Building Damage Segmentation After Natural Disasters in Satellite

Imagery with Mathematical Morphology and Convolutional Neural

Networks

Ant

onio dos Santos Ramos Neto

and Daniel Oliveira Dantas

Departamento de Computac¸

ao, Universidade Federal de Sergipe, S

ao Crist

ao, SE, Brazil

Keywords:

Morphological Operations, Registered Images, Damage Level Classiﬁcation, Unet, BDANet, CutMix.

Abstract:

In this study, our main motivation was to develop and optimize an image segmentation model capable of

accurately assessing damage caused by natural disasters, a critical challenge today where the frequency and

intensity of these events are increasing. In order to predict damage categories, including no damage, minor

damage, and major damage, we compared several models and approaches. we explored and compared several

models, focusing on the Unet architecture employing BDANet and other architectures such as ResNet18,

VGG16, and ResNet50. Layers with mathematical morphology operations were applied as a ﬁltering strategy.

The results indicated that the Unet model with the BDANet backbone had the best performance, with an F1-

score of 0.761, which increased to 0.799 after applying mathematical morphology operations.

1 INTRODUCTION

Image segmentation is a fundamental technique in the

ﬁeld of image processing and computer vision that in-

volves dividing an image into meaningful and non-

overlapping regions. The process of image segmenta-

tion is essential for natural scene understanding, as

it allows for the identiﬁcation of objects and their

boundaries within an image (Yu, 2023).

The importance of image segmentation lies in its

ability to extract relevant information from an image,

which can be used for various applications such as ob-

ject recognition, image compression, and image en-

hancement (Li et al., 2020).

One such application of image segmentation is in

assessing damage to buildings and in identifying land-

scape changes before and after natural disasters. This

allows the location and identiﬁcation of buildings and

other structures, which can then be analyzed for dam-

age.

A neural network can be used to locate the most

affected areas and segment damage to buildings (Da

et al., 2022). Similarly, Wang (Wang and Li, 2022)

used automatic image segmentation technology to

identify comprehensive disaster reduction capability

assessment of regional disaster hotspots. However, to

https://orcid.org/0009-0000-3122-8292

https://orcid.org/0000-0002-0142-891X

perform this task of image segmentation after natural

disasters, the availability of datasets that have pre and

post-disaster quality annotations is crucial.

The xBD dataset, is widely used to evaluate

building damage segmentation from satellite im-

agery (Gupta et al., 2019). This dataset includes high-

resolution satellite images from 19 natural disaster

events, such as hurricanes, earthquakes, ﬂoods, for-

est ﬁres, volcanic eruptions, and tsunamis, covering

more than 45,000 square kilometers of various natu-

ral disasters that have happened around the world.

Various neural network architectures have been

used in satellite image segmentation after natural dis-

asters. UNet is a popular architecture that uses a

contraction path to capture context and a symmetric

expansion path to enable accurate localization (Ma

et al., 2020).

Multiscale convolutional neural network with

cross-directional attention (BDAnet) is an architec-

ture that uses multidirectional attention and multi-

scale feature fusion to improve building damage as-

sessment from satellite imagery (Shen et al., 2021).

Like BDAnet, the Siamese hierarchical transformer

framework is another architecture that uses a segmen-

tation network (Da et al., 2022).

The successful application of architectural mod-

els for segmenting pre and post-disaster images de-

pends on using registered images. This presents sub-

828

Neto, A. and Dantas, D.

Building Damage Segmentation After Natural Disasters in Satellite Imagery with Mathematical Morphology and Convolutional Neural Networks.

DOI: 10.5220/0012706300003690

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

In Proceedings of the 26th International Conference on Enterprise Information Systems (ICEIS 2024) - Volume 1, pages 828-836

ISBN: 978-989-758-692-7; ISSN: 2184-4992

stantial challenges, such as the identiﬁcation of sta-

ble features in multitemporal images (Lyu and Jiang,

2017) and the segmentation and clustering of planar

images, both critical issues in the ﬁeld of computer vi-

sion (Burdescu et al., 2014). It is also crucial to main-

tain the spatial relationship between these images for

accurate segmentation (Da et al., 2022).

