Rainfall looks like a simple number, millimetres of water on the ground. It is hard to measure well, because rain falls unevenly. One storm can drop 80 mm on one village and 10 mm on the next one, 5 km away. Every rainfall dataset is therefore a compromise between what is measured, what is estimated and what is modelled, and the compromise shapes what the data can answer.
Four sources of rainfall numbers
Rain gauges catch the water in a funnel and measure it. A gauge is the most direct measurement, and it is the reference that the other products are checked against. It tells you the rain at one point only. Gauges are sparse in many regions, and the wind blows some rain past the funnel, so gauges tend to read a little low in storms.
Weather radar sends out a beam and measures how much of it bounces back from raindrops. It sees whole storms at a fine scale, every few minutes. It covers a radius of a few hundred kilometres around the radar, and the conversion from echo to millimetres needs calibration with gauges.
Satellite estimates infer rain from what a satellite sees. Thermal infrared sensors see the cold tops of tall rain clouds. Microwave sensors see through cloud to the ice and water inside it. Neither sees rain at the ground, so the estimate is calibrated against gauges and radar. Satellites cover the whole globe, which is their strength, and they handle light, warm rain over hills less well.
Reanalysis is a weather model that is run over the past and nudged toward the observations it is given. It gives a complete, consistent record of many variables on a regular grid, for any place and any hour, including those with no instruments. The rainfall in a reanalysis is a model product, produced by the model's physics between observations. It is not a measurement.
Three products you will meet
| Product | Type | Grid | Period | Account |
|---|---|---|---|---|
| CHIRPS | Satellite estimates blended with gauge data | 0.05 degrees, about 5 km | 1981 to the present | None |
| GPM IMERG | Satellite estimates, adjusted with gauges in the final run | 0.1 degrees, about 11 km | 2000 to the present | Free NASA Earthdata account |
| ERA5 and ERA5-Land | Reanalysis | 0.25 degrees (ERA5) and 0.1 degrees (ERA5-Land) | 1940 and 1950 to the present | Free account for the Climate Data Store, none for Open-Meteo |
CHIRPS combines infrared satellite data with a long-term rainfall climatology and with gauge records, and it is made for drought and agriculture work on land between 50 degrees north and south. IMERG merges data from the constellation of GPM microwave satellites, and it comes in an early, a late and a final run. The early run appears within hours and has the least gauge correction, and the final run arrives after months and has the most. ERA5 is the reanalysis of the European Centre for Medium-Range Weather Forecasts, and ERA5-Land is a finer land-only version of it.
Resolution, latency and units
Resolution means the size of one cell. A value is an average over the whole cell, so a 5 km cell cannot tell you the rain of one field. A coarser product smooths out the storms and the peaks.
Latency is the delay between the rain and the data. A reanalysis or a final satellite run can be days or months behind. For a flood warning you need an early run, and for a study of last year you want the final one.
Units and time are where mistakes happen. IMERG monthly files give a mean rate in millimetres per hour, and you multiply by the hours in the month. CHIRPS monthly files give millimetres per month. Daily files in some products start at midnight UTC, which is 5:30 in the morning in India, so a storm in the evening can land in the next day. Always check the unit, the accumulation period and the time zone.
Why two good products disagree
The rainfall guide compares CHIRPS and IMERG for one place in Punjab over 2023. The two put the peak in the same month, July, and have a correlation of 0.84 across the twelve months. They differ by 67 mm in the year total, and by large factors in some months. This is normal, for four reasons: the cells are different sizes and sit on different ground, the two use different satellites and different gauges, the ways of correcting with gauges differ, and rain itself changes within a cell.
Weather variables other than rain are more reliable in a reanalysis. Air temperature and pressure change smoothly in space, and the model reproduces them well. Rain, wind gusts and cloud are the hard ones, because they depend on small-scale processes.
How to choose
- To follow drought, vegetation response or crop seasons over decades, use a long blended product such as CHIRPS.
- To follow a recent event or a near-real-time situation, use IMERG early or late, and say that it is a first estimate.
- To get temperature, wind, humidity and soil moisture together for one place, use a reanalysis, such as ERA5 through the weather guide.
- To get the truth at one place, use a rain gauge, and use the grid to understand the surroundings.
Common mistakes
- Treating a gridded value as the rain at a point.
- Mixing products in one time series. A change in the numbers can then be a change of product.
- Reading a model rainfall as a measurement.
- Forgetting the unit and the accumulation period.
- Comparing two products without saying the cell size.
What to report
State the product and its version, the cell size, the period, the unit, the date you downloaded the data and how you picked the cell for your place. With that, another person can repeat your work, and can see what the numbers can and cannot support.