2024 IEEE 30th International Conference on Telecommunications (ICT)

Building Damage Detection Using Deep Learning

Models
Kibitok Abraham∗, Moataz Abdelwahab∗, and Mohammed Abo-Zahhad∗†
Email: {abraham.bett, moataz.abdelwahab, mohammed.zahhad}@ejust.edu.eg
∗School of Electronics, Communications & Computer Engineering, Egypt-Japan University of Science and Technology (E-

JUST), New Borg El Arab City, Alexandria 21934, Egypt.

†
Electrical Engineering Department, Faculty of Engineering, Assiut University, Assiut 71515, Egypt.

Abstract— Human life is significantly impacted by ongoing algorithms can adjust and acquire knowledge from current
natural disasters. The global impact of the building damage they data, enhancing their predicting skills as time progresses and
inflict is deeply significant. Deep learning can effectively analyze environmental variables fluctuate. However, the dataset for
the extent of damage to buildings. However, the challenge is the training the deep learning models must be present and
availability of the dataset. An innovative, challenging dataset
adaptive to the scenarios in which they happen [4], [7]. The
collected from satellites has been presented, encompassing the
extent of structural harm caused by the earthquake in Morocco. development of the dataset is a cumbersome experience. The
The dataset is from a complex, practical, arid environment lack of a well-curated dataset is a significant challenge
where the building colour is almost like the background. The researchers face in developing deep learning models.
2024 IEEE 30th International Conference on Telecommunications (ICT) | 979-8-3503-5669-4/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICT62760.2024.10606139

dataset has been tested using U-Net, FCN, and ResUnet models.
The ResUnet demonstrates superior performance metrics with
Analyzing the impact of natural disasters using deep
pixel accuracy, recall, precision, F1-score, kappa score, and IoU learning is gaining traction [12], [13], [14]. The analysis can
values of 96.2%, 86.4%, 86%, 86.2%, 75.7%, and 85.5%, include financial loss, infrastructural damage, and the extent
respectively, which promotes it for practical application. of the loss. Buildings damaged by natural disasters affect
livelihood and societal norms. The prompt quantification of
Keywords— Natural Disaster, Building Damage, CNN, Deep losses can help mitigate their impacts . Earthquakes continue
Learning, Dataset to cause extensive damage to buildings [15], [16], [17].
However, the challenge in analyzing the building damage is
I. INTRODUCTION
the availability of a well-annotated dataset. The already
Natural disasters continue to present significant challenges available dataset has the challenge of covering all the areas,
to the world. They impact human lives, infrastructure, especially the arid regions. Moreover, the building blocks in
communities, and the environment around them. With the the arid regions are almost like the environmental
ongoing expansion of metropolitan areas into susceptible background, making it very difficult to determine the extent
regions, the frequency and intensity of natural disasters are of damage in their analysis.
projected to continue, worsening their impact on society and
the economy [1], [2], [3]. This paper introduces a new annotated dataset for
detecting building damage, especially for arid regions. The
Natural disaster prediction and mitigation have depended dataset has been collected from satellite imagery after the
on empirical models and analysis of historical data [4]. Morocco earthquake in September 2023. Human experts
Although these methods have offered significant have curated the dataset to ensure that it is well annotated. A
perspectives, they frequently face challenges in grasping comparative analysis of the dataset performance using
natural systems’ intricate dynamics and nonlinear structures. various deep-learning models has been done.
Machine learning algorithms have revolutionized the analysis
of natural disasters. However, machine learning models do The paper is divided into six sections. Section two presents
not extract features well when faced with deep and large the existing related works, with part three introducing the
complex datasets. Deep learning algorithms have proposed collected dataset. The models used are analyzed in
significantly transformed multiple fields by providing section four. Section five highlights the performance metrics.
advanced abilities in data-driven modeling, pattern The experimental results and their analysis are done in
identification, and decision-making [5]. The development of section six. Section seven highlights the conclusion and
computer vision has dramatically increased the availability of future works.
data that deep learning algorithms need to process. Deep II. RELATED WORKS
learning algorithms utilize extensive datasets and computing
capabilities to reveal complex patterns and connections in Analyzing natural disaster damage on buildings using deep
data, leading to more precise predictions and proactive learning provides an avenue for better understanding and
actions. analysis. In early attempts, machine learning algorithms like
the K-nearest neighbour, decision trees, and random forests
Deep learning shows potential for improving forecast have been proposed to classify earthquake damage in
accuracy, early warning systems, and disaster response buildings [18]. The data used in the research were collected
tactics in natural disasters [5], [6], [7], [8], [9], [10], [11]. from the 2014 South Napa earthquake, which entailed
Deep learning algorithms can extract valuable insights and spectral acceleration and building features like floor area,
detect early signs of upcoming disasters by examining many plan irregularities, and age. They achieved an accuracy of
data sources such as satellite imagery, sensor data, social 66%.
media feeds, and historical records. Furthermore, these