The studies from Shen (Shen et al., 2021) and

Weber (Weber and Kan, 2020) propose new con-

volutional neural network architectures for building

damage assessment from satellite imagery after nat-

ural disasters. Shen (Shen et al., 2021) proposes

the BDANet, a two-stage CNN that uses multidimen-

sional and multidirectional attention mechanisms to

improve the F1-score of image segmentation. We-

ber (Weber and Kan, 2020) proposes an improved

CNN Inception V3 architecture that combines remote

sensing imagery and block vector data to assess the

degree of damage of groups of buildings.

Both studies aimed to improve the F1-scores of

building damage segmentation models from satellite

imagery using deep learning techniques. Despite the

contributions of the mentioned studies, there is still a

gap regarding the exploration of mathematical mor-

phology and different approaches for registered im-

ages in the evaluation of building damage in post-

disaster satellite images.

Mathematical morphology, with its operations of

dilation, erosion, opening, and closing, allows for the

improvement of the mask predicted by the model,

removing noise and imperfections generated by the

model, thus facilitating the identiﬁcation of damaged

buildings.

In addition, we seek to develop and implement

new approaches for handling registered images while

maintaining the spatial relationship between pre and

post-disaster images. This strategy will allow a more

accurate comparison of the state of buildings before

and after the catastrophic event, contributing to a

more accurate damage assessment.

The goal of this study is to improve the segmen-

tation of building damage from satellite images us-

ing the xBD dataset (Gupta et al., 2019), particularly

in post-disaster situations, by testing different neu-

ral network architectures and by using mathematical

morphology techniques as layers in the neural net-

works and new approaches for registered images.

2 BACKGROUND

2.1 Data Preparation Methods for

Image Segmentation

Data preparation is a crucial step in image segmen-

tation that involves preparing the image conditions to

meet the segmentation requirements.

1. Cropping: select a part of the image, perform a

cropping, and use it as input to the segmentation

model. This technique helps to increase the diver-

sity of the training data and avoid overﬁtting (He

et al., 2022).

2. Data augmentation transformations: applying var-

ious transformations to the input image to create

new training samples. Common transformations

include rotation, scaling, ﬂipping, color jittering,

shear, translation, and color distortions. These

transformations help to increase the size of the

training dataset and improve the robustness of the

model (He et al., 2022).

3. Resize: changing the size of the input image to

match the input size of the segmentation model.

This technique is useful when the input image size

is different from the model’s input size (Alamin

et al., 2016).

4. CutMix: the technique is a data augmentation

strategy proposed by Yun (Yun et al., 2019) that

involves cutting and pasting patches among train-

ing images, where the ground truth labels are also

mixed proportionally to the area of the patches.

2.2 Approaches for Registered Images

There are several possible approaches for segmenting

registered images, such as disaster images where we

have one image pre and one post-disaster.

One possible approach for segmenting registered

images, such as pre and post-disaster images, is to

concatenate the two input images in the channels di-

mension, i.e., stacking them generating an image with

six channels. This approach has been used by Li (Li

et al., 2019) and Muhadi (Muhadi et al., 2020). In this

approach, the concatenated image is fed into a seg-

mentation network to obtain the segmented output.

Another possible approach is to feed the pre and

post-disaster separated images into a Siamese net-

work. This approach has been used by Da (Da

et al., 2022) and Chowdhury (Chowdhury and Rah-

nemoonfar, 2021). In this approach, the Siamese net-

work learns a similarity metric between the pre and

post-disaster images, which can be used to identify

changes in the scene.

Building Damage Segmentation After Natural Disasters in Satellite Imagery with Mathematical Morphology and Convolutional Neural

Networks

829

A third possible approach is to calculate the differ-

ence between the pre and post-disaster images. This

approach has been used by Rudner (Rudner et al.,

2019) and Yang (Yang et al., 2020). In this approach,

the difference image is fed into a segmentation net-

work to obtain the segmented output.

2.3 Neural Network Architectures in

Image Segmentation

UNet is a fully convolutional encoder-decoder net-

work architecture with skip connections between

encoder-decoder modules, which was introduced by

Gaj (Gaj et al., 2019).

Feature pyramid network (FPN), is a neural net-

work architecture that was originally proposed for ob-

ject detection but has also been applied to segmenta-

tion tasks (Minaee et al., 2021).

LinkNet is a neural network architecture that

was proposed for semantic segmentation tasks (Basu,

2023). It is a fully convolutional network that uses a

combination of encoder and decoder modules to ex-

tract features from the input image and generate the

segmentation map.

PSPNet is a neural network architecture that

stands for pyramid scene parsing network. It was pro-

posed for semantic segmentation tasks and is based on

a pyramid pooling module (PPM) that captures mul-

tiscale contextual information (Li, 2023).

BDAnet is a neural network architecture that was

proposed for image segmentation tasks (Mochalov

and Mochalova, 2019). It is a fully convolutional net-

work that uses a combination of dilated convolutions

and atrous spatial pyramid pooling (ASPP) to capture

multiscale contextual information.

2.4 Mathematical Morphology

Mathematical morphology is a ﬁeld of image process-

ing that deals with the analysis and manipulation of

geometric structures in images. The two primitive

operations of mathematical morphology are dilation

and erosion. Additionally, there are the operations

of opening and closing, which are derived from the

primitives (Chen et al., 2021).

Dilation is an operation that expands the bound-

aries of a bright region in an image, while erosion

shrinks them. Opening is an erosion followed by a

dilation, which can be used to remove small bright

objects from an image. Closing is a dilation followed

by an erosion, which can be used to ﬁll small holes in

an image (Banon and Barrera, 1994).

2.5 Visualization and Evaluation

There are several visualization and evaluation tech-

niques that can be used to measure the results of

image segmentation models. For visualization we

use visual semantic segmentation (Benkhoui et al.,

2021) and for evaluation of the results we use the F1-

score (Laine et al., 2021).

3 METHODOLOGY

This section begins with a description of the dataset,

followed by the four steps of the methodology. Data

preparation, models training, application of mathe-

matical morphology ﬁlters, and model evaluation.

The code was developed in Python and is available

in GitHub for future reference

3.1 xBD Dataset from xView2 Challenge

The xBD database from xView2 challenge (Defense

Innovation Unit, DoD, 2023) contains high resolution

satellite imagery of six different types of natural dis-

asters around the world, covering a total area of over

45,000 square kilometers (Gupta et al., 2019).

Each pixel in the image has a value that corre-

sponds to one of the ﬁve labels deﬁned to classify

the state of each building. These labels were deﬁned

according to the degree of damage presented by the

building and are as follows: background, no dam-

age, building with minor damage, building with ma-

jor damage, and destroyed building. Table 1 shows

the values corresponding to each damage level. The

levels range from 1, no damage, to 4, destroyed. The

0 corresponds to the background.

Table 1: Scale of damage noted on buildings. Adapted

from (Defense Innovation Unit, DoD, 2023).

Score Label

Visual description

of the structure

1 No damage

Undisturbed. No sign of water,

structural damage, shingle damage,

or burn masks.

2 Minor damage

Building partially burnt, water surrounding

the structure, volcanic ﬂow nearby,

roof elements missing, or visible cracks.

3 Major damage

Partial wall or roof collapse, encroaching

volcanic ﬂow, or the structure is

surrounded by water or mud.

4 Destroyed

Structure is scorched, completely collapsed,

partially or completely covered with water

or mud, or no longer present.

The xBD dataset was divided into 9,162 training

images, 906 test images and 906 validation images,

https://github.com/ddantas-ufs/2024 building

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

830

for a total of 10,974 images. This division ensures

effective learning, evaluation and validation of the

model, avoiding overﬁtting and improving general-

ization.

3.2 Data Preparation

The data preparation step plays a crucial role in build-

ing robust image segmentation models, allowing the

model to capture the diversity of the input images and

minimize the risk of overﬁtting. The image and mask

are cropped to a speciﬁc size (512×512 pixels). After

cropping, we use three approaches to prepare the data

input:

1. Stacked images: concatenate two RGB input im-

ages into a single one with six channels, three

channels from each input image.

2. Separated images: passing pre and post-disaster

images separately to feed a Siamese network.

3. Difference between images: the difference be-

tween the two input images, forming a single in-

put image with three channels corresponding to

the intensity difference between the images.

3.3 Models Training

In the model training step, four neural net-

work models, available in the Python libraries

segmentation models and Pytorch, were tested:

UNet, FPN, Linknet, and PSPNet. Each of these

models was implemented using seven backbones:

BDANet, VGG16, ResNet18, ResNet50, ResNet34,

ResNeXt50, and SENet154. A total of 28 combina-

tions were trained. All of these models were trained

using pre-trained weights from ImageNet.

For training the models, the Dice loss function

was used, which is a commonly used metric to eval-

uate the overlap between the predicted mask and the

groundtruth mask. We also used the focal loss func-

tion, which aims to solve the class unbalance prob-

lem.

Additionally, a weight was applied to the loss

function to deal with class unbalance. These weights

were set based on the frequency of the classes in the

training images. The optimizer used was AdamW

with a learning rate of 10

−4

and a weight decay of

−6

. The metric used to evaluate the model was the

F1-score.

In each epoch of the training, the model receives

every image of the training set. Transformations are

randomly applied to each image to increase the vari-

ability of the data. Increasing the variability of the

data allows the model to generalize better to new and

unknown data. The following transformations were

applied.

1. Horizontal mirroring: mirrors the image along the

vertical axis.

2. Vertical mirroring: mirrors the image along the

horizontal axis.

3. Rotation: rotates the image by a random angle be-

tween -10 and 10 degrees.

4. Scale: applies a random scale to the image be-

tween 0.8 and 1.2.

5. Additive gaussian noise: adds Gaussian noise to

the image with a random scale between 0 and

0.05.

6. Contrast normalization: adjusts the contrast of the

image to a random value between 0.8 and 1.2.

7. Elastic transformation: applies an elastic defor-

mation to the image.

To increase the amount of examples per class, the

CutMix technique was implemented. This technique

consists of selecting a random rectangular part, with

a probability of 87%, from an image and replacing it

with a corresponding part from another image, with

the respective annotation masks adjusted similarly as

shown in Figure 1.

3.4 Morphological Filters

The application of mathematical morphology opera-

tions was performed in order to improve the F1-score

of the best segmentation model that was obtained in

the training. The use of morphological layers was

composed of tests considering the sizes of the struc-

turing elements 3×3 pixels (SeSize) and the shapes of

the structuring elements squares (SeShape).

We tested four traditional morphological opera-

tions: erosion, dilation, opening and closing. These

layers were inserted immediately after the Unet model

softmax layer with the BDANet backbone. This layer

has an image of dimension 512×512×5, each one of

ﬁve channels in this image predicted represents a seg-

mentation label. The training was carried out with a

set of all the training images and the convolutional

layers frozen.

3.5 Model Evaluation

We used the F1-score metric to evaluate the perfor-

mance of the segmentation models used in the pre and

post-disaster image segmentation problem.

The F1-score is deﬁned as the harmonic mean be-

tween precision and recall and can vary from 0 to 1,

Building Damage Segmentation After Natural Disasters in Satellite Imagery with Mathematical Morphology and Convolutional Neural

Networks

831

(a) (b) (c)

Figure 1: Example of CutMix application, which is represented in the red rectangles, on images. a) image pre-disaster, b)

image post-disaster and c) post-disaster plus mask image.

Table 2: Models results in test dataset (Top 5 F1-score).

Nº

Different approaches

for registered

images

Model Backbone

F1-score

overall

1 Separated images Unet BDANet 0.761

2 Stacked images Linknet ResNet18 0.433

3 Stacked images Linknet VGG16 0.401

4 Stacked images Linknet ResNet50 0.368

5 Difference between images Linknet ResNet50 0.203

with values closer to 1 indicating a model with better

performance. To calculate the F1-score, the segmen-

tation masks produced by the models were compared

with the grountruth segmentation masks.

In addition, to visualize the model results, we plot-

ted the predictions that the model provided to verify

if the model was generating the masks properly.

4 RESULTS

4.1 Experiments

We ran 28 experiments to identify which models have

a better F1-score. Table 2 shows the F1-score of the

ﬁve best models for pre and post-disaster image seg-

mentation. The models were evaluated using the F1-

score with 25 epochs.

The Unet model with the BDANet backbone,

which treated the separated images, performed the

best, with an F1-score of 0.761. This result highlights

the effectiveness of the separated image approach,

suggesting that maintaining image individuality can

retain critical features that may be lost in other ap-

proaches.

Furthermore, the F1-score obtained can also be at-

tributed to the use of the BDANet backbone, since

the model showed high F1-scores compared to other

models tested. The BDANet was proposed by Shen

Figure 2: Train loss and test score by epochs of the best

model.

