A probability of snow is not automatically a fraction of snowfall water
A model assigns a 70% chance that precipitation falls as snow. Does that mean 70% of the water in this particular event is solid? Not necessarily. The number might describe the frequency of an event category, a product-specific estimate, or a fraction accumulated over many samples. Its definition determines what can be calculated from it.
Rain–snow partition matters because precipitation can enter immediate liquid pathways or temporary snow storage. This guide follows a phase estimate into water-equivalent accounting while keeping classification, probability, mixed phase and amount separate. The examples use fictional records and explicitly stated toy models, not operational weather forecasts.
The task is narrower than predicting every form of winter precipitation. Surface conditions, the atmospheric column and the observational method all matter. This educational explanation does not establish road, mountain or ice safety. Use current responsible-authority information for real weather decisions.
Hydrology subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Route 1: Define the output · Chapters 1–3
Route 2: Compare temperature rules · Chapters 4–5
Route 3: Convert to water amounts · Chapters 6–8
Route 4: Check observations and transfer · Chapters 9–10
Route 5: Practise and report · Chapters 11–12
Test elevation, climate and observation changes before transfer
Full chapter index · Independent worked case · Questions and practical checks
Chapter index
Define the output · Chapters 1–3
Start with the phase quantity being predicted
Use water equivalent before comparing rain with snow
Recognise why one surface air threshold is a simplification
Compare temperature rules · Chapters 4–5
Keep wet-bulb temperature distinct from air and dew point
Read a smooth probability model without making it a law
Convert to water amounts · Chapters 6–8
Distinguish an event probability from an event mixture
Weight by precipitation amount when estimating accumulated water
Preserve sub-interval conditions instead of partitioning a mean blindly
Check observations and transfer · Chapters 9–10
Check the observation labels and their support
Evaluate calibration and error consequences separately
Practise and report · Chapters 11–12
Test elevation, climate and observation changes before transfer
Carry phase uncertainty into a bounded hydrological statement
CHAPTER 1 OF 12
1. Start with the phase quantity being predicted
A binary classification chooses between two categories, such as rain and snow, for a defined observation. A probability assigns uncertainty to such a category. A mixed-phase fraction describes how an amount is divided. These are related quantities, but none can silently substitute for the others.
Write the output definition before examining the colour map. Does a value of 0.7 mean snow occurred in 70% of comparable observations? Does it mean an estimated 70% of water-equivalent amount was solid over an interval? Or is it a confidence score that has not been calibrated as a probability? The same numerical value can support different conclusions.
The distinction is not merely terminology. A hydrological model needs amounts entering rain and snow stores. A classifier evaluated on occurrence labels needs a category or category probability. Connecting them requires an explicit interpretation and, where needed, an amount-weighted aggregation or an additional model of mixed precipitation.
CHAPTER 2 OF 12
2. Use water equivalent before comparing rain with snow
Rainfall depth is already a liquid-water amount per area. Snow depth includes air and depends on density. To add snowfall to a water budget, use snow water equivalent or another consistent water-mass measure rather than adding centimetres of snow directly to millimetres of rain.
For a fictional layer 0.10 m deep with mean density 100 kg/m³, water mass per square metre is 10 kg/m². Using approximately 1,000 kg/m³ for liquid water gives 0.010 m, or 10 mm water equivalent. A different density would change the water amount even if visible snow depth stayed the same.
Phase partition divides a known or estimated total water-equivalent precipitation amount. If total amount is 20 mm and the model explicitly assigns an amount fraction of 0.6 to snow, the modelled inputs are 12 mm snow water equivalent and 8 mm rain. Their sum remains 20 mm. A partition rule does not independently correct a biased total-precipitation estimate.
CHAPTER 3 OF 12
3. Recognise why one surface air threshold is a simplification
A threshold model labels precipitation snow below a chosen air temperature and rain above it. This can be a convenient approximation, but the selected threshold is a model assumption, not a universal physical boundary for every observation at every site.
Jennings and colleagues examined variation in rain–snow temperature thresholds using a large observational dataset. Their research supports checking local conditions and humidity rather than assuming one surface-air threshold transfers unchanged everywhere. A concise research result does not replace the method’s definitions, observation basis and scope.
Surface air temperature is one measurement near the end of a falling particle’s path. The particle’s history through the atmospheric column can differ. A two-category surface rule also leaves other precipitation forms outside its explicit model. State those exclusions, especially when observations include mixed categories, freezing rain or ice pellets that should not simply be relabelled to make a binary table complete.
CHAPTER 4 OF 12
4. Keep wet-bulb temperature distinct from air and dew point
Wet-bulb temperature reflects a cooling balance associated with evaporation under specified measurement or calculation conditions. It differs from dry-bulb air temperature and from dew point. A humid or dry environment can therefore change the relationship between an air-temperature reading and a phase inference.
For the following exercises, wet-bulb temperatures are supplied values. No psychrometric formula is implied. If an actual method calculates them from temperature, humidity and pressure, retain the method, input units and uncertainty. Treating a dew-point value as a wet-bulb value because both are expressed in degrees is a quantity error.
A model incorporating humidity can contain information omitted from an air-temperature-only rule, but that does not guarantee perfect classification or universal transfer. Validate the actual rule against suitable observations. More physically relevant inputs improve an opportunity for inference; they do not remove the need to test what the output means.
CHAPTER 5 OF 12
5. Read a smooth probability model without making it a law
Consider the fictional rule p = 1/[1 + exp((T − a)/b)], where p is the probability assigned to snow, T is a specified temperature variable, a is the midpoint and b is a positive transition-width parameter. This is a teaching model, not a fitted formula recommended for any climate.
At T = a, the model gives p = 0.5. With b = 1°C, a temperature one degree above a gives approximately 0.269, while one degree below gives approximately 0.731. Changing a shifts the transition; changing b alters how quickly it changes with temperature. Identify the variable and parameters before substituting data.
If T denotes wet-bulb temperature, substituting dry-bulb temperature changes the model. If a was fitted using one observation definition, applying it to another may change its calibration. A smooth curve can look scientific while still being wrong for the intended quantity. Preserve the fitted domain and test on observations not used to choose its parameters.
CHAPTER 6 OF 12
6. Distinguish an event probability from an event mixture
A probability of 0.7 for snow in a binary event model can mean that, across comparable events, about seven in ten are expected to be labelled snow. It does not assert that each event physically contains exactly 70% snow water. A single event might be entirely rain or entirely snow under that binary definition.
Across many calibrated binary cases with known amounts, an expected snow-water amount can be computed as the sum of pᵢPᵢ if the conditional interpretation is appropriate. Here Pᵢ is each case’s water-equivalent amount. This is an expectation over uncertainty, not an observation of the phase split in each individual event.
A model of genuinely mixed precipitation instead needs a fraction definition suited to that mixture. Do not evaluate a 60% solid-water fraction against a simple present-or-absent label as though they were identical observations. The relationship between labels and amounts must be part of the evaluation, not an assumption hidden after the calculation.
CHAPTER 7 OF 12
7. Weight by precipitation amount when estimating accumulated water
Suppose two events have 2 and 18 mm total precipitation. Their snow-category probabilities are 0.9 and 0.1. The arithmetic mean probability is 0.5, but the amount-weighted expected snow fraction is (2 × 0.9 + 18 × 0.1)/20 = 0.18. The small snowy-likely event and large rainy-likely event do not contribute equal water.
Expected snow input is 3.6 mm and expected rain input is 16.4 mm under the stated interpretation. Multiplying the unweighted mean probability by total precipitation would give 10 mm snow, substantially changing the modelled storage input without changing any measured precipitation amount.
NASA’s explanation of the IMERG ProbabilityLiquidPhase field is a useful reminder to inspect product definitions and aggregation. Do not transplant an interpretation from that product into every phase classifier. A field name, its time scale and its documented meaning determine how the weighting should be handled.
CHAPTER 8 OF 12
8. Preserve sub-interval conditions instead of partitioning a mean blindly
A threshold or nonlinear probability rule applied to an average temperature need not reproduce partitioning each shorter interval. If precipitation amount and temperature change together, the averaging choice can alter the modelled snow input.
In a simple threshold exercise, let a cold hour receive 1 mm and a warm hour receive 9 mm. The rule assigns the cold hour to snow and the warm hour to rain. Snow input is 1 mm out of 10 mm. Applying a single category to the two-hour mean temperature may instead allocate all 10 mm to one phase, depending on the chosen threshold and averaging.
Preserve aligned observations of amount and phase predictors at the resolution needed by the model. This does not mean the shortest available interval is always best: noisy predictors, observation resolution and missing records matter too. The important step is to state how the temporal aggregation changes the quantity being estimated.
CHAPTER 9 OF 12
9. Check the observation labels and their support
A phase observation has a place, time and method. A human report, a station sensor and a satellite-derived product need not sample the same space or interval. Disagreement can arise from real variation as well as classification error. Match observation support before computing a model score.
A crowdsourced report may identify what fell at one place during a short visit. It cannot automatically label every pixel around it for a full hour. Conversely, an area-average field can reasonably differ from a single point. Record the matching rule, excluded ambiguous cases and any spatial or temporal tolerance used.
Jennings and colleagues’ crowdsourced phase study illustrates the importance of observations in the transition zone. Use it as an example of testing methods against evidence, not a reason to assume every volunteer label is exact or that one successful rule has been validated for all regions and seasons.
CHAPTER 10 OF 12
10. Evaluate calibration and error consequences separately
For probability calibration, gather comparable cases assigned a similar probability and compare predicted probability with observed category frequency. If cases assigned 0.8 snow probability show snow about 80% of the time under the same label definition, that supports calibration for that group. A small group supplies limited precision.
Classification accuracy asks a different question: how often does a chosen category agree with the observation? A method that predicts rain every time can score well in a mostly rainy sample while being useless for rare snowfall. Inspect class balance and the errors relevant to the hydrological task.
Hydrological impact also depends on precipitation amount. Misclassifying a 0.2 mm event and a 40 mm event counts as one error each in a simple event-accuracy score, but they can have very different effects on modelled snow storage. Report both the classification evidence and an amount-relevant consequence check rather than assuming one score captures both.
CHAPTER 11 OF 12
11. Test elevation, climate and observation changes before transfer
A rule fitted to one station carries its input and observation history. Applying it at another elevation, climate or sensor regime can change the relationship between predictors and phase. A transfer test should name the changed conditions and preserve independent observations where possible.
Do not tune a threshold on a complete record and then call performance on that same record independent validation. Hold out meaningful periods or locations before fitting, and ensure that tuning does not indirectly use the holdout. A good fit shows compatibility with the training record, not automatic predictive success elsewhere.
A boundary change can be observational as well as climatic. If the definition changes from any snow observed to predominantly snow, the target itself has changed. Apparent phase trends may then include label drift. Keep revisions and method transitions visible so comparisons remain about the same quantity.
CHAPTER 12 OF 12
12. Carry phase uncertainty into a bounded hydrological statement
A phase estimate affects how a model allocates incoming water, but it does not determine all subsequent runoff. Snow retention, melt energy, soil conditions and routing remain part of the water system. An estimate of snow input is not a complete forecast of flood response.
A useful report says: total water-equivalent precipitation was estimated as P; the specified phase method allocated or expected S as snow under the stated interpretation; uncertainty was tested using plausible alternative phase rules. Keep the amount, method, domain and time scale attached to the result.
For independent understanding, explain why a 50% mean probability can correspond to only 18% expected snowfall water in the two-event example. Then change the event amounts and calculate again. If the explanation follows the weights rather than repeating the first percentage, the distinction between occurrence and amount has become usable.
Independent worked case and evidence checks
Consider four fictional intervals with water-equivalent precipitation amounts of 2, 8, 5 and 15 mm. Total precipitation is 30 mm. A binary snow-category probability model assigns probabilities 0.9, 0.6, 0.4 and 0.1. For the first exercise, assume these probabilities have the conditional interpretation needed to calculate expected phase amounts.
Expected snow input is 2 × 0.9 + 8 × 0.6 + 5 × 0.4 + 15 × 0.1 = 10.1 mm. Expected rain input is 19.9 mm. The two amounts sum to 30 mm because each interval’s snow and rain expectations add to its total water amount.
