climate-weather/ai produced the result/PLoS ONE 2026 · v2
Machine learning turns satellite cloud images into hourly rainfall maps for central Vietnam
Researchers built an hourly rainfall product for four coastal provinces of central Vietnam by combining satellite infrared images, weather reanalysis data and elevation. Tree-based machine learning models decided which pixels were raining and how hard.
spectrum · one line per step, placed by what the step does · bright lines used AI
Enhancing rainfall estimation accuracy with machine learning, cloud masking, and multi-source data: A case study of four coastal provinces in central Vietnam
PLoS ONE, 2026
doi:10.1371/journal.pone.0342825 · record aix-00288 v2 · checked 2026-11-05
- AI was for
- Classification, Property prediction
- Model family
- Gradient-boosted trees, Random forest
- Checked by
- Held-out7851 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Knowing how much rain is falling, where and when, is harder than it sounds. Rain gauges measure it accurately but only at the spots where they stand, and gaps between them can be wide. Weather satellites see everywhere, but they cannot see rain directly. Instead they measure infrared radiation from the tops of clouds, which tells you how cold and therefore how high a cloud top is. Cold, tall clouds often mean heavy rain, but not always: some high clouds give nothing, and some modest ones pour. Turning cloud-top temperature into a rainfall rate is therefore an estimate built on an indirect clue.
The team set out to make that estimate for four coastal provinces in central Vietnam, a region exposed to heavy rain. They assembled infrared imagery from the Himawari-8 satellite, atmospheric fields from the ERA5 reanalysis (a reconstruction of past weather), and ground elevation from a satellite terrain map. Readings from 175 automatic rain gauges between 2019 and 2023 served as the answers to learn from. Everything was put on a shared grid of about four kilometres, one step per hour.
Where AI came in
The rainfall maps are the direct output of the trained models. Three tree-based learning methods, LightGBM, XGBoost and Random Forest, were each trained from scratch on the data. Rather than predicting rainfall in one go, the researchers split the job into stages: first decide whether a pixel is raining at all, then whether the rain is light or heavy, then sort it into one of four rainfall classes. A separate model for each class then estimated the actual rate in millimetres per hour, and the four sets of estimates were stitched into one map.
The models stand in for the fixed formulas and lookup rules that have traditionally converted cloud-top temperature into rainfall. Alongside them the team kept simple rule-based filters, thresholds on infrared brightness, to screen out pixels whose clouds were unlikely to be raining. The results were checked against gauge readings held back from training, against weather radar imagery, and against four existing satellite rainfall products, using a set of 7,851 rainfall maps.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The study builds an hourly rainfall product for four coastal provinces of central Vietnam from Himawari-8 infrared brightness temperatures, ERA5 reanalysis fields, ASTER DEM elevation and 175 automatic rain gauges over 2019–2023. Gradient-boosted and random-forest models were trained in a three-stage scheme that first separates rain from no-rain, then low- from high-intensity rain, then four rainfall classes, with a separate regression model estimating rate within each class; brightness-temperature cloud masks were applied to the first two classification steps. Evaluated against gauge observations on a 20% held-out set, the LightGBM version of the product reached a Critical Success Index of 0.55, a Probability of Detection of 0.74, a correlation coefficient of 0.47, a modified Kling-Gupta Efficiency of 0.47, a mean absolute error of 2.66 mm/h and a root mean squared error of 5.48 mm/h. Among the four comparison satellite products, GSMaP_MVK_Gauge had a lower mean absolute error than the proposed product, and IMERG Early Run had a higher Probability of Detection than it in some individual events.
How AI was used
Multi-source inputs were resampled to a common 4 km, 1-hour grid, with gauge readings as labels. Thirteen input features were chosen either as a fixed set of cloud- and rain-forming variables applied to every model or, in a comparison configuration, by Random Forest importance ranking from Scikit-learn. Class imbalance was handled by randomly augmenting minority samples across eleven rainfall intervals for the rain/no-rain model and by class weighting for the three intensity classifiers. LightGBM, XGBoost and Random Forest were each trained from scratch with grid-searched hyperparameters and five-fold cross-validation on 80% of the data, producing four classifiers (M1–M4) in a three-stage cascade and four class-specific regression models (M5–M8). Thresholds on Himawari-8 infrared brightness temperatures and channel differences formed two rule-based cloud masks that were intersected with the outputs of the rain/no-rain and intensity classifiers. Samples surviving classification were passed to the corresponding regression model, and the four regression outputs were merged into the final rainfall maps.
The shape of the work
Structural · the record, drawn
Assemble and resample multi-source inputs
Cleaning, filtering, normalising or labelling data already obtained.
both the input datasets and the comparative datasets were resampled to a common spatial resolution of 4 km and a temporal resolution of 1 hourwhere the paper describes this · verbatim
Select features and balance classes
Encoding data into features, descriptors, embeddings or graphs.
the common feature set used to train models M1 through M8 consists of 13 featureswhere the paper describes this · verbatim
Train classification and regression models M1–M8
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
the k-fold cross-validation technique (k = 5) is employed for model training and validation to prevent overfittingwhere the paper describes this · verbatim
Classify rain/no-rain and rainfall intensity classes
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
the first model classifies the data into rain and no-rain regions, while the second model further classifies the rain regions into different rainfall intensity classeswhere the paper describes this · verbatim
Apply brightness-temperature cloud masks CM1 and CM2
Reducing a candidate set by filtering or ranking, in a single pass.
A pixel is classified as a rain-bearing cloud if the BT values of all the aforementioned channels fall within their respective ranges.where the paper describes this · verbatim
Estimate rainfall rates per class
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
the corresponding samples were transferred to models M5 through M8 to perform rainfall regressionwhere the paper describes this · verbatim
Evaluate against gauges, radar and satellite products
Testing outputs against ground truth.
The dataset used for evaluation accounts for 20% of the original dataset and is independent from the training dataset.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The rainfall product the paper reports is produced by the trained classification and regression models; the reported classification and error metrics are metrics of those models' outputs.
Three machine learning algorithms were investigated in this study: LGBM, XGB, and RF.where the paper describes this · verbatim
as well as with four regional satellite-based rainfall products using a dataset of 7,851 rainfall mapswhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of LightGBM (LGBM)Which version of the model was used is not stated.
- Version of XGBoost (XGB)Which version of the model was used is not stated.
- Version of Random Forest (RF)Which version of the model was used is not stated.
About this article
Record aix-00288, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Machine learning turns satellite cloud images into hourly rainfall maps for central Vietnam. Record aix-00288, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00288/