Using Generative Adversarial Networks for Enhanced Augmentation

on Natural Disaster

Shivam Pandey

, Vijay Bhardwaj

, Kavita Thukral

and Mohit Yadav

AIT-CSE, Chandigarh University, Mohali, Punjab, India

Department of Mathematics, Chandigarh University, Mohali, Punjab, India

Keywords: Image Processing, Deep Learning, Generative Adversarial Network, Computer Vision, Artificial Intelligence.

Abstract: Regarding effectively allocating relief as well as resources in tragic circumstances and catastrophic events,

rapid damages recognition and categorization is essential. Numerous studies have been conducted as a result

of the growth of methods for deep learning as well as the accessible nature of pictures on social networking

sites. It has concentrated on assessing damages. Using geographic information from such instances, those

pictures' visual qualities enable immediate assessment of the region's security condition. This study suggests

a system for categorizing disasters; this includes a variety of catastrophic photos that have been synthesized

using generating adversarial particular adjusting of a deep segmentation neuronal network, and adversarial

generative networks using a model. In this research, a structure for categorizing disasters is proposed. It blends

a collection of synthetic, different calamity photos produced by generative adversarial networks, with domain-

dependent adjusting of a deep convolutional neural network -based system. Due to the fact that previous

research in this field has mostly been hampered by a lack of data materials, an example dataset high- lighting

the problem of the unbalanced categorization of several catastrophic events has been created and enhanced.

Investigations, qualitative as well as quantitative information, demonstrate the effectiveness of the

information enhancement technique used to create a data set with equilibrium. Additional tests con- ducted

with different metrics to assess confirmed the suggested framework's precision and generalizability across

multiple categories when compared with additional cutting-edge approaches for the objective of catastrophe

categorization. The structure performed better than the remaining algorithms by an additional 11% annually

in terms of validating reliability.

1 INTRODUCTION

Tremendous detrimental impact on destruction of

property, lives of people, including the planet,

catastrophes and emergencies necessitate rapid

action. To ensure that damages and modifications to

the environment are kept to a least, rescuers and relief

organizations must make strong reaction and

restoration efforts (Dong et al., 2021).

Communication networks have become more

significant in the work of categorizing disasters since

ongoing surveillance of information on various web

channels can result in rapid recognition of risky

circumstances. Utilizing these tools aids the delivery

https://orcid.org/0000-0002-2542-9490

https://orcid.org/0000-0001-6989-1006

https://orcid.org/0000-0001-5988-5083

https://orcid.org/0000-0002-9332-8480

of vital situational rescues as well as urgent

catastrophe relief (Alam et al., 2020).

Communication sites like Facebook and Twitter are

viewed as essential providers of written and visual

material. The effectiveness of multisensory designs

has been improved over starting points in additional

operates, notwithstanding substantial study that

mostly concentrates on written material for obtaining

useful data (Hossain et al., 2022). While prior study

has shown that using pictures is effective, a few

investigations have concentrated exclusively on the

use of vision contents. Following emergency, online

information might be utilized on its own to create

efficient methods for categorizing disasters (Aamir et

664

Pandey, S., Bhardwaj, V., Thukral, K. and Yadav, M.

Using Generative Adversarial Networks for Enhanced Augmentation on Natural Disaster.

DOI: 10.5220/0013583400004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 1, pages 664-670

ISBN: 978-989-758-763-4

al., 2021). Disaster-related postings and pictures are

able to be used to analyses the geographical

connections among mishaps and tragic occurrences

from various perspectives. Emer- gency aid workers

may be sent directly to the scene of the accident if

such knowledge is paired up with the physical details

of what happened. In order to identify catastrophic

incidents in instantaneous fashion, the present

research suggests a method for gathering spatial data

from social media platforms and using classified

images. This strategy can help first-aid rescuers in

catastrophe situations if the outcomes are paired with

location data. Data-level sampled strategies,

including the synthetic minority oversampling

technique, that boost the total amount of classes by

integrating neighboring indications, are commonly

used as remedies versus imbalanced. Instances

loading or changing the elimination functions are

additional algorithm-level strategies to punish

minority class errors (Johnson et al., 2019).