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

Authorized licensed use limited to: Makerere University Library. Downloaded on July 31,2024 at 21:25:10 UTC from IEEE Xplore. Restrictions apply.
979-8-3503-5669-4/24/$31.00 ©2024 IEEE
 2024 IEEE 30th International Conference on Telecommunications (ICT)

Semantic segmentation has been applied in building IV. MODELS

damage analysis as it can consider spatial relationships The deep learning models used in this study include the U-
between objects in an image[19]. A multi-task framework to Net [20], the ResUnet [26], and the Fully Connected Network
detect the damaged and those of good standing by (FCN) [27] network. The U-Net model is the most commonly
segmentation has been proposed in [15]. The proposed used deep learning model for segmentation. We apply the
network has been trained using the dataset collected from the ResUnet to leverage the residual backbones and FCN for its
2010 Haiti earthquake. U-Net [20] has been presented to lightweight nature. U-Net was introduced to solve the issue
determine the building damage from the earthquake in Haiti of limited annotated data, especially in the medical field. It
using VHR satellite images [21]. They achieve an overall consists of two paths, contracted and expansive, in a U-
accuracy of 68.71%. shaped format, as shown in Figure 1. The contracting path
Several networks have been used to categorize damage in extracts features on the image through encoding layers, while
buildings. In Zhao et al. [22] an Efficient Channel Attention the expansive paths generate the segmentation maps. The
and Depthwise Separable Convolutional Neural Network encoder layer reduces the spatial resolution and increases the
(ECADS-CNN) network is used to assess building damage in depths through the convolution. The decoder layers upsample
two stages using the xBD dataset [23], achieving 82.14% F1- the data with the help of skip connections, which pass spatial
score. The xBD and Haiti earthquake datasets have been information from the contracting path to create the feature
analyzed using the Squeeze and Excitation high-resolution maps.
model to achieve an F1-score of 82.71% [24]. ResNet-34,
Squeezeand excitation network (SENet), and Dual Path
Network (DPN) backbones, in conjunction with attention-
based U-Net, have been evaluated using the xBD datasets by
[25]. However, the challenge with the xBD dataset is the
availability of images from the exclusive earthquake cases.
The Haiti earthquake dataset cannot be applied in an arid
region environment. This leads to a need for a dataset that can
address the challenge within the regions.
III. DATASET Figure 1 U-Net architecture [20].

A. Data Collection The ResUnet model combines the advantages of U-Net

and residual network algorithms to develop a segmentation
The Morocco Earthquake happened in 2023 near the model [26]. The models incorporate residual blocks to help
Marrakech region. We collected the images from the Maxar solve the gradient vanishing problem, enabling better
Open data platform to develop our dataset. The portal has the
information exchange across the network. The residual block,
pre- and post-disaster images. We obtained 1000 large
GeoTIFF images of the post-disaster category. with skip connections, is added to the contracting path, as
shown in Figure 2. The encoder uses a 3x3 convolution layer
B. Data Preprocessing with a stride of 2 instead of a pooling layer. The bridge
The dataset downloaded comes in the form of GeoTIFF applies a preactivated residual block to connect the decoder.
files. The files are stored in folders that correspond to the The decoder path up samples with a 2x2 layer to double the
Catalog ID. GeoTIFF files come in sizes of 17408 by 17408 spatial dimensions of the feature map.
pixels. This size is too big for processing. Each image has
been split into smaller images “1024 x 1024” with PNG
format. This resulted in 189,000 images. Some resulting
images can be blank; hence, we remove them by filtering.
This reduces the number of images to 34,000. Images with
buildings are then selected, and this reduces our dataset to
34,000. We select the buildings with damaged buildings,
resulting in 1,100 images. The images are curated and then
annotated to obtain the labels.
Figure 2 ResUnet architecture [30].
C. Data Annotations The FCN model is a modification of the classification
Supervised deep learning models require labels for their model for segmentation purposes. FCN uses locally
learning process. The labels help the models make the correct connected convolution layers, pooling, and up-sampling
decisions. To create the labels, we label the pixels into three while removing the dense layer. The up-sampling is
categories: background, damaged, and no damaged buildings. processed via deconvolution using fractional stride. Since the
The dataset labelled comes to 1100 images with their model does not have dense layers, the network is lightweight
respective masks. The dataset is then subdivided into train, and has few parameters.
test, and validation at a ratio of 80%, 10%, and 10%, The use of vision transformers in the analysis of natural
respectively. disasters has gained traction due to their ability to extract long
features. In [28], incorporation of multi-stage Transformer
blocks in the encoder is applied in the development of

