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

Asanobu Kitamoto; Victor Le Louarn; Noelia Falczuk; Ngoc Anh Tong; Anja Veleva

cs.CVOct 15, 2026

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

Authors: Asanobu Kitamoto, Victor Le Louarn, Noelia Falczuk, Ngoc Anh Tong, Anja Veleva

Organizations: National Institute of Informatics, Japan · Typhoon Science and Technology Research Center, Yokohama National University, Japan · University of Artois, France · University of Buenos Aires, Argentina ·

Abstract

This paper presents the Digital Typhoon Dataset V3, a new version of the longest typhoon satellite image dataset spanning nearly 50 years, designed to benchmark machine learning models on long-term spatio-temporal data. The latest addition to Dataset V3 is multispectral satellite data, enabling multi-channel models. We investigate data-splitting strategies and present benchmarking results for the Northern and Southern Hemispheres. Experimental results show that, under the current dataset construction and experimental setting, multi-channel inputs do not clearly outperform single-channel inputs for pressure regression. Among the evaluated model architectures, image types, and spectral channels, ResNet-18, full-resolution images, and Channel 1 generally provide the best single-channel performance for pressure regression. Performance also differs between the WP and AU datasets by regression target, with lower RMSE on WP for pressure regression and on AU for wind regression. These results highlight the importance of considering dataset-specific and task-specific characteristics when interpreting benchmark performance. Additionally, we analyzed systematic bias in intensity regression, which leads to underestimation of intensity for strong typhoons. We tested imbalance-aware weighting strategies, including inverse-density weighting inspired by BalancedMSE, and showed that these approaches are not sufficient to remove the bias. The dataset is publicly available at http://agora.ex.nii.ac.jp/digital-typhoon/dataset/ and https://github.com/kitamoto-lab/digital-typhoon/.

Figures & tables

Feature	WP dataset	AU dataset
Temporal coverage	1978-2025	1979-2025
Spatial coverage	100E to 180E in the Northern Hemisphere	90E to 160E in the Southern Hemisphere
Season	January to December	July (the previous year) to June
Best track	Japan Meteorological Agency (JMA)	Bureau of Meteorology (BoM), Australia

Table 1: Characteristics of the WP and AU datasets.

Ch.	WP TCs	WP images	AU TCs	AU images
1	1,169	202,986	511	76,211
2	744	142,583	339	58,865
3	744	142,591	339	58,842
4	502	96,611	223	40,238

Table 2: Number of TCs and Images for the WP and AU datasets for V3-2025.R1.

Ch.	Type	Himawari-1/2/3/4	Himawari-5	GOES-9	Himawari-6/7	Himawari-8/9
1	Infrared	10.5-12.5	10.5-11.5 (11.0)	10.2-11.2 (10.7)	10.3-11.3 (10.8)	$11.2\pm 0.2$
2	Infrared	-	11.5-12.5 (12.0)	11.5-12.4 (11.95)	11.5-12.5 (12.0)	$12.35\pm 0.3$
3	Water Vapor	-	-	6.5-7.0 (6.75)	6.5-7.0 (6.75)	$6.25\pm 0.37$
4	Near Infrared	-	-	-	3.5-4.0 (3.75)	$3.85\pm 0.22$

Table 3: Channel numbers assigned in the dataset and the duration of each channel across the Himawari satellite series.

Figure 1: Overview of the data-processing pipeline.

Ch.	Total images	Train images	Evaluation images (%)
1	202,986	192,956	10,030 ( $4.94\%$ )
2	142,583	132,555	10,028 ( $7.03\%$ )
3	142,591	132,554	10,037 ( $7.04\%$ )
4	96,611	86,551	10,060 ( $10.41\%$ )

Table 4: SS-L2 split (2024–2025) for the WP dataset (V3-2025.R1).

