~/aixsci
weekly updates

climate-weather/ai produced the result/arXiv 2026 · v2

Satellite image dataset extended to four channels for typhoon strength prediction

Researchers expanded the Digital Typhoon Dataset to four satellite channels, then trained convolutional neural networks to estimate storm central pressure and wind speed from images, and examined how the errors varied with storm strength.

1. Assemble satellite images and best-track labels2. Build single-channel and multi-channel datasets3. Split by season into training and evaluation sets4. Train ResNet regression models5. Predict pressure and wind on held-out seasons6. Compute benchmark RMSE across configurations7. Analyse intensity-dependent bias by sliding-window retraining8. Retrain with imbalance-aware weighting schemes

spectrum · one line per step, placed by what the step does · bright lines used AI

Digital Typhoon Dataset V3: Benchmarking for Multi-spectral Extension and Analyzing Systematic Bias in Intensity Regression
arXiv, 2026

doi:10.48550/arxiv.2610.21654 · record aix-00287 v2 · checked 2026-11-05

ai-resultrole of AI
AI was for
Property prediction
Model family
Convolutional neural network
Checked by
Held-out
Code
available

The finding the paper is about came from the AI.

read as

The science is explained before the AI appears. Switch to field specialist to go straight to the method.

Assumes the discipline and goes straight to the method.

The diagram, the record and what the paper did not report are identical in both modes. Only the framing changes, never the evidence.

Introduction by AIxSci · plain language

What this research was about

Weather agencies record the strength of a typhoon by two numbers: the central pressure at its eye, where lower pressure means a stronger storm, and the highest sustained wind speed near it. Out over open ocean there is rarely an aircraft or a buoy to measure either directly, so forecasters read the storm's strength from satellite pictures instead. That reading is hard because the link between what a storm looks like from above and how strong it is underneath is indirect. Weather satellites such as the Japanese Himawari series see in several separate channels, each sensitive to different wavelengths of light, from visible sunlight to infrared heat and to the wavelengths absorbed by water vapour high in the atmosphere.

The team behind the Digital Typhoon Dataset, an archive of satellite images cropped and centred on individual storms, released a third version. It adds three further Himawari channels, a second infrared, a water vapour and a near-infrared one, to an archive that previously held just one, covering storms in the Western Pacific and the Australian region. They then set out to measure how well standard image models can read storm strength from these images, and whether feeding in several channels at once helps compared with a single one.

Where AI came in

The AI here is a pair of ResNet image networks, ResNet-18 and ResNet-50, widely used designs that learn from many labelled examples which visual patterns go with which answer. The researchers trained them from scratch on the dataset's images, each labelled with the pressure or wind value from official best-track records, the after-the-fact storm histories compiled by weather agencies. The 2024 and 2025 seasons were kept aside, so the trained models were tested on storms they had never seen. Every finding in the paper, which channels and which network did best, and how the errors behaved, comes from models the authors trained and ran themselves.

Multi-channel inputs did not clearly beat a single channel for pressure: the lowest single-channel error was 6.56 hectopascals in the Western Pacific against 6.83 for the best multi-channel setting. A further set of runs retrained one network separately for each of 47 consecutive-year pairs, which showed the error depending on the true pressure, with the strongest storms estimated as weaker than they were. Reweighting the training examples so that rare, intense storms counted for more shrank that error but did not remove the pattern.

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

Version 3 of the Digital Typhoon Dataset adds three further Himawari channels (a second infrared, water vapour and near-infrared) to the previously single-channel archive for the Western Pacific and the Australian region. The authors trained ResNet-18 and ResNet-50 on typhoon-centered images to regress central pressure and maximum wind, holding out the 2024 and 2025 seasons, and compared single-channel against multi-channel inputs on the four-channel intersection dataset. Multi-channel inputs did not clearly improve pressure regression; the lowest reported single-channel pressure RMSE was 6.56 hPa for the WP dataset and 7.20 hPa for the AU dataset, both with ResNet-18, Channel 1 and full-resolution images, while the best multi-channel results were 6.83 hPa on WP and 8.44 hPa on AU. A sliding-window analysis over 47 consecutive-year pairs of the WP Channel 1 data showed prediction error depending on true pressure, with strong typhoons underestimated, and inverse-frequency and inverse-density weighting reduced the magnitude of this error without removing the pattern.

How AI was used

Satellite images and best-track records were assembled into per-channel datasets, and a multi-channel dataset was formed from observation times at which all four channels were available; observations with missing channels were discarded rather than filled. Data were split at season level, reserving 2024 and 2025 as the held-out evaluation set. ResNet-18 and ResNet-50 were trained with the Adam optimizer for 50 epochs to regress central pressure, and wind for the single-channel comparison, across four channels, three image types (full 512x512, resized 224x224, centre-cropped 224x224) and three multi-channel input combinations (Channels 1+2, 1+3, 1+2+3), with each experiment repeated five times on the same partitions. Trained models were run over the held-out seasons and RMSE was aggregated per configuration. For the bias analysis, ResNet-18 with full-resolution Channel 1 images was retrained for each consecutive-year pair from 1978 to 2024, excluding the target pair from training, and the mean error was examined against true pressure in 10 hPa bins. Further runs retrained the same model with weights inversely proportional to bin frequency, an epoch-wise adaptive update using per-bin training RMSE, a weighted random sampler, and a loss-only inverse-density weighting inspired by BalancedMSE.

The shape of the work

Structural · the record, drawn

ACQUISITIONPREPARATIONPREPARATIONTRAININGINFERENCEVALIDATIONINTERPRETATIONTRAINING12345678AIAIAIAIAssemblesatellite imagesand best-track l…Buildsingle-channeland multi-channe…Split by seasoninto training andevaluation setsTrain ResNetregression modelsPredict pressureand wind onheld-out seasonsCompute benchmarkRMSE acrossconfigurationsAnalyseintensity-dependentbias by sliding-…Retrain withimbalance-awareweighting schemes
AI stepNo AI↤ what the AI stood in for

Assemble satellite images and best-track labels

Obtaining raw data, whether by measurement, download or retrieval.

For the Northern Hemisphere, we added data for the 2025 typhoon season (January-December 2025) using best-track data from the Japan Meteorological Agency (JMA).where the paper describes this · verbatim
in the paper

Build single-channel and multi-channel datasets

Cleaning, filtering, normalising or labelling data already obtained.

An observation is retained only when Channels 1, 2, 3, and 4 are simultaneously available at the same observation time.where the paper describes this · verbatim
in the paper

Split by season into training and evaluation sets

Cleaning, filtering, normalising or labelling data already obtained.

we reserve the last two seasons, 2024 and 2025, as the held-out evaluation set, and use all earlier seasons for trainingwhere the paper describes this · verbatim
in the paper

Train ResNet regression models

Fitting model parameters, including fine-tuning an existing model.

we performed all subsequent experiments in this paper using Adamwhere the paper describes this · verbatim
in the paper

Predict pressure and wind on held-out seasons

Running a trained model over new data to predict, classify or score.

we evaluated the Digital Typhoon Dataset V3 for pressure and wind regression taskswhere the paper describes this · verbatim
in the paper

Compute benchmark RMSE across configurations

Testing outputs against ground truth.

report the mean RMSE and standard deviation across the five runswhere the paper describes this · verbatim
in the paper

Analyse intensity-dependent bias by sliding-window retraining

Extracting understanding from model behaviour.

we trained the WP-1 dataset using ResNet-18 with full-resolution images for each consecutive-year pair from 1978 to 2024where the paper describes this · verbatim
in the paper

Retrain with imbalance-aware weighting schemes

Fitting model parameters, including fine-tuning an existing model.

we assigned initial weights inversely proportional to the sample frequency in each pressure binwhere the paper describes this · verbatim
in the paper

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.

~Role of AI
AI produced the resultour reading

The paper's reported findings (channel and architecture comparisons, intensity-dependent bias) are properties of models the authors trained and ran; no result is independent of them

+What the AI was for
we compare base model architectures between ResNet-18 and ResNet-50, two widely used baseline vision modelswhere the paper describes this · verbatim
+Model families
+How it was taught
Supervisedin the paper
+Models named
ResNet-18 · Trained from scratchResNet-50 · Trained from scratchin the paper
+How results were checked
Held-outin the paper
We use the held-out evaluation set to monitor model performance during training and report final benchmark resultswhere the paper describes this · verbatim
+Code · weights · data
code availableweights not reporteddata availablein the paper
under an open data license: the Creative Commons Attribution 4.0 International (CC BY 4.0) Licensewhere the paper describes this · verbatim
+Compute
not reportedin the paper

What this paper did not report

Technical · absence is published deliberately

Reported as not stated: 9 items
  • Trained model weightsWhether the trained model is available is not stated.
  • ComputeThe hardware or time used is not stated.
  • How many were testedThe paper gives no count of what was tested.
  • Version of ResNet-18Which version of the model was used is not stated.
  • Version of ResNet-50Which version of the model was used is not stated.
  • What step 4 replacedThe paper gives no basis for what the AI stood in for.
  • What step 5 replacedThe paper gives no basis for what the AI stood in for.
  • What step 7 replacedThe paper gives no basis for what the AI stood in for.
  • What step 8 replacedThe paper gives no basis for what the AI stood in for.

About this article

Record aix-00287, 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). Satellite image dataset extended to four channels for typhoon strength prediction. Record aix-00287, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00287/