Authorized licensed use limited to: Makerere University Library. Downloaded on July 31,2024 at 21:25:10 UTC from IEEE Xplore. Restrictions apply.
 2024 IEEE 30th International Conference on Telecommunications (ICT)

lightweight model for medical imaging. The model, LeViT- on a core i7, Graphical processing unit RTX 5000, and
UNet, combines the U-Net features and vision tranformers. 132GB RAM. The maximum number of epochs is capped at
Determing the boundaries of the image features is an 200, with model checkpoint and early stopping implemented.
important persperctive in image analysis. In [29], a model for We analyze the model’s performance using 512 x 512 and
skin lesion identification is enhanced by utilizing channel 1024 by 1024 image sizes. A comparative analysis using the
reverse attention and gated convolutions iteratively to merge xBD dataset was also done to evaluate the parameters of our
the features from the current layer with the prediction proposed dataset obtained compared to the commonly used
outcomes of the neighboring next layer. dataset. Table I shows the performance metrics of the various
models.
V. PERFORMANCE METRICS Two experiments were conducted using 512 by 512 and
1024 by 1024 dataset image size. In experiment I, the U-Net
The performance of deep learning models helps determine
model performs better than the other models when using a
the appropriate one for the application. The most common
512 by 512 image size. It achieves 94.2%, 53.8%, 73.2%,
metrics used are the pixel classification accuracy (Acc),
62.2%, 60.8%, and 45.1% for pixel accuracy, recall,
precision (Pre), recall (Re), F1-score (F1), Kappa score (k),
precision, F1-score, kappa score, and IoU, respectively. The
and Intersection over Union (IoU). The calculations of these
ResUnet achieves comparatively good results compared to
metrics are based on the model’s ability to determine the true
the FCN. The FCN model has a better recall metric than all
positive (TP), true negative (TN), false positive (FP), and
other metrics. This shows a better ability to identify true
false negative (FN) values. The metrics can be defined as per
positives. However, its precision is too low, resulting in a
the following equations. The model’s performance can also
lower kappa score and F1 score.
be measured in terms of computational complexity when
deployed in embedded devices. Testing time also influences In experiment II, all models have improved performance
the choice of the model. when evaluating our dataset with 1024 by 1024 image size.
In FCN, the kappa score, F1-score, precision, and IoU are
= (1) greatly improved to 48.3%, 50.3%, 55.2%, and 33.6% from
6.8%, 12.8%, 7%, and 6.8%, respectively. The ResUnet
= (2) model achieves the best results comparatively to other models
with 96.2%, 86.4%, 86%, 86.2%, 75.7%, and 85.5% for pixel
= (3) accuracy, recall, precision, F1-score, kappa score, and IoU,
( )
respectively. The ResUnet model achieves comparatively
1 = (4) better performance metrics due to residual networks’ ability
to go deeper without the vanishing gradients problem. This
= (5) improved performance is due to the ample spatial space for
the models to extract the features. However, this increases the
= (6) inference time.
As shown in Table I, the proposed deep learning model has
VI. EXPERIMENTAL RESULT AND ANALYSIS
achieved excellent results. These consistent results promote
The proposed U-Net, FCN, ResUnet, LeViT-UNet [28], it for challenging cases, as in our new dataset. The models
GFANet [29] models were implemented and tested on our can understand and extract features from our dataset. Fig. 2
collected dataset in its challenging environment. To achieve below shows the segmentation map of the three models used
this, we carry out the experiments in a PyTorch environment in the study.

