This article has been accepted for publication in IEEE Geoscience and Remote Sensing Letters.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/LGRS.2024.3358733

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1

Automated Segmentation of tropical cyclone clouds

in geostationary infrared images
Joshua May, Liang Hu, Elizabeth A. Ritchie, Mehrtash Harandi Senior Member, IEEE, and J. Scott
Tyo, Fellow, IEEE

Abstract—We demonstrate that a convolutional neural network was used to identify tropical cyclone clouds from geostationary
(CNN) based on the U-Net architecture can be used to create infrared brightness temperature images[2], [3]. Motivated by
a cloud mask data set that accurately identifies the clouds earlier work on more general mesoscale convective systems
associated with tropical cyclones (TCs). The CNN can be trained
using a single year of cloud masks produced by an earlier first- [4], we developed a spatiotemporal algorithm that tracked the
principles algorithm, and the results are insensitive to the specific cloud pixels that could be attributed to a particular storm. The
year of training data used. These masks were originally created in net upwelling radiation attributable to TC clouds was then
order to compute the upwelling radiation due to TC clouds, and computed from upwelling radiation data. We first examined
we show that the predicted masks result in both pixel areas and all TCs globally in 2016 by comparing with the clear sky
radiation calculations that are nearly identical to those computed
using the earlier masks. condition[3], and we later extended that work by updating
the labeling algorithm and examining all TCs globally from
Index Terms—Satellite Remote Sensing, Supervised Segmenta- 2000 – 2020 [5]. Our findings were that TC clouds have a net
tion, Tropical Cyclones
cooling effect on average, but that this cooling effect depended
on both location and season of any particular storm.
I. I NTRODUCTION Our previous work has two shortcomings . First, the
cloud pixel mask only considers specific pixels identified
T ROPICAL cyclones (TCs) are mesoscale convective sys-
tems that form between late Spring and early Autumn in
six primary basins around the globe (North Atlantic, Eastern
as cold TC clouds and excludes clear and very low-level
cloud pixels within the tropical cyclone circulation. In addition
to influencing the area covered by TC clouds, a tropical
and Western North Pacific, South Pacific, North Indian, and
cyclone also influences regional weather and even the global
South Indian Ocean basins). Recent research has seen signif-
weather pattern [6], [7]. Thus, the cloud pixel mask likely
icant interest in the relationship between TCs and climate,
misses significant areas whose cloud cover (or lack thereof)
and that research has largely focused on the future frequency
is impacted by TCs. We are developing mask identification
and intensity of TCs in a warming climate. However, there is
methods that will allow better labelling of cyclone vs non-
relatively little research that explores the reverse, namely: How
cyclone areas regardless of cloud temperature. Identifying
does TC activity affect the prevailing climatic conditions?
this region currently requires an expert observer to create a
One mechanism by which TCs impact the climate system is
mask for each image of each storm (comprising hundreds
through their contribution to the Earth’s energy balance (EEB).
of thousands of images in our 20-year data set). The second
The main driving source of energy for the climate system
problem is that our existing algorithm is extremely slow and
is solar radiation. The net radiation flux at the top of the
requires supervision and manual intervention. It is therefore
atmosphere, which is the difference between the downwelling
a daunting task to try to manually create a full 20-year
flux incident from the sun and the upwelling flux reflected
record of TC masks to improve upon our current masks that
from the Earth in the short wave (¡ 3 µm) and emitted in the
only highlight clouds. The goal of the present paper is to
long wave (¿3 µm), is currently negative at about 0.3 PW. This
demonstrate that a stable CNN-based system can be trained
means that the system is trapping energy and hence the current
using a relatively small set of expert-created training data to
warming conditions [1]. If the change in the clouds caused by
reproduce the original 20-year dataset of TC cloud masks
TCs is significant enough to impact the average Earth albedo,
in [5]. We demonstrate that a method based on the U-net
then it can potentially act as a feedback mechanism (positive
convolutional neural network (CNN) architecture[8] can be
or negative) on climate.
trained to create such a dataset with only a single year of
Accurate identification of the regions covered by TCs at
manually labeled training data. While the study presented here
any given time is necessary to study the contribution of TCs
specifically reproduces the 20-year cloud mask dataset we have
to the Earth’s radiation budget. Previously our group presented
reported previously [5], we demonstrate that the masks created
a supervised, first-principles cloud segmentation algorithm that
by the CNN system produce essentially equivalent results for
J. May, M Harandi, L. Hu and J. S. Tyo are with the Electrical and the radiation calculation. This gives us confidence that in the
Computer Systems Engineering Department, and E. Ritchie is with the future we can create TC area masks that include both clear
Civil Engineering and the Earth, Atmospheric, and Environmental Sciences and cloudy pixels using no more than one year’s worth of
Departments, Monash University, Clayton, Australia. This work was supported
by the Australian Research Council under award DP210103203. The authors manually labeled data and then use that training set to train a
acknowledge Himashi Peiris for her assistance in tuning the CNN. CNN system to produce a full 20-year dataset.