Ch.	Total images	Train images	Evaluation images (%)
1	76,042	70,274	5,768 ( $7.58\%$ )
2	58,865	53,097	5,768 ( $9.80\%$ )
3	58,842	53,074	5,768 ( $9.80\%$ )
4	40,238	34,501	5,737 ( $14.26\%$ )

Table 5: SS-L2 split (2024–2025) for the filtered AU dataset (V3-2025.R1).

Figure 2: Comparison of the convergence behavior of SGD and Adam on WP-1. Curves show the average validation RMSE on the held-out 2024–2025 evaluation set over 5 independent runs, with shaded regions indicating one standard deviation.

Hyper-parameters	Values
Base model	ResNet-18, ResNet-50
Pretrained weights	ImageNet
Image type	Full ( $512\times 512$ ), Resized ( $224\times 224$ ), Cropped ( $224\times 224$ )
Data split	SS-L2
Optimizer	Adam, lr 1e-4 (constant)
Batch size	16

Table 6: Experimental settings for benchmarking single-channel models.

Ch.	Model	Full	Resized	Cropped
1	ResNet-18	6.5618 (±0.06)	6.9591 (±0.09)	6.9054 (±0.05)
1	ResNet-50	7.3367 (±1.13)	7.0446 (±0.22)	8.0222 (±2.53)
2	ResNet-18	6.7844 (±0.09)	7.2420 (±0.06)	7.0621 (±0.09)
2	ResNet-50	8.1232 (±1.29)	8.0211 (±0.43)	7.0262 (±0.26)
3	ResNet-18	6.9385 (±0.05)	7.4771 (±0.15)	7.3650 (±0.10)
3	ResNet-50	7.1517 (±0.24)	7.8674 (±0.21)	7.3509 (±0.19)

Table 7: RMSE (hPa) for the pressure regression task for the WP dataset for four channels and three image types.

Ch.	Model	Full	Resized	Cropped
1	ResNet-18	7.2015 (±0.12)	7.6933 (±0.15)	7.3887 (±0.31)
1	ResNet-50	7.3342 (±0.27)	8.2380 (±0.35)	7.5673 (±0.39)
2	ResNet-18	7.2708 (±0.29)	7.8543 (±0.51)	7.2427 (±0.12)
2	ResNet-50	8.5820 (±0.70)	8.1601 (±0.42)	8.2518 (±0.68)
3	ResNet-18	8.5809 (±1.36)	7.9710 (±0.27)	8.7664 (±1.56)
3	ResNet-50	9.5891 (±1.46)	9.0027 (±0.71)	8.4390 (±0.63)

Table 8: The RMSE (hPa) of the pressure regression task for the AU dataset for four channels and three image types.

Ch.	Model	Full	Resized	Cropped
1	ResNet-18	15.7493 (±1.09)	16.9485 (±0.41)	16.1151 (±0.27)
1	ResNet-50	15.9613 (±0.32)	15.9931 (±0.06)	15.8842 (±0.87)
2	ResNet-18	15.6748 (±0.80)	17.0907 (±0.46)	16.0942 (±0.28)
2	ResNet-50	15.9355 (±0.33)	17.1484 (±1.00)	16.1344 (±0.18)
3	ResNet-18	16.0743 (±0.67)	17.4122 (±0.66)	16.3196 (±0.15)
3	ResNet-50	16.1426 (±0.40)	16.8639 (±0.24)	16.4675 (±0.60)

Table 9: RMSE (kt) for the wind regression task for the WP dataset for four channels and three image types.

Ch.	Model	Full	Resized	Cropped
1	ResNet-18	11.4344 (±0.13)	11.6104 (±0.17)	11.5828 (±0.24)
1	ResNet-50	10.7545 (±0.17)	11.9051 (±0.12)	11.0413 (±0.86)
2	ResNet-18	13.2135 (±0.33)	13.4230 (±0.18)	12.6630 (±0.19)
2	ResNet-50	13.6150 (±0.42)	13.7778 (±0.54)	13.7715 (±0.63)
3	ResNet-18	12.9419 (±0.54)	13.2351 (±0.74)	12.7642 (±0.44)
3	ResNet-50	12.5016 (±0.31)	13.1745 (±0.52)	13.1644 (±0.50)

Table 10: RMSE (kt) for the wind regression task for the AU dataset for four channels and three image types.