Information-dependent approaches, nevertheless,

may result from duplication and be unfeasible with

data with large dimensions (Ma et al., 2022).

According to this, computational approaches must

choose an appropriate cost or punishment which

constantly changes depending on the assignment at

issue. Artificial information that replicates statistical

features of the data set that was originally collected

can be produced for resolving the issues with

conventional remedies (Singh et al., 2025). The

accuracy of the synthesized information is

significantly influenced by the technique employed to

generate the information. In order to produce

adequate data sets, data enhancement has become an

essential component of neural networks.

The majority of prior strategies rely on data shifting

supplementation techniques, such as conventional

geometrical augmenting and color

modifications(Shorten and Khoshgoftaar, 2019), as

well as ad- vanced strategies like Cut Mix

augmenting, who swaps out updates from sample

pictures against those coming from various groupings

to produce extra data (Yun et al., 2019). Despite these

techniques increase the number of samples, they

merely use crude tweaks that don't result in any

additional significant findings.

2 RELATED WORK

2.1 Catastrophe Categorization

It has been well researched in prior study. But

empirical research on the catergorization of

inaccurate data sets that contain fewer common

calamities, including shortages and storms, is few.

Recent investigations have concentrated around

gathering of catastrophe data using communication

channels to build databases for researchers because

the field of catastrophe categorization lacks labelled

photos. These databases could contain textual data,

photos of disasters, or an assortment of the two. The

CrisisLex information set contains comments relating

to six distinct catastrophic incidents and is one of the

previous papers (Olteanu et al., 2014). In order to help

fire-fighters fight fires in forests, messages paired

with 5,000 photographs are employed to group

together burns having a precision of eighty-six

percent as the categorization re- sults increase

whenever the datasets include pictures of what

happened (Lagerstrom et al., 2016). A large

collection of recorded Tweets postings from

Hurricane Sandy in 2012 has been described by Yang

et al. although the dataset is devoid of any manually

entered comments (Yang et al., 2021). Three

alternative forms of annotating were used by Alam et

al. to obtain CrisisMMD, a sizable heterogeneous

information set, using Twitter throughout several

catastrophic events (Alam et al., 2018). For all kinds

of catastrophe, additional data sets merged and

merged online communication ma- terial with aerial

photographs (Bischke et al., 2017). Additionally, the

researchers re-labeled the already existing

CrisisMMD information in order to identify the kind

of catastrophe, the accuracy of the de- scription

associated with the catastrophe, as well as the extent

of destruction. The resulting dataset was subsequently

evaluated for categorization versus novel models for

deep learning to serve as a starting point to upcoming

research on problems of comparable kind. Pre-trained

neural network algorithms have performed well while

used alongside transfer learning methodologies. A

compact neural network with two separate heads was

developed by Valdez and Godmalin to categories

catastrophe photos and gauge catastrophe severity

(Valdez and Godmalin, 2021). In a study, Hong et al.

used following the catastrophe aerial pic- tures and a

suggested Network that integrates worldwide and

historical context to calculate the extent of harm and

recognize structures destroyed by tremors(Hong et

al., 2022). Utilizing multimedia outputs of text-image

combinations, Liang et al. adjusted already trained

Using Generative Adversarial Networks for Enhanced Augmentation on Natural Disaster

665

neural and linguistic algorithms that obtained

successful results when contrasted with multimedia

categorization metrics (Liang et al., 2022). We

combined the reference sets we already talked about

to create an equitable catastrophe database in this

research, and we generated more cases using a

generative adversarial algorithm for augmenting the

data to boost the amount of data collected for

marginalized categories.