Authorized licensed use limited to: Nanjing University of Information Science and Technology. Downloaded on April 01,2024 at 01:29:14 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Geoscience and Remote Sensing Letters. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/LGRS.2024.3358733

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 2

II. DATA S OURCES & R ADIATION C ALCULATION

A. Best Track TC Information
The NOAA International Best Track Archive for Climate
Stewardship (IBTrACS) best-track dataset provides all TC
locations over all basins [9]. IBTrACS is a compilation of
TC best-track archives from all operational weather forecast
centers globally. The time resolution is 3 hours and we
linearly interpolate the data in position to match the temporal Fig. 1. Left: Raw data from GPM MERGIR, centered using IBTrACS
information. Middle: After using mean image-downscaling (ignoring missing
resolution of the satellite data used in this study. Data available values) from 1024x1024 to 512x512. Right: Filling in missing information at
from https://www.ncei.noaa.gov/products the periphery of the TC with information from 30-minutes prior.
/international-best-track-archive.

contained missing data. The segmentation model requires three

B. Infrared Brightness Temperature Images images spaced apart 90 minutes, so data that failed this test
The NCEP/CPC Global Merged IR Brightness Temperature for any of those three images were excluded from the training
Dataset (GPM MERGIR) contains merged geostationary and testing datasets. The training and testing datasets only used
satellite brightness temperature images of the Earth times while the TCs were above 35 knots.
surface from 60◦ S to 60◦ N using all available operational The preprocessing of the GPM MERGIR data can be
geostationary satellites [10]. The half-hourly images have 4 km summarised with the following steps as illustrated in Fig. 1:
spatial resolution and cover all meridians. GPM MERGIR 1) Identify the TC in the IBTrACS database and retrieve 3-
data are preprocessed to adjust resolution and remove hourly latitude, longitude, timestamps, USA wind values
artifacts as described below. Data are available from 2) Identify the first and last time the TC reaches above 35
https://disc.gsfc.nasa.gov/datasets/GPM MERGIR 1/summary. knots, crop the other values.
3) Find the GPM MERGIR files relevant to these times.
C. Upwelling Radiation Dataset 4) Linearly interpolate the 3-h latitude, longitude, times-
The Clouds and Earth’s Radiant Energy System (CERES) tamps and wind values to 0.5 h. (Fig. 2 shows a
SYN1 Top-Of-Atmosphere Dataset fluxes are computed from preprocessed image frame with TC track locations and
the measurements by CERES instruments combined with intensities superimposed.)
broadband fluxes observed by geostationary satellites [11]. The 5) Create an image:
resolution is 1 hour and 1 degree. Radiation calculations are a) Crop the GPM MERGIR images to be 1024x1024
accomplished through a pixel matching algorithm that relates pixels, centered on the interpolated latitude and
the cloud mask at the resolution of the GPM MERGIR dataset longitude (Figure 1A)
to CERES so that CERES radiation is computed proportional b) Affine transform the images so they can be neatly
to the amount of the CERES pixel occupied by TC cloud. The fit into 0-255 values for 8-bit unsigned integers.