Input channels	RMSE (hPa)	Single-channel dataset
Channel 1	6.9270 (±0.12)	6.5618 (±0.06)
Channel 2	6.9605 (±0.07)	6.7844 (±0.09)
Channel 3	7.3958 (±0.58)	6.9385 (±0.05)
Channel 4	7.3973 (±0.17)	7.3386 (±0.08)
Channel 1 + 2	6.9265 (±0.11)	-
Channel 1 + 3	7.1512 (±0.24)	-

Table 11: Results of the pressure regression task for the WP dataset, comparing single-channel and multi-channel inputs. Results from the single-channel dataset with the same configuration are taken from Table 7 .

Input channels	RMSE (hPa)	Single-channel dataset
Channel 1	8.2839 (±0.33)	7.2015 (±0.12)
Channel 2	8.2370 (±0.62)	7.2708 (±0.29)
Channel 3	8.4316 (±0.60)	8.5809 (±1.36)
Channel 4	9.7118 (±0.27)	9.7814 (±0.31)
Channel 1 + 2	8.4443 (±0.10)	-
Channel 1 + 3	8.4714 (±0.42)	-

Table 12: Results of the pressure regression task for the AU dataset, comparing single-channel and multi-channel inputs. Results from the single-channel dataset with the same configuration are taken from Table 8 .

Figure 3: Systematic bias (Prediction - True) across seasonal splits for the pressure regression task (WP-1 dataset, ResNet-18, Channel 1, Full images). Gray curves represent individual splits (each corresponding to a consecutive-year pair from 1978–1979 to 2024–2025), while the blue curve shows the mean bias across all 47 overlapping temporal splits. The shaded area represents ±1 standard deviation.

Panel of Figure 4 — Figure 4: Trade-off between bias reduction for strong typhoons and global RMSE for different weighting schemes for the WP-1 dataset. Results are averaged over five independent training runs using the same fixed training protocol. The red dashed line indicates zero error.

Figure 5: Mean prediction error (Pred - True) for the WP-1 dataset using inverse-density weighting (global RMSE = 8.42). Results are averaged over five independent training runs using the same fixed training protocol. The red dashed line indicates zero error.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Ch.	WP TCs	WP images	AU TCs	AU images
1	1,169	202,986	511	76,188
2	744	142,583	339	58,842
3	744	142,591	339	58,842
4	502	96,611	223	40,238

Appendix

Table 13: Number of TCs and images for the WP and AU datasets for V3-2025.R2. The only differences from Table 2 (V3-2025.R1) are the AU images in Channels 1 and 2, which are 23 fewer because of the removed GOES-9 images (Section 2.5 ); the number of TCs is unchanged.

Dataset	Corrupted	Release	Last 10	Last 5	Epoch 50	Epoch 50
			Evaluation RMSE			Training RMSE
WP-1	577 (0.30%)	R1	6.63 ( $\pm$ 0.01)	6.61 ( $\pm$ 0.05)	6.66 ( $\pm$ 0.08)	2.03
		R2	6.54 ( $\pm$ 0.02)	6.52 ( $\pm$ 0.04)	6.51 ( $\pm$ 0.05)	1.33
WP-3	1,252 (0.94%)	R1	7.10 ( $\pm$ 0.08)	7.09 ( $\pm$ 0.09)	7.07 ( $\pm$ 0.11)	2.89
		R2	6.98 ( $\pm$ 0.05)	7.04 ( $\pm$ 0.10)	7.05 ( $\pm$ 0.23)	1.34
AU-1	39 (0.06%)	R1	7.53 ( $\pm$ 0.19)	7.60 ( $\pm$ 0.23)	7.72 ( $\pm$ 0.37)	1.44

Appendix

Table 14: Comparison of V3-2025.R1 and V3-2025.R2: mean (

\pm

standard deviation) over five runs. “Corrupted” is the number (fraction) of training observations whose image was corrected in V3-2025.R2; in addition, 23 training images were removed from each of AU-1 and AU-2. None of them are in the evaluation set. “Last 10” and “Last 5” are the medians of the evaluation RMSE over epochs 41–50 and 46–50 of each run. All values are in hPa. A large AU-3 R2 value is due to 2 runs that ended on a spike.