2.2 Generative Enhancement

After the time that Ian Goodfellow and others

introduced a generative adversarial network design in

2014 to create artificial sensory specimens, scientific

literature has put forth a number of variations on the

initial structure. For man- aging numerous categories

in the same system, an adaptive architecture was

presented. By pumping a neural network with class

labels for every data point and employing

unconditional batches normalisation, Shahbazi et al.

investigated the use of prepared generative

adversarial network for transferring information

between categories (Shahbazi et al., 2021). A

Catastrophe generative adversarial net- work was

developed by Rui et al. to produce information set for

the identification of building damages from satellite

photography. Furthermore, by feeding the initial little

dataset and Gaussian white noise through the

generative adversarial network method, which

generated replicated body position samples, a

Generative adversarial network-based system was

utilised to substitute the time-consuming gathering of

information procedure for estimating human posture

(Rui et al., 2021).

3 METHODOLOGY

In order to categories unbalanced catastrophe

information into three primary periods, as shown in

the first figure, this research recommends an

approach. First, in order to produce false examples

with characteristics that are as plausible as

achievable, we built our information augmenting on

the most recent conditioned generative adversarial

network architecture. The examples are created,

assessed utilizing Fréchet Inception Proximity as well

as Inception Scores, and subsequently utilised to add

improved in- formation to the learning set in order to

prevent excessive fitting (Borji, 2022). Forming

judgements using this approach is easier and avoids

the requirement for further actual gathering.

Secondly, utilising both the starting point as well as

the added information set, we improved a group of

prepared VGG16 classifications for performing

disaster prediction. We additionally evaluated how

well the representations performed with various

augmentations. It should be highlighted that the

artificial specimens were solely utilized throughout

the training stage, and that every one of the

experiments were carried out on a single small

number of actual data examples to prevent excess

fitting and guarantee an equitable evaluation of

various approaches. For the purpose of helping in

directing the teaching procedure across all of the

deployed designs, hyper-parameters are individually

set as values. A grid-based look, essentially specifies

the search area as a matrix of hyperparameter

readings and assesses each place in the structure in

order to get the ideal numbers, was used to acquire all

of the variables. As employing images or text-based

information sets, collective classifiers can generate

outcomes that are better and are typically more

resilient that individual algorithms. By merging

projections form many models that were educated on

different portions of the information using

bootstrapping, which utilises stratification

information replication with substitution, we used a

distributed learning methodology. By using

stratification random selection with substitution,

every category is accurately and fairly represented.

Figure 1: The suggested methodology for categorizing

disasters

INCOFT 2025 - International Conference on Futuristic Technology

666

Every framework is prepared to produce predictions

following the instruction phase, in which transferred

learning is used to benefit from the already acquired

parameters. When it comes to making the final

forecast, an overwhelming hard vote is utilised, in

which every classifier individually selects the group

that has the greatest result likelihood. The group that

receives the greatest number of votes determines the

outcome. The generative framework that will be

utilised for enhancement and the sophisticated

convolution classifiers that will undergo training are

described in the parts that follow. Figure 1 depicts the

suggested methodology for categorizing disasters.

3.1 Data Enhancement

For heterogeneous or tabulated information, data

generating strategies utilising over or

underestimating have been successful. Generative

adversarial net- work provides extra advantages for

image production, though. The dependence on an

ample training set is removed by producing excellent

synthesized pictures. Using the building design

depicted in Figure 2, we have established the

suggested structure for a disastrous conditioned

Generative adversarial network. To enable the

envisioned production of the photos that be- long of a

specific class, the model's accuracy is dependent on

the labelling of the classes attached to every frame.

The very first of each of the three convolutional layers

is used by the generator algorithm to transform the

consistent noisy dispersion source into a large feature

vector that represents the newly created picture. To

bring together the data inputs while avoiding each

sample from converging into just one particular,

batch normalization is utilised. Lastly, the system for

discrimination will receive 64*64* 3 photos from

the subject. Comparable convolution stacking is used

to build the tool, which is then finished with

numerous layers of information for categorization.

The last three extremely dense ones receive aver-

aged downward examples of the vectors of features,

which are then gradually mapped to a space with

fewer dimensions for categorization by another Soft-

Max structure. The probable forecast is produced

after the model generates a distribution of odds

spanning the category markers.