Figure 3: F1-score per class in each epoch of the best model.

and obtained an F1-score higher than other architec-

tures in the damage segmentation task (Shen et al.,

2021).

After the training step, the model with the high-

est F1-score was chosen to analyze the results and

ﬁnd opportunities to improve the F1-score. Figure 2

shows the train loss and test score over the training

epochs.

The train loss is consistently decreasing over the

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

832

(a)

(b) (c) (d) (e) (f)

Figure 4: Best model prediction on an image example. a) Classes of masks, b) Image pre-disaster, c) Image post-disaster, d)

Mask, e) Mask predicted and f) Mask predicted plus mathematical morphology.

epochs, suggesting that the model is continuously

learning and improving. This is a positive sign that

training is progressing as expected.

The test score was calculated as a weighted av-

erage of these two metrics: 0.3 times the Dice score

plus 0.7 times the F1-score. The test score generally

increases over time. This indicates that the perfor-

mance of the model on the test set is improving.

In addition to analyzing the train loss and test

score. Figure 3 shows the evolution of the F1-score

Building Damage Segmentation After Natural Disasters in Satellite Imagery with Mathematical Morphology and Convolutional Neural

Networks

833

Table 3: Best model results after applying morphological

layers.

Operation F1-score

Increase

F1-score

Dilation

0.799 0.038

Erosion

0.781 0.020

Closing

0.776 0.015

Opening

0.761 0.000

per class in each epoch of the best model.

The overall F1-score had an increasing trend in

the test dataset, although there were some oscilla-

tions. This suggests that the model improved over the

epochs.

We note that the no damage class consistently

has the highest F1-score, suggesting that the model

is more effective at predicting this class which repre-

sents the background.

In contrast, the minor damage and major damage

classes have the lowest F1-scores, suggesting that the

model has more difﬁculty predicting these classes or

that it may have too few examples of these classes.

In addition to observing the trend by class of the

overall F1-score over the epochs, visual results of the

predictions were also evaluated, as shown in Figure 4.

Despite the model’s F1-score having a satisfactory

result, the model predictions had small errors in the

completeness of the polygons. One of these errors can

be clearly seen in the second row of Figure 4, where

we have some predicted rectangles with small holes

or inadequate ﬁlls.

From these observations, an experiment was car-

ried out with the application of layers with mathemat-

ical morphology operations to improve the F1-score.

Table 3 shows the best overall results with mathemat-

ical morphology experiments.

The last column of Table 3 shows the improve-

ment in F1-score obtained with the addition of the

morphological layers. When the morphology layer is

applied after decision layer, there is an increase in the

F1-score, leading to an improvement of 0.038 over the

best model, reaching an F1-score of 0.799.

4.2 Comparison with Other Studies

Table 4, shows how the proposed model in this study

compares to other models from related works.

The proposed model obtained an F1-score of

0.799. This model outperformed BDANet in the no

damage and destroyed classes with an F1-score of

0.954 and 0.879 respectively. However, the proposed

model underperformed in all other classes when com-

pared to BDANet. We may try other techniques to im-

prove the F1-score, such as ensemble methods, Cut-

Mix concentrated on the classes that had lower F1

scores or ﬁne-tuning the hyperparameters with a grid

search.

The FCN (Long et al., 2015; Shen et al., 2021),

MTF (Weber and Kan, 2020), WNet (Hou et al., 2019;

Shen et al., 2021) and Baseline model (Gupta et al.,

2019) models achieved lower F1-scores than the pro-

posed model and BDANet.

5 CONCLUSIONS

In this study, several models were evaluated regard-

ing their F1-score in segmenting pre and post-disaster

images.

The Unet model with the BDANet backbone ob-

tained the best performance, achieving an F1-score of

0.761. This result indicates the effectiveness of the

separated image approach, preserving their individual

features that may be lost in other approaches.

BDANet may have contributed to the high perfor-

mance of this model in extracting relevant features

from images, as evidenced in the study by Shen (Shen

et al., 2021).

The stacked image approach obtained inferior per-

formance, with lower F1-scores than the other tested

models. It was observed that the ResNet18 backbone

architecture obtained a higher F1-score than VGG16

and ResNet50.

The results of these experiments indicate that ap-

plying layers with mathematical morphology opera-

tions can improve the F1-score of the model. The

overall F1-score increased from 0.761 to 0.799.

When comparing the performance of the proposed

model with other studies, it was observed that the

proposed model outperformed the BDANet in the no

damage and destroyed classes. In addition, the pro-

posed model has higher F1-score in all classes com-

pared to FCN, MTF, WNet and the baseline model.

However, the proposed model underperformed in

all other classes when compared to BDANet, suggest-

ing that there is still room for improving the F1-score.

Future works may include strategies to further im-

prove the performance of the proposed model. The

ﬁrst one would be the implementation of an ensem-

ble method. Secondly, we may apply CutMix concen-

trated on the classes that had lower F1-scores, such as

the minor damage and major damage classes. Finally,

ﬁne tuning the hyperparameters with a grid search or

random search could be an additional strategy to im-

prove performance.

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

834

Table 4: Comparison of the F1-score among other studies. Table is in descending order by F1-score.

Models Overall

damage

Minor

damage

Major

damage

Destroyed

BDANet

(Shen et al., 2021)

0.806 0.925 0.616 0.788 0.876

Proposed

model

0.799 0.954 0.601 0.762 0.879

FCN

(Long et al., 2015; Shen et al., 2021)

0.765 0.919 0.532 0.708 0.861

MTF

(Weber and Kan, 2020)

0.741 0.906 0.493 0.722 0.837

WNet

(Hou et al., 2019; Shen et al., 2021)

0.737 0.884 0.518 0.684 0.855

Baseline

model (Gupta et al., 2019)

0.265 0.663 0.143 0.009 0.465

REFERENCES

Alamin, I. M. J., Jeberson, W., and Bajaj, H. (2016). Im-

proved region growing based breast cancer image seg-

mentation. International Journal of Computer Appli-

cations, 975:888.

Banon, G. J. F. and Barrera, J. (1994). Bases da Morfolo-

gia Matem

atica para a an

alise de imagens bin

arias.

UFPE-DI.

Basu, A. (2023). A survey on recent trends in deep learning

for nucleus segmentation from histopathology images.

Evolving Systems, pages 1–46.

Benkhoui, Y., El-Korchi, T., and Reinhold, L. (2021). Ef-

fective pavement crack delineation using a cascaded

dilation module and fully convolutional networks. In

Geometry and Vision: First International Symposium,

ISGV 2021, Auckland, New Zealand, January 28-29,

2021, Revised Selected Papers 1.

Burdescu, D. D., Stanescu, L., Brezovan, M., and Spahiu,

C. S. (2014). Efﬁcient volumetric segmentation

method. In 2014 Federated Conference on Computer

Science and Information Systems, pages 659–668.

Chen, Q., Cheng, Q., Wang, J., Du, M., Zhou, L., and

Liu, Y. (2021). Identiﬁcation and evaluation of ur-

ban construction waste with vhr remote sensing using

multi-feature analysis and a hierarchical segmentation

method. Remote Sensing, 13(1):158.

Chowdhury, T. and Rahnemoonfar, M. (2021). Self atten-

tion based semantic segmentation on a natural disas-

ter dataset. In 2021 IEEE International Conference on

Image Processing (ICIP), pages 2798–2802.

Da, Y., Ji, Z., and Zhou, Y. (2022). Building damage as-

sessment based on siamese hierarchical transformer

framework. Mathematics, 10(11):1898.

Defense Innovation Unit, DoD (2023). xview2: Assess

building damage. https://xview2.org /dataset.

Acessed: 30/04/2023.

Gaj, S., Yang, M., Nakamura, K., and Li, X. (2019). Au-

tomated cartilage and meniscus segmentation of knee

mri with conditional generative adversarial networks.

Magnetic Resonance in Medicine, 84(1):437–449.

Gupta, R., Goodman, B., Patel, N., Hosfelt, R., Sajeev, S.,

Heim, E., Doshi, J., Lucas, K., Choset, H., and Gas-

ton, M. (2019). Creating xBD: A dataset for assessing

building damage from satellite imagery. In Proceed-

ings of the IEEE/CVF conference on computer vision

and pattern recognition workshops, pages 10–17.

He, W., Liu, M., Tang, Y., Liu, Q., and Wang, Y.

(2022). Differentiable automatic data augmentation

by proximal update for medical image segmentation.

IEEE/CAA Journal of Automatica Sinica, 9(7):1315–

1318.