TABLE I. COMPARATIVE PERFORMANCE METRICS OF THE DATASET

Models Dataset Accuracy IoU k F1-score Recall Precision Test Time (sec)
FCN xBD 78.2% 2.9% 3.7% 5.7% 58.8% 3% 0.4
Proposed 512x512 57% 6.8% 7.1% 12.8% 77.8% 7% 0.22
Proposed1024 x1024 92.2% 48.3% 33.6% 50.3% 46.1% 52.2% 0.35
ResUnet xBD 99.2% 59.6% 74.5% 74.7% 71.5% 78.2% 0.57
Proposed 512x512 93.6% 45.7% 61.4% 62.8% 57.1% 69.8% 0.25
Proposed 1024 x1024 96.2% 85.5% 75.7% 86.2% 86.4% 86% 0.394
U-Net xBD 99.2% 55.9% 71.5% 71.7% 63.5% 80.4% 0.465
Proposed 512x512 94.2% 45.1% 60.8% 62.2% 53.8% 73.2% 0.29
Proposed 1024 x1024 95% 59.2% 73.3% 74.5% 67.3% 83.3% 0.435
LeViT-UNet xBD 99% 46.1% 62.8% 63.1% 54.9% 74.3% 0.35
Proposed 512x512 92.9% 38.1% 39.3% 49.9% 39.3% 68.3% 0.13
Proposed 1024 x1024 94% 44.8% 60.6% 61.8% 48.6% 85% 0.28
GFANet xBD 99% 41.5% 58.3% 44.7% 85.1% 44.7% 0.47
Proposed 512x512 76.1% 27.3% 39.2% 39.5% 65.9% 28.2% 0.25
Proposed 1024 x1024 79.1% 42.3% 44.7% 44.9% 75.3% 32% 0.36

Authorized licensed use limited to: Makerere University Library. Downloaded on July 31,2024 at 21:25:10 UTC from IEEE Xplore. Restrictions apply.
 2024 IEEE 30th International Conference on Telecommunications (ICT)

VII. CONCLUSION kappa score, and IoU, respectively. The developed dataset
Natural disasters continue to affect human life extensively. can detect damaged buildings in a complex environment,
The building damage they cause is profoundly felt across the especially in arid regions where the buildings are almost like
world. A novel dataset has been introduced that covers the environment.
building damage from the Morocco earthquake. The dataset ACKNOWLEDGMENT
is of two sizes: 1024 by 1024 and 512 by 512. Increasing the
image size of the dataset improves the performance metrics. The support from the Mission Department of Egypt’s
U-Net, FCN, and ResUnet models have comparatively Ministry of Higher Education, the Japanese International
analyzed the dataset. The ResUnet achieves better Cooperation Agency (JICA), and the Science, Technology,
performance metrics of 96.2%, 86.4%, 86%, 86.2%, 75.7%, and Innovation Funding Authority (STDF) under Project
and 85.5% for pixel accuracy, recall, precision, F1-score, #38285 is greatly appreciated.

Figure 3 Sample images building damage detection using different deep learning models. The deep blue colour indicates the background,
the red colour represents the undamaged buildings, and the green colour the damaged buildings.

social media,” Software Impacts, vol. 13, p. 100319,

REFERENCES Aug. 2022, doi: 10.1016/j.simpa.2022.100319.
[1] H. J. Caldera and S. C. Wirasinghe, “A universal [6] M. Aamir et al., “Natural Disasters Intensity Analysis
severity classification for natural disasters,” Natural and Classification Based on Multispectral Images
Hazards, vol. 111, no. 2, pp. 1533–1573, Mar. 2022, Using Multi-Layered Deep Convolutional Neural
doi: 10.1007/s11069-021-05106-9. Network,” Sensors, vol. 21, no. 8, p. 2648, Apr. 2021,
[2] J. P. Amezquita-Sanchez, M. Valtierra-Rodriguez, doi: 10.3390/s21082648.
and H. Adeli, “Current Efforts for Prediction and [7] Z. Zheng, Y. Zhong, J. Wang, A. Ma, and L. Zhang,
Assessment of Natural Disasters: Earthquakes, “Building damage assessment for rapid disaster
Tsunamis, Volcanic eruptions, Hurricanes, response with a deep object-based semantic change
Tornados, and Floods,” Scientia Iranica, vol. 0, no. detection framework: From natural disasters to man-
0, pp. 0–0, Nov. 2017, doi: 10.24200/sci.2017.4589. made disasters,” Remote Sens Environ, vol. 265, p.
[3] R. Pleninger, “Impact of natural disasters on the 112636, Nov. 2021, doi: 10.1016/j.rse.2021.112636.
income distribution,” World Dev, vol. 157, p. [8] M. S. Abdalzaher, M. Krichen, D. Yiltas-Kaplan, I.
105936, Sep. 2022, doi: Ben Dhaou, and W. Y. H. Adoni, “Early Detection of
10.1016/j.worlddev.2022.105936. Earthquakes Using IoT and Cloud Infrastructure: A
[4] K. Abraham, M. Abdelwahab, and M. Abo-Zahhad, Survey,” Sustainability, vol. 15, no. 15, p. 11713, Jul.
“Classification and detection of natural disasters 2023, doi: 10.3390/su151511713.
using machine learning and deep learning techniques: [9] O. M. Saad, A. G. Hafez, and M. S. Soliman, “Deep
A review,” Earth Sci Inform, Dec. 2023, doi: Learning Approach for Earthquake Parameters
10.1007/s12145-023-01205-2. Classification in Earthquake Early Warning System,”
[5] F. K. Sufi, “AI-SocialDisaster: An AI-based software IEEE Geoscience and Remote Sensing Letters, vol.
for identifying and analyzing natural disasters from