Figure 2: The catastrophic Generative adversarial network

construction, which includes the generation and a

discriminator algorithm

3.2 Classifiers with Convolutions

Convolutional neural networks frequently utilize

designs to analyse multivariate vector and produce

extremely precise outcomes. Deep learning is a

prominent method for gathering data. Additionally,

there are several distinct CNN structures, and every

network has a different set of inner components and

processing methods. Since the aforementioned

networks can tackle visual challenges while still

requiring a smaller number of parameters than

conventional systems, we chose to employ the

Inception-V4 design (Szegedy, et al., 2017). Every

design is briefly addressed in the section.

Using Generative Adversarial Networks for Enhanced Augmentation on Natural Disaster

667

4 INCEPTION TECHNIQUE

The initial structure was implemented for training and

evaluation for catastrophe categorization when it first

appeared in 2015 by Szegedy et al. [20]. Despite

being some- what small in terms of dimensions

compared to similarly cutting-edge designs, the

algorithm offers sufficient precision on the ImageNet

information set. GoogleNet com- ponents as well as

Inception V3 are both expanded in Inception V4. The

main objective of the Inception V4 algorithm is to

reduce the number of training variables and there-

fore, computationally complex. The design relies on

the idea of utilising convolutional filtering techniques

of different widths working at the same distance to

construct an overall bigger network opposed to a

more detailed network. The framework is made up of

three separate kinds of inception blocks of

information, each of which has an assortment of

filters in every level, as illustrated in Figure 3,

because the inception design is extremely

configurable.

Figure 3: Inception V4 Architecture with Less Complexity

We examine the initial information in the

following paragraphs before moving on to the

expanded dataset. To confirm the caliber of the

synthesized characteristics, we quantitatively

evaluated the resulting images. 7 separate groups of

catastrophes, with their own unique characteristics,

have been incorporated in the dataset of catastrophes.

Fortunately, two classes (storm and drought photos)

revealed a clear unbalance. Testing is done on

algorithm Inception V4. The information set was

initially matched by learn- ing a generative

adversarial network to generate more examples into

the minority classes in order to create a successful

categorization system. As shown by the various aug-

mentation strategies used in Figure 4, we further

contrasted the efficacy of generative adversarial

network augmentation against conventional and Cut

Mix expansions to confirm its advantages.

Figure 4: The many techniques for augmentation employed

for research includes rotations and twisting of geometry

5 RESULTS

A set of photographs with 24,000 photos spread over

seven categories was used in all of the initial

investigations, and it was unbalanced. The last studies

used the GAN- produced photographs, producing a

collection of 29,000 pictures that was remarkably

matched. the programming language Python, Tensor

Flow Software 2.3.1, and Keras software were used

for implementing each suggested framework. The

standard deviation of the period time to completion of

5000 s was achieved by running every one of the

practice trials on a Jupiter notebook with a local CPU,

however a GPU-based method might significantly

accelerate up computing. Images tumbling, 30% each

direction shifting, and 30% magnification were all

used as part of the geometrical enhancement.

Additionally, the information set had been divided

into 70% learning, 20% confirmation, and 10% tests

in accordance with generally accepted best practices,

with a total number of batches of 8 observations. It

was vital to allot adequate verification examples

because the verification data gave knowledge that

guided the tuning of the algorithm's variables as well

as settings. The evaluation set used the fewest

examples because it evaluated the precision within

the finished classifier. We used several performance

measures, which included accuracy, precision, re-

call, and F1 score, that are computed as follows, to

assess the effectiveness of each catastrophe

categorization structure.

INCOFT 2025 - International Conference on Futuristic Technology

668

Accurac

TP + TN

)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =

𝑇𝑃

𝐹𝑃

)

𝑅𝑒𝑐𝑎𝑙𝑙 =

𝑇𝑃

𝐹𝑁

)

F1 Score =

∗



)

Where FP stands for false favourable, FN for false

unfavourable TP stands for true positive. Evaluation

metrics for the top performing algorithm has been

broken down by classes. Table 1 depicts the

evaluation metrics for the top-performing framework

inception group, broken down by classes.

