Choose how to read

The full view retains the technical content, figures and sources. The plain-language explanation introduces the ideas without changing the findings.

Terms explained

Rainfall research dashboard

Mockup using executed historical results. Not a live forecast, warning service or validated impact tool.

Coarse NOAA cell centred 17.5°N, 77.5°W. Daily targets 2020–2025. Model selected by validation data 2017–2019; seed 29, physics penalty 0.1.

Rainfall prediction error3.83 mm/dayRoot mean squared error; persistence 4.26 and equal-input nonnegative-clipped linear model 3.81.
Absolute prediction error2.45 mm/dayMean absolute error; persistence 2.54.
Budget alignment3.47 mm/dayApproximate frozen-forcing residual. Not a physical-truth certificate.
Heavy-rain detection37.3%Recall at ≥10 mm/day; persistence 45.8%. Misses 126 of 201 reference days.

Historical rainfall forecast replay

Each point predicts the next daily reanalysis estimate using prior-day information. Revised archives do not prove real-time issue-date availability.

Daily rainfall reference, neural hindcast and persistence over the final 60 days of 2025

Physics and accuracy are separate checks

Across three random seeds, stronger budget penalty lowers residuals but does not uniformly improve prediction accuracy

The largest penalty worsened column-water error. The reference data themselves have 9.96 mm/day residual under this approximate frozen-forcing budget. Better agreement with this simplified constraint is not proof of better atmospheric physics.

All three selected runs

Rainfall root mean squared error: seed 13, 3.945; seed 29, 3.827; seed 47, 3.877 mm/day. The replay uses seed 29, which also has the lowest validation error. Full results for all nine networks.

A simpler model did better

The validation-selected nonnegative-clipped linear predictor has rainfall error 3.81 mm/day, lower than every selected neural run. More model complexity did not establish better accuracy.

Satellite inputs were tested

A separate matched experiment used 640 calibrated infrared images, with 409 training, 76 validation and 155 evaluation dates. Twelve additional networks were trained. Adding image features worsened rainfall error in all three paired selected runs. This does not establish an imagery benefit.

Weather-only versus infrared-feature model errors; image features worsen all three selected runs, beside an actual calibrated infrared input

These are pooled infrared features, not a convolutional image model. Sparse sampling, retrospective availability, calibration transitions and additional parameters limit attribution. All satellite-ablation results.

Actual archived infrared brightness-temperature image, not a rainfall or flood forecast

Simulated impacts

Separate from rainfall forecasting. Crop and flood scenarios require their own assumptions and field validation. No damage, yield or inundation prediction is produced by this dashboard.

Terms explained

Select a term to read its meaning in this study.

climatology

The usual weather for a place and time of year, calculated from a chosen historical period.

hindcast

A prediction-style test using past data. It is not automatically proof that the same inputs were available in real time.

reanalysis

A weather model combined with historical observations to estimate past conditions on a grid.

baseline

A simple comparison method that a more complicated model should be tested against.

persistence

A baseline that assumes the next value will be the same as the latest known value.

root mean squared error

An error measure that squares each difference, averages those squares and takes the square root. Large mistakes count more heavily. Lower is better, but the score alone does not show every kind of failure.

mean absolute error

The average size of the mistakes, ignoring whether the model was too high or too low.

water vapour

Water in its gaseous form in the atmosphere.

moisture budget

An accounting of water entering, leaving and remaining within a chosen part of the atmosphere over time.

regularisation

An added training rule that discourages certain kinds of model behaviour. A rule can help some scores while hurting others.

physics-informed neural network

A neural network trained using information from a physical equation as well as other evidence. A physical penalty does not guarantee that its predictions are correct.

linear model

A model that combines its input features using fitted weights without hidden nonlinear layers.

infrared brightness temperature

A temperature-like value calculated from infrared radiation measured by a satellite. It can describe cloud tops or the surface seen through clear air; it is not a direct rainfall measurement.

training

The stage in which a model adjusts its parameters using designated examples.

validation

A separate set of examples used to choose settings without using the final evaluation results.

holdout

Data set aside for evaluation. Previously studied years are not a completely untouched scientific test.

recall

The fraction of reference cases of interest that a method successfully identifies. Here it describes how many heavy-rain days were detected.

bootstrap interval

An uncertainty estimate made by repeatedly resampling data. Its meaning depends on how the samples are chosen and which sources of uncertainty are included.

ablation

A comparison that changes or removes one part of a modelling pipeline to study its effect. Other differences, such as parameter count, can still complicate interpretation.

field validation

Checking model outputs against relevant independent observations collected in the real setting where the model would be used.