details of this calculation are reported in Nguyen, et al. [3]. c) Take the average of 2 × 2 pixels, ignoring missing
Data are available from https://ceres.larc.nasa.gov/data/. pixels. Results in a 512 × 512 image. (Figure 1B)
6) Create a label: Crop the relevant label for this timestamp
D. TC Cloud Masks and TC to 1024 × 1024 and resample to 512 × 512
7) Repeat steps 5–6 for all times.
Cloud masks created using our semi-supervised algo-
8) Where there was missing data in the images and labels,
rithm for the years 2001 - 2020 [3], [5] are used
copy the previous timestamps values. (Figure 1C)
to train the U-Net system. Data are available from
9) Analyse the TC visually, if there are image and label
https://bridges.monash.edu/articles/dataset/tropical
combinations that don’t match, flag them for the future.
cyclone cloud labels in 2001-2020/21436017.
10) Turn the images into a timeseries:
a) Skipping flagged images, examine images 1.5 h
III. M ETHODOLOGY
before and after the image in question.
A. Input Data Preparation b) If neither of the adjacent images are missing 20%
The GPM MERGIR dataset has missing data pixels as of the data in the centre (middle 100 × 100 pixels)
artifacts spread throughout the images. To address these ar- combine it into a timeseries image.
tificats and simultaneously reduce the trainable parameters c) Repeat steps 10a–10b until all images are turned
of the model, the data was downsized from 1024 × 1024 to into 3-image time series.
512 × 512 by averaging out non-missing data pixels in 2 × 2-
pixel blocks resulting in a spatial resolution of 8km for the
model. When larger spatial blocks of data were missing from B. Model Structure and Training
GPM MERGIR dataset, data was copied from 30 minutes A U-Net was chosen as the architecture as it has histor-
prior if data at that pixel existed at that time. Images were ically shown to be successful with semantic segmentation
removed if more than 20% of the middle 100 × 100 pixels [8]. Though it was originally developed for use with medical

Authorized licensed use limited to: Nanjing University of Information Science and Technology. Downloaded on April 01,2024 at 01:29:14 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Geoscience and Remote Sensing Letters. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/LGRS.2024.3358733

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 3

training loss until approximately epoch 12. The testing loss is

essentially the same for all single-year training sets, and the
testing loss using all four years to train is roughly 20% better.
Based on these results, the model chosen for the rest of the
results presented here was trained using 20 epochs. The data
involving 2001 as either the training or testing set has different
behavior. The testing loss is more variable and decreases more
slowly through the first eight epochs, but have comparable
behavior with other years after that. We are uncertain as to
the reasons for this discrepancy for this particular year, but as
the performance of the CNN is comparable for larger numbers
of epochs, so we have not looked into the differences in the
middle epochs.
Fig. 2. Demonstration of using IBTrACS in conjuction with GPM MERGIR.
Showing the track of Typhoon Mitag (2002).

Fig. 3. Architecture of the U-Net.