Hou, B., Liu, Q., Wang, H., and Wang, Y. (2019). From

W-Net to CDGAN: Bitemporal change detection via

deep learning techniques. IEEE Transactions on Geo-

science and Remote Sensing, 58(3):1790–1802.

Laine, R. F., Arganda-Carreras, I., Henriques, R., and

Jacquemet, G. (2021). Avoiding a replication cri-

sis in deep-learning-based bioimage analysis. Nature

Chemical Biology, 18(10):1136–1144.

Li, K. (2023). Upernet optimisation and application to

mousehole segmentation. In International Conference

on Artiﬁcial Intelligence, Virtual Reality, and Visual-

ization (AIVRV 2022), volume 12588, pages 157–163.

Li, T.-T., Jiang, C., Bian, Z., Wang, M., and Niu, X. (2020).

Semantic segmentation of urban street scene based

on convolutional neural network. Journal of Physics

Conference Series, 1682(1):012077.

Li, W., Jafari, O., and Rother, C. (2019). Deep object co-

segmentation. In Computer Vision–ACCV 2018: 14th

Asian Conference on Computer Vision, Perth, Aus-

tralia, December 2–6, 2018, Revised Selected Papers,

Part III 14, pages 638–653.

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-

volutional networks for semantic segmentation. In

Building Damage Segmentation After Natural Disasters in Satellite Imagery with Mathematical Morphology and Convolutional Neural

Networks

835

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 3431–3440.

Lyu, C. and Jiang, J. (2017). Remote sensing image regis-

tration with line segments and their intersections. Re-

mote Sensing, 9(5):439.

Ma, H., Liu, Y.-L., Ren, Y., Wang, D.-C., Yu, L., and Abell,

A. D. (2020). Improved CNN classiﬁcation method

for groups of buildings damaged by earthquake, based

on high resolution remote sensing images. Remote

Sensing, 12(2):260.

Minaee, S., Boykov, Y., Porikli, F., Plaza, A., Kehtarnavaz,

N., and Terzopoulos, D. (2021). Image segmenta-

tion using deep learning: A survey. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

44(7):3523–3542.

Mochalov, V. and Mochalova, A. (2019). Extraction of

ionosphere parameters in ionograms using deep learn-

ing. E3s Web of Conferences, 127:01004.

Muhadi, N. A., Abdullah, A. Z., Bejo, S. K., Mahadi, M. R.,

and Mijic, A. (2020). Image segmentation methods

for ﬂood monitoring system. Water, 12(6):1825.

Rudner, T. G. J., Rußwurm, M., Fil, J. E., Pelich, R.,

Bischke, B., Kopackova, V., and Bilinski, P. (2019).

Multi3net: Segmenting ﬂooded buildings via fusion of

multiresolution, multisensor, and multitemporal satel-

lite imagery. Proceedings of the Aaai Conference on

Artiﬁcial Intelligence, 33(1):702–709.

Shen, Y., Zhu, S., Yang, T., Chen, C., Pan, D., Chen, J.,

Xiao, L., and Du, Q. (2021). Bdanet: Multiscale con-

volutional neural network with cross-directional atten-

tion for building damage assessment from satellite im-

ages. IEEE Transactions on Geoscience and Remote

Sensing, 60:1–14.

Wang, J. and Li, M. (2022). Evaluation model of regional

comprehensive disaster reduction capacity under com-

plex environment. Journal of Environmental and Pub-

lic Health, 2022.

Weber, E. and Kan, H. (2020). Building disaster damage as-

sessment in satellite imagery with multi-temporal fu-

sion. arXiv preprint arXiv:2004.05525.

Yang, M.-D., Tseng, H.-H., Hsu, Y., and Tsai, H. J. (2020).

Semantic segmentation using deep learning with veg-

etation indices for rice lodging identiﬁcation in multi-

date uav visible images. Remote Sensing, 12(4):633.

Yu, Y. (2023). Techniques and challenges of image segmen-

tation: A review. Electronics, 12(5):1199.

Yun, S., Han, D., Chun, S., Oh, S. T., Yoo, Y., and Choe,

J. (2019). CutMix: Regularization strategy to train

strong classiﬁers with localizable features. In Pro-

ceedings of the IEEE/CVF international conference

on computer vision.

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

836