Table 1: Evaluation metrics for the top-performing

framework inception group, broken down by classes.

Classes Origi

nal

data

preci

sion

Origin

al data

recall

Gan

argumentati

on precision

Gan

argument

ati-on

recall

Land

Slide

0.775 0.77 0.855 0.825

Floo

0.852 0.751 0.84 0.867

Fire 0.76 0.722 0.824 0.840

Structur

0.72 0.669 0.805 0.827

Non-

dama

0.65 0.814 0.838 0.818

Drou

ht 0.114 0.077 0.691 0.764

Hurrica

0.322 0.314 0.770 0.700

Figure 5: Inception- GAN augmentation analysis for

accuracy and epochs.

The goal of the research was to develop a deep

learning platform for catastrophe categorization

which performed better than the latest modern

algorithms. The outcomes demonstrate the viability

of our method for categorizing damaged occurrences.

Figure 5 depicts the inception- GAN augmentation

analysis for accuracy and epochs

6 CONCLUSION

By utilising data gathered from online platforms, we

developed an extensive structure for categorizing

catastrophes in the present article. We addressed the

significant class imbalance in the first information by

training a generative adversarial network to pro- duce

excellent artificial data to strengthen the initial

information. By using a bagging strategy, we carried

out numerous tests and trained combined classifiers

of Inception systems. By contrasting the structure's

outcomes without those obtained from the stand- ard

convolutional neural network designs, we were able

to confirm the effectiveness of the suggested

ensembles method when used in alongside

information enhancement. The concluded

architecture outperformed all previous methods for

solving the identical task by a median of 11%, and it

obtained a precision of 88.5%. This structure may be

used to gather current information across every social

network and carry out geographical categorization

and analysis. Future studies can focus on examining

how the structure can be improved by incorporating

multisensory elements. We think that combining

catastrophe photographs alongside topographical and

written accounts of each major catastrophe could

enhance the categorization outcomes.

REFERENCES

Aamir, M., Ali, T., Irfan, M., Shaf, A., Azam, M. Z.,

Glowacz, A., ... Rahman, S. (2021). Natural disasters

intensity analysis and classification based on

multispectral images using multi-layered deep

convolutional neural network. Sensors, 21(8), 2648.

https://www.mdpi.com/1424-8220/21/8/2648/notes

Alam, F., Ofli, F., Imran, M. (2018, June). Crisismmd:

Multimodal twitter datasets from natural disasters. In

Proceedings of the international AAAI conference on

web and social media (Vol. 12, No. 1).

https://doi.org/10.1609/icwsm.v12i1.14983

Alam, F., Ofli, F., Imran, M., Alam, T., Qazi, U. (2020,

December). Deep learning benchmarks and datasets for

social media image classification for disaster response.

In 2020 IEEE/ACM International Conference on

Using Generative Adversarial Networks for Enhanced Augmentation on Natural Disaster

669

Advances in Social Networks Analysis and Mining

(ASONAM) (pp. 151-158). IEEE.

https://doi.org/10.1109/ASONAM49781.2020.938129

Bischke, B., Helber, P., Schulze, C., Srinivasan, V., Dengel,

A., & Borth, D. (2017, September). The Multimedia

Satellite Task at MediaEval 2017. In MediaEval.

Borji, A. (2022). Pros and cons of GAN evaluation

measures: New developments. Computer Vision and

Image Understanding, 215, 103329.

https://doi.org/10.1016/j.cviu.2021.103329

Dong, Z. S., Meng, L., Christenson, L., Fulton, L. (2021).

Social media information sharing for natural disaster

response. Natural hazards, 107(3), 2077-2104.

https://doi.org/10.1007/s11069-021-04528-9

Hong, Z., Zhong, H., Pan, H., Liu, J., Zhou, R., Zhang, Y.,

... Zhong, C. (2022). Classification of building damage

using a novel convolutional neural network based on

post-disaster aerial images. Sensors, 22(15), 5920.

https://doi.org/10.3390/s22155920

Hossain, E., Hoque, M. M., Hoque, E., Islam, M. S. (2022).