Fig. 4. The performance of the AI architecture after being trained with the
2001, 2002, 2003, 2004, 2005 and all but 2005, tested on the 2005 dataset.
images, our segmentation task is similar to the location of
objects in a medical image. The U-Net CNN architecture is de- In order to quantify the performance of the model in
signed to provide high-quality segmentation maps as it benefits matching the training cloud masks, we used the intersection
from multiple stages of its encoder during mask generation. over union (IoU) metric [12].
In this paper, we used batch sizes of 12 (limited by GPU area({T } ∩ {P })
RAM), a learning rate of 0.0002, ReLu activation, binary cross IoU = (1)
area({T } ∪ {P })
entropy loss and Adam optimiser. The code used to train and
test the model is available at https://github.com/joshmay47/AI- with {T} representing a ground truth mask and {P} represent-
cloud-segmentation. The model was constructed as a U-Net ing a predicted mask. A histogram of single-frame IoU values
architecture with five encoder and decoder blocks using 3 × 3 was computed and the IoU values for the 15th percentile
convolution kernels and 2 × 2 max pooling at each layer. The is 77.73%, for the 50th percentile is 88.82%, and for the
model takes three images as input: 512 × 512 satellite images 85th percentile is 93.09%. The three images closest to these
for the time in question and the images 90 minutes before and percentiles are shown in Fig. 5 comparing the outline of the
after. The model outputs a 512 × 512 prediction of the cloud training data (“ground truth”) and that predicted by the model.
mask of the TC at the central time. The network architecture The model not only has a high IoU, but it also captures the
is shown schematically in Fig. 3 with the detailed parameters detailed shape of the ground truth with high fidelity. In fact,
of each block shown. The architecture results in 7,775,618 it is our subjective opinion that, in most cases, the model may
trainable parameters. be doing a “better” job of detecting the TC clouds than the
“true” cloud masks from our earlier work[2], [5] that were
IV. R ESULTS used to train it.

The model was trained using all TC data globally for each
year between 2001 – 2005 one year at a time. The models V. C OMPUTATION OF U PWELLING R ADIATION
were tested on all five years. Figure 4 shows the training and The ultimate goal is to assess whether cloud masks created
testing loss using the 2005 data as the testing data as function using the CNN are useful in computing the net upwelling
of epoch on single-year training data as well as using all years radiation from TCs [5]. The upwelling radiation is com-
2001 – 2004 to train. With the exception of 2001, the year-to- puted by matching the cloud mask pixels obtained from
year performance is nearly identical. Both training and testing GPM MERGIR data to upwelling radiation in the CERES
loss decrease rapidly through the fourth epoch, and then the dataset and comparing it with background climatology. The
testing loss continues to decrease but more slowly than the details of this computation are presented in our earlier work

Authorized licensed use limited to: Nanjing University of Information Science and Technology. Downloaded on April 01,2024 at 01:29:14 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Geoscience and Remote Sensing Letters. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/LGRS.2024.3358733

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 4

% Example 1 Example 2 Example 3

15th

50th

85th

Fig. 5. Examples of the AI predictions. Red = ground truth, Blue = prediction

from CNN. These images demonstrate representative performance of the AI
at the 15% percentile IoU, 50% IoU and 85% IoU. Fig. 7. TC pixel number and net radiation from 2130UTC 6th to 2130UTC
10th July 2005 for Dennis over NA. Each line corresponds to the model in
Fig. 6 trained on a particular year of data.

Fig. 8. TC pixel number and net radiation for 44 TCs and their average
diurnal cycle in 2005 computed using the cloud masks predicted by each of
the models in Fig. 6.

predicts cloud size and net radiation for all 44 TCs in 2005
regardless of training set, and the diurnal cycle of net radiation
is almost identical to that computed using the ground truth
cloud masks. Even though at some times there are differences
of as much as 10% in the pixel area of the predicted masks,
Fig. 6. TC labels and predictions for Dennis at 2130UTC on 6th July 2005.
“True” is the mask computed using our earlier algorithm [3]. Images labeled
the impact on the radiation calculation is smaller and does not
2001 – 2004 are masks computed using each single year as the training set change the conclusions of our earlier study that TC clouds
in the U-Net model. “All” is the mask computed using all four years 2001 – are net cooling. The integrated results for the 2005 season are
2004 as training data.
given in Table I compared with the earlier computed values.
The new model has the best result when the entire 2001 - 2004
data set is used to train, but the results are not significantly
[3], [5]. Fig. 6 compares the earlier ground truth cloud different from the models trained on individual years. This
masks for TC Dennis, a storm in the North Atlantic basin gives us high confidence that cloud masks can be computed
at 2130 UTC on July 6, 2005, at the resolution of the CERES from any single year of training data.
dataset. The number of pixels and the net upwelling radiation
computed for each of these masks are plotted in Fig. 7.
Compared with results using our earlier algorithm[3], [5], VI. C ONCLUSIONS
Figs. 6 and 7 indicate that CNN correctly captures the cloud We demonstrate that a U-Net CNN architecture trained
structure, location, net radiation, and time variance of TC only on a single year of data can predict TC cloud masks
Dennis throughout its lifetime. Fig. 8 shows the aggregated that are nearly identical to our earlier supervised method [3].
results for the entire 2005 TC data set using the various Subjective examination indicates that the new masks may
training sets indicated in Fig. 6. The U-Net model accurately better capture the TC clouds than the old ones. These masks