Remote Sensing Benchmarks

Remote sensing benchmarks measure how well models interpret Earth-observation data across sensors, regions, time, tasks, and operating conditions.

See all 193 papers

Physics-Informed Diffusion Model for Generating Synthetic Extreme Rare Weather Events Data
Mar 6, 2026Marawan Yakout, Tannistha Maiti, Monira Majhabeen +1Synthetic Data Generation Synthetic Data Augmentation
Department of Computer Science, University of London, London · Deepkapha AI Lab, Assen, The Netherlands
MotifGen: Spatiotemporal interpolation of misaligned satellite images via multi-source generative modeling, in an application to tropical cyclones
Jun 23, 2026Clément Dauvilliers, Claire MonteleoniRemote Sensing Generative Modeling
Inria · INRIA, Paris, Île-de-France, France · University of Colorado Boulder, Boulder, CO, USA.
Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields
Aug 8, 2026Meheru Zannat, Sk. Md. Masudul AhsanRectified Flow Tropical Cyclone Intensity Forecasting
Department of Computer Science and Engineering Khulna University of Engineering & Technology Khulna 9203, Bangladesh

Mar 6, 2026J/K move · Enter open · S save

Physics-Informed Diffusion Model for Generating Synthetic Extreme Rare Weather Events Data

Marawan Yakout, Tannistha Maiti, Monira Majhabeen, Tarry Singh

Department of Computer Science, University of London, London · Deepkapha AI Lab, Assen, The Netherlands

Synthetic Data Generation Synthetic Data Augmentation Physics-Informed Diffusion Models Conditional Diffusion Models

Open paper AskPDF arXiv

Data scarcity is a primary obstacle in developing robust Machine Learning (ML) models for detecting rapidly intensifying tropical cyclones. Traditional data augmentation techniques (rotation, flipping, brightness adjustment) fail to preserve the physical consistency and high-intensity gradients characteristic of rare Category 4-equivalent events, which constitute only 0.14% of our dataset (202 of 140,514 samples). We propose a physics-informed diffusion model based on the Context-UNet architecture to generate synthetic, multi-spectral satellite imagery of extreme weather events. Our model is conditioned on critical atmospheric parameters such as average wind speed, type of Ocean and stage of development (early, mature, late etc) -- the known drivers of rapid intensification. Using a controlled pre-generated noise sampling strategy and mixed-precision training, we generated

16\times16

wind-field samples that are cropped from multi-spectral satellite imagery which preserve realistic spatial autocorrelation and physical consistency. Results demonstrate that our model successfully learns discriminative features across ten distinct context classes, effectively mitigating the data bottleneck. Specifically, we address the extreme class imbalance in our dataset, where Class 4 (Ocean 2, early stage with average wind speed 50kn hurricane) contains only 202 samples compared to 79,768 samples in Class 0. This generative framework provides a scalable solution for augmenting training datasets for operational weather detection algorithms. The average Results yield an average Log-Spectral Distance (LSD) of 4.5dB, demonstrating a scalable framework for enhancing operational weather detection algorithms.

Abstract

Remote Sensing Benchmarks

Explore similar work

Physics-Informed Diffusion Model for Generating Synthetic Extreme Rare Weather Events Data

MotifGen: Spatiotemporal interpolation of misaligned satellite images via multi-source generative modeling, in an application to tropical cyclones

Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields