement except the computational cost?

Answer 1

My experience with the WRF is in tropical cyclone simulation, but here are some things to consider:

• Smaller scale information might already be resolved in your higher resolution initial conditions, improving the spin-up time of effects you are trying to model
• I believe the ECMWF resolution has been statistically downscaled from a 0.75 degree dataset, I might be wrong on that one though. That would mean that your 0.25 degree is just some interpolated data
• Some models perform better than others in specific regions
• Higher resolution is not better if your parameterization is wrong
• You can find caveats of specific datasets here: PSD Gridded Climate Datasets.

Answer 2

Yes, of course finer input data can lead to better model accuracy, but only if the model's conceptualization is close to the mark. Of the many problems of climatic model misinterpretation, I would emphasize the following:

a) The implied precision of most climatic modelling output is absurd, so take a long, hard and skeptical look at the data in relation to realistic error bounds.

b) Down-scaled models tend to give an impression of higher accuracy. For the most part, this is illusory.

c) Halve the grid dimension and you require four times as much input data. In reality there usually isn't that much good quality data available, and even if there is (in principle), there usually isn't the manpower capability to process it.

d) Reduce the grid size, and you have to pay much closer attention to the conceptualization. For example, physiographic impacts upon rainfall, land-water ratio, and microclimatic factors become more important.

In my experience of working on AOGCM outputs in East Africa, I found that smaller grids only highlights local climatic discrepancies between model and evidence-based instrumental data. Even with smaller grids, the models totally fail to capture meso-climatic impacts of large lakes, or the steep climatic gradients over a scale of about 100 km. So the message is clear: global climatic models are great for general climatic trends, dubious for rainfall trends, and downright misleading for most small scale climate trend analysis. My worry is that one of leading categories of GCM users, water resources analysts, don't yet seem to have a clear grasp of GCM limitations. There are dangers in treating climatic computer models as a 'magical black box'. There are inherent dangers in gridding the available data realistically (a major source of error!). And whenever you choose a finer grid size, always pay very close attention to differences between hind-cast outputs and the actual spread of climatic data on the ground.

Answer 3

To begin, climatology does not have real 25km x 25km data, since they have only about 6000 meteostations, while the 25x25 grid on dry land amounts to 240,000 stations. Artificial make-up of this kind of data must use variety of assumptions that might be not necessary true or coherent. Feeding a model with goofy data would accomplish nothing.

Second, finer input data do nothing if the model does not resolve dynamic patterns of the same scale. Unless you see a model that correctly reproduces main details of global atmospheric circulation (I mean just crude topology of vortices-hurricanes and their average frequency), fine input is meaningless. More, a good correct model should start with any realistic and non-detailed data, and develop a stable realistic trajectory with realistic means. Climate and transport data should not need any "input", the model must generate the data on its own.