The mean of the four probabilities is 0.5. Multiplying that mean by 30 mm gives 15 mm, not 10.1 mm. The shortcut weights intervals equally even though their water amounts differ. The larger, rain-likely interval carries much of the total precipitation.
The amount-weighted expected snow fraction is 10.1/30, approximately 0.337. It is an expectation under the model, not a measurement that exactly 33.7% of each interval physically fell as snow. Preserve that distinction in any graph caption.
Apply a binary decision threshold of 0.5 to the same probabilities. The first two intervals are labelled snow and the last two rain. This produces a hard model allocation of 10 mm snow and 20 mm rain. Its closeness to the 10.1 mm expectation is specific to these numbers, not a general property of thresholding.
Change only the last interval from 15 to 50 mm. Expected snow becomes 13.6 mm out of 65 mm total, about 20.9%. Hard threshold allocation remains 10 mm snow, about 15.4%. Larger uncertain events can make the difference between hard and expected allocations more consequential.
Now use the toy logistic model with a = 1°C and b = 1°C, explicitly interpreting T as supplied wet-bulb temperature. At T = 0°C, p is about 0.731; at 1°C it is 0.5; at 2°C it is about 0.269. These values follow from the stipulated equation, not from a claimed fitted regional rule.
A fictional observation supplies air temperature 2°C and wet-bulb temperature 0°C. Substituting the correct wet-bulb value gives 0.731. Substituting air temperature into the same rule gives 0.269. The nearly reversed prediction results from using the wrong input quantity, not from a calculation error in the exponential.
For a sensitivity exercise, keep T = 1°C and b = 1°C but change the midpoint a from 0 to 2°C. The predicted probability changes from about 0.269 to 0.731. A threshold parameter uncertainty can matter greatly near the transition. Do not give a highly precise phase amount while treating an untested midpoint as exact.
Test the effect of temporal averaging with the same toy rule and equal precipitation amounts. Two intervals have supplied temperatures −1°C and 3°C. Their probabilities are approximately 0.881 and 0.119, averaging 0.5. The average temperature is 1°C and also produces 0.5 in this symmetric example.
Change the temperatures to 0°C and 4°C. Their probabilities are about 0.731 and 0.047, averaging 0.389. Applying the rule to their mean temperature, 2°C, gives about 0.269. The symmetric agreement in the previous example did not establish a general averaging rule.
Add unequal water amounts of 1 and 9 mm to that changed pair. Expected snow is approximately 1 × 0.731 + 9 × 0.047 = 1.154 mm, allowing for rounding. The corresponding fraction is about 11.5%, neither the unweighted probability average nor the result from a single mean-temperature calculation.
A separate occurrence record contains ten cases, each assigned snow probability 0.8. Eight are observed as snow. This group is compatible with its nominal probability, but ten observations do not prove exact calibration. A second set from a changed region might behave differently. Keep sample size and transfer scope visible.
Now attach amounts to the ten observed cases: the eight snow events each contain 1 mm, while the two rain events each contain 20 mm. Observed snowfall frequency is 80%, but observed snowfall-water fraction is 8/(8 + 40), about 16.7%. Occurrence frequency alone does not describe the amount partition.
Consider an accuracy trap. In a sample of 100 observations, 95 are rain and five snow. A rule that always predicts rain has 95% category accuracy. It detects none of the snowfall cases. Whether it is useful depends on the task, and the headline accuracy should not hide its snow-detection failure.
Change the error weighting. Suppose one missed snow event carries 30 mm and four missed snow events each carry 0.5 mm. The five misses total 32 mm of snowfall water. Treating all five as equal observations is valid for an event-count metric, but it does not preserve their different contributions to water input.
Test a mixed-phase observation explicitly. An independently established 10 mm interval contains 4 mm solid and 6 mm liquid water equivalent. Its solid amount fraction is 0.4. A binary observer who reports snow present has supplied a different label, which cannot by itself verify or reject the 0.4 fraction.
A useful dataset should preserve both when available: presence categories and water-equivalent phase amounts. Do not manufacture a fractional mass from the presence label alone. Similarly, a rain or snow icon cannot automatically represent every minute in an interval that changes phase.