Authorized licensed use limited to: Nanjing University of Information Science and Technology. Downloaded on April 01,2024 at 01:29:14 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
 This article has been accepted for publication in IEEE Geoscience and Remote Sensing Letters. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/LGRS.2024.3358733

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 5

TABLE I [9] K. Knapp, S. Applequist, H. Diamond, J. Kossin, M. Kruk, and

T HE TURE VALUE AND AI PREDICTION OF TC PIXELS AND NET C. Schreck, “NCDC international best track archive for climate stew-
RADIATION IN 2005. ardship (IBTrACS) project, version 3,” Tech. Rep., 2010.
[10] J. Janowiak, B. Joyce, and P. Xie, “NCEP/CPC l3 half hourly 4km global
True 2001 2002 2003 2004 All (60S - 60N) merged IR v1,” Greenbelt, MD, Tech. Rep., 2017.
Pixel number 9.31 9.29 9.64 9.25 9.40 9.30 [11] D. R. Doelling, “CERES level 3 syn1deg-1hour terra-aqua-MODIS
(×107 ) HDF4 file–edition 4a, data set,” Tech. Rep., 2017.
Radiation 28.99 29.13 29.46 29.61 29.59 28.93 [12] H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese,
(TW) “Generalized intersection over union: A metric and a loss for bound-
ing box regression,” in Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), June 2019.