A deep attentive multimodal learning approach for

disaster identification from social media posts. IEEE

Access, 10, 46538-46551.

https://doi.org/10.1109/ACCESS.2022.3170897

Johnson, J. M., Khoshgoftaar, T. M. (2019). Survey on deep

learning with class imbalance. Journal of big data, 6(1),

1-54. https://doi.org/10.1186/s40537-019-0192-5

Lagerstrom, R., Arzhaeva, Y., Szul, P., Obst, O., Power, R.,

Robinson, B., Bednarz, T. (2016). Image classification

to support emergency situation awareness. Frontiers in

Robotics and AI, 3, 54.

https://doi.org/10.3389/frobt.2016.00054

Liang, T., Lin, G., Wan, M., Li, T., Ma, G., Lv, F. (2022).

Expanding large pre-trained unimodal models with

multimodal information injection for image-text

multimodal classification. In Proceedings of the

IEEE/CVF conference on computer vision and pattern

recognition (pp. 15492-15501).

https://doi.org/10.1109/CVPR52688.2022.01505

Ma, F., Li, Y., Ni, S., Huang, S. L., Zhang, L. (2022). Data

augmentation for audio-visual emotion recognition

with an efficient multimodal conditional GAN. Applied

Sciences, 12(1), 527.

https://doi.org/10.3390/app12010527

Olteanu, A., Castillo, C., Diaz, F., Vieweg, S. (2014, May).

Crisislex: A lexicon for collecting and filtering

microblogged communications in crises. In

Proceedings of the international AAAI conference on

web and social media (Vol. 8, No. 1, pp. 376-385).

https://doi.org/10.1609/icwsm.v8i1.14538

Rui, X., Cao, Y., Yuan, X., Kang, Y., Song, W. (2021).

Disastergan: Generative adversarial networks for

remote sensing disaster image generation. Remote

Sensing, 13(21), 4284.

https://doi.org/10.3390/rs13214284

Shahbazi, M., Huang, Z., Paudel, D. P., Chhatkuli, A., &

Van Gool, L. (2021). Efficient conditional gan transfer

with knowledge propagation across classes. In

Proceedings of the IEEE/CVF Conference on

Computer Vision and Pattern Recognition (pp. 12167-

12176).

Shorten, C., Khoshgoftaar, T. M. (2019). A survey on

image data augmentation for deep learning. Journal of

big data, 6(1), 1-48. https://doi.org/10.1186/s40537-

019-0197-0

Singh, K., Yadav, M., Singh, Y., Malik, P. S., Siwach, V.,

Khurana, D., ... Elngar, A. A. (2025). IoT Networks:

Integrated Learning for Privacy-Preserving Machine

Learning. In Artificial Intelligence Using Federated

Learning (pp. 250-275). CRC Press.

http://dx.doi.org/10.1201/9781003482000-13

Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A. (2017,

February). Inception-v4, inception-resnet and the

impact of residual connections on learning. In

Proceedings of the AAAI conference on artificial

intelligence (Vol. 31, No. 1).

https://doi.org/10.1609/aaai.v31i1.11231

Valdez, D. B., & Godmalin, R. A. G. (2021, December). A

deep learning approach of recognizing natural disasters

on images using convolutional neural network and

transfer learning. In Proceedings of the international

conference on artificial intelligence and its

applications (pp. 1-7).

https://doi.org/10.1145/3487923.3487927

Yang, W., Zhang, X., Luo, P. (2021). Transferability of

convolutional neural network models for identifying

damaged buildings due to earthquake. Remote Sensing,

13(3), 504. https://doi.org/10.3390/rs13030504

Yun, S., Han, D., Oh, S. J., Chun, S., Choe, J., & Yoo, Y.

(2019). Cutmix: Regularization strategy to train strong

classifiers with localizable features. In Proceedings of

the IEEE/CVF international conference on computer

vision (pp. 6023-6032).

http://dx.doi.org/10.48550/arXiv.1905.04899

INCOFT 2025 - International Conference on Futuristic Technology

670