For an amount sensitivity check, total precipitation is estimated as 20 mm and the explicit snow-amount fraction as 0.6. Snow input is 12 mm. If total precipitation is instead 24 mm with the same fraction, input becomes 14.4 mm. If total stays 20 mm but the fraction changes to 0.4, input becomes 8 mm. Amount uncertainty and phase uncertainty are separate levers.
Combine alternatives carefully rather than calling them a statistical confidence interval. Total precipitation from 18 to 22 mm and a snow-amount fraction from 0.4 to 0.7 produce corner values of 7.2 and 15.4 mm. These are sensitivity bounds under the stated joint possibilities, not a 95% interval supplied by data.
Finally place the result into a toy snow store. Begin with 25 mm snow water equivalent, add the original 10.1 mm expected snow input and subtract an independently stipulated 4 mm snow-water loss. The expected end store is 31.1 mm under that simplified account. The 19.9 mm rain input must be routed or retained according to a separate model; it does not disappear because the snow account focuses on solid input.
Test sensitivity to a hard decision boundary as well. Two fictional cases each contain 10 mm precipitation and receive snow probabilities 0.49 and 0.51. At a 0.5 threshold, one is assigned all rain and the other all snow. Their expected snow amounts are instead 4.9 and 5.1 mm under the stipulated conditional model. A small change in probability can cause a large categorical change without an equally large change in expectation.
This is not an argument that hard categories are never useful. It is a reason to retain the underlying probability and decision rule when a downstream calculation depends on them. If a report keeps only icons, another reader cannot recover the difference between a marginal classification and a confident one.
The defensible conclusion is therefore a modelled or expected water allocation with a defined interpretation. A student should be able to identify which examples concerned frequency, probability, amount fraction, total amount and subsequent storage. Their connections matter precisely because they are not the same quantity.
Worked comparison
| Number | Meaning in its stated example | Incorrect automatic interpretation |
|---|---|---|
| p(snow) = 0.7 | Probability of a binary observation label | Every event contains 70% solid water |
| 8 snow cases out of 10 | 80% observed occurrence frequency | 80% of precipitation water is snow |
| 4 mm solid out of 10 mm | 40% solid-water amount fraction | Only 40% confidence that snow occurred |
| Mean probability 0.5 | Equal weighting of interval probabilities | Amount-weighted fraction without checking amounts |
| Snow input 12 mm | Water-equivalent allocation under a specified model | Complete prediction of runoff |
Questions and practical checks
Does snow require surface air below 0°C?
A single surface-air cutoff is a modelling simplification. Humidity, particle history and atmospheric structure matter. This guide does not offer a universal threshold.
Is wet-bulb temperature the same as dew point?
No. They represent different physical quantities. Use the variable required by the documented method, with compatible input units and assumptions.
Can a category probability be used for expected water amounts?
Only with an appropriate conditional interpretation and amount weighting. The expected amount is not a direct observation of a mixed-phase fraction in each event.
Does high overall accuracy prove snowfall detection?
No. A rain-dominated sample can reward an always-rain classifier. Inspect class-specific errors, probability calibration and water-amount consequences.
What should be checked before moving a rule to another site?
The predictor definitions, observation labels, climatic domain, time and spatial support, and independent performance. A good training fit is not independent transfer evidence.
Continue with the connected subjects
Follow snow input into storage and melt
Snow water equivalent: compare water mass rather than snow depth
Snowmelt and rain-on-snow: phase input is only the beginning
Canopy snow: temporary stores and water delivered below
Check weather quantities and observations
Dew point versus wet-bulb temperature: keep predictors distinct
Freezing rain and ice pellets: why a binary surface rule does not cover every precipitation form
Rain-gauge undercatch: phase partition does not repair total precipitation bias
Apply the connected model
Snow lysimeters: follow phase inputs through storage and basal drainage
Sources and scope
Jennings and colleagues, 2018: spatial variation in rain–snow temperature thresholds
Jennings and colleagues, 2023: crowdsourced observations in the rain–snow transition zone
NASA GPM: interpreting IMERG ProbabilityLiquidPhase and aggregation
The numerical cases in this guide are original fictional teaching examples, not measurements from a real site. This is an educational explanation, not a field safety instruction, engineering approval or an official certification. For a real project, use qualified practitioners, current product documentation and site-specific evidence. Sources were checked on 30 October 2026.