are used to compute the net upwelling radiation due to TC

clouds [5], and for that purpose the masks produce essentially
identical results.
The importance of this result is that we have high confidence
in the ability of our framework to accurately create cloud Joshua May Joshua W. May is a researcher with the Electronic and Computer
Systems Engineering Department at Monash University. In 2023, he received
masks based on a relatively small amount of high-quality his BE (Hons.), specialising in electrical and computer systems engineering,
training data. The main limitation of our current radiation and his BSc, with a major in astrophysics and a minor in mathematics. He
study is that we only compute the radiation due to TC clouds. is currently completing a Masters in Engineering Science (research). His
research interests include machine learning development for remote sensing
However, TCs impact the atmospheric dynamics in clear-sky and computer vision.
regions both within and immediately surrounding the cloudy
regions, and these areas should be taken into account in trying
to quantify the true radiation contribution of TCs to the EEB.
Our current research is aimed at developing a method using
IR Brightness temperature, horizontal winds, vorticity, and Liang Hu Liang Hu received the B.S. and M.Sc. degrees in atmospheric
other meteorological parameters to more accurately capture science from the Nanjing University of Information Science and Technology,
the domain of influence of the TC. These methods, while still Nanjing, China, in 2004 and 2007, respectively, and the Ph.D. degree in
meteorology from the Institute of Atmospheric Physics, Chinese Academy
under development, require and expert observer to hand-label of Sciences, Beijing, China, in 2010. His research interests include mesoscale
cloud masks for individual images. We are already working to convective, tropical cyclone, monsoon, precipitation, and satellite meteorol-
expand the AI method presented here to be able to predict TC ogy. He joined Geoscience Australia as a Severe Wind Hazard Scientist in
2023 after the completion of this work.
area masks from these hand-drawn training data objectively.
Such a method will enable us to more accurately quantify the
true impact of TCs on the upwelling radiation and how it may
affect the EEB in a future climate.
Elizabeth Ritchie is a Professor in the School of Earth, Atmosphere and
R EFERENCES Environment and Department of Civil Engineering at Monash University.
Previously she was a Professor and Associate Dean at UNSW Canberra, and
[1] G. L. Stephens, J. Li, M. Wild, C. A. Clayson, N. Loeb, S. Kato, a faculty member in the Atmospheric Sciences Department of the University
T. L’ecuyer, P. W. Stackhouse Jr, M. Lebsock, and T. Andrews, “An of Arizona. She is a Fellow of the American Meteorological Society (AMS)
update on earth’s energy balance in light of the latest global observa- and has served as editor of the AMS Journals Monthly Weather Review and
tions,” Nature Geoscience, vol. 5, no. 10, pp. 691–696, 2012. Weather and Forecasting, chaired the AMS Scientific and Technological and
[2] K. T. Nguyen, A. S. Alenin, E. A. Ritchie, and J. S. Tyo, “Quantifying Activities Commission tropical meteorology and tropical cyclone committee,
the contribution of tropical cyclones to the earth’s outgoing radiation,” in and served as an elected Councillor of the Society.
IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing
Symposium. IEEE, 2019, pp. 10 107–10 110.
[3] K. T. Nguyen, L. Hu, A. S. Alenin, E. A. Ritchie, and J. S. Tyo,
“A satellite-based remote-sensing framework to quantify the upwelling
radiation due to tropical cyclones,” IEEE Journal of Selected Topics in
Applied Earth Observations and Remote Sensing, vol. 14, pp. 5488–
5500, 2021. Mehrtash Harandi Mehrtash Harandi is an associate professor with the
[4] T. Fiolleau and R. Roca, “An algorithm for the detection and tracking Department of Electrical and Computer Systems Engineering at Monash
of tropical mesoscale convective systems using infrared images from University. He is also a contributing research scientist in the Machine Learning
geostationary satellite,” IEEE transactions on Geoscience and Remote Research Group (MLRG) with Data61/CSIRO. His current research interests
Sensing, vol. 51, no. 7, pp. 4302–4315, 2013. include theoretical and computational methods in machine learning, computer
[5] L. Hu, E. A. Ritchie, and J. S. Tyo, “Quantifying the cooling effect vision, and signal processing.
of tropical cyclone clouds on the climate system,” NPJ Climate and
Atmospheric Science, vol. 6, p. 99, 2023.
[6] I. Tobin, S. Bony, C. E. Holloway, J. Y. Grandpeix, and G. Seze, “Does
convective aggregation need to be represented in cumulus parameteri-
zations?” Journal of Advances in Modeling Earth Systems, vol. 5, pp.
692–703, 2013. J Scott Tyo is the Head of the Electrical and Computer Systems Department
[7] S. Bony, A. Semie, R. J. Kramer, B. Soden, A. M. Tompkins, and at Monash University. Prof. Tyo’s research interests are in areas of applied
K. A. Emanuel, “Observed modulation of the tropical radiation budget optics and electromagnetics, especially as applied to imaging and sensing
by deep convective organization and lower-tropospheric stability,” AGU problems. He is a Fellow of the IEEE, Optica, and SPIE. He served on the
advances, vol. 1, p. e2019AV000155, 2020. AdCom of the IEEE Antennas and Propagation Society from 2010 – 2016
[8] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- and the SPIE Board of Directors from 2020 - 2022.
works for biomedical image segmentation,” in Proc. Int. Conf. Med.
Image Comput. Comut.-Asist. Intervent., 2015, pp. 234 – 241.

Authorized licensed use limited to: Nanjing University of Information Science and Technology. Downloaded on April 01,2024 at 01:29:14 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.