Authorized licensed use limited to: Makerere University Library. Downloaded on July 31,2024 at 21:25:10 UTC from IEEE Xplore. Restrictions apply.
 2024 IEEE 30th International Conference on Telecommunications (ICT)

18, no. 7, pp. 1293–1297, Jul. 2021, doi: [20] O. Ronneberger, P. Fischer, and T. Brox, “U-Net:
10.1109/LGRS.2020.2998580. Convolutional Networks for Biomedical Image
[10] F. Lamas, K. Duguet, J. E. Pezoa, G. A. Montalva, S. Segmentation,” May 2015, doi:
N. Torres, and W. Meng, “Crowd detection and 10.48550/arXiv.1505.04597.
estimation for an earthquake early warning system [21] M. Moradi and R. Shah-Hosseini, “Earthquake
using deep learning,” in Pattern Recognition and Damage Assessment Based on Deep Learning
Tracking XXXIII, M. S. Alam and V. K. Asari, Eds., Method Using VHR Images,” in IECG 2020, Basel
SPIE, May 2022, p. 32. doi: 10.1117/12.2622392. Switzerland: MDPI, Nov. 2020, p. 16. doi:
[11] K. Abraham, M. Abdelwahab, and M. Abo-Zahhad, 10.3390/IECG2020-08545.
“Image Classification of Natural Disasters Using [22] Z. Zhao, F. Wang, S. Chen, H. Wang, and G. Cheng,
Different Deep Learning Models,” in 2022 10th “Deep object segmentation and classification
International Japan-Africa Conference on networks for building damage detection using the
Electronics, Communications, and Computations xBD dataset,” Int J Digit Earth, vol. 17, no. 1, Dec.
(JAC-ECC), Alexandria, Egypt: IEEE, Dec. 2022, 2024, doi: 10.1080/17538947.2024.2302577.
pp. 191–196. doi: https://doi.org/10.1109/JAC- [23] D. Lam et al., “xView: Objects in Context in
ECC56395.2022.10043965. Overhead Imagery,” Feb. 2018, doi:
[12] L. Yang and G. Cervone, “Analysis of remote 10.48550/arXiv.1802.07856.
sensing imagery for disaster assessment using deep [24] C. Liu, S. M. E. Sepasgozar, Q. Zhang, and L. Ge, “A
learning: a case study of flooding event,” Soft novel attention-based deep learning method for post-
comput, vol. 23, no. 24, pp. 13393–13408, Dec. 2019, disaster building damage classification,” Expert Syst
doi: 10.1007/s00500-019-03878-8. Appl, vol. 202, p. 117268, Sep. 2022, doi:
[13] S. Jones and J. Saniie, “Using Deep Learning and 10.1016/j.eswa.2022.117268.
Satellite Imagery to Assess the Damage to Civil [25] C. Wu et al., “Building Damage Detection Using U-
Structures After Natural Disasters,” in 2019 IEEE Net with Attention Mechanism from Pre- and Post-
International Conference on Electro Information Disaster Remote Sensing Datasets,” Remote Sens
Technology (EIT), Brookings, South Dakota, USA: (Basel), vol. 13, no. 5, p. 905, Feb. 2021, doi:
IEEE, May 2019, pp. 189–193. doi: 10.3390/rs13050905.
10.1109/EIT.2019.8833724. [26] F. I. Diakogiannis, F. Waldner, P. Caccetta, and C.
[14] J. Sublime and E. Kalinicheva, “Automatic Post- Wu, “ResUNet-a: a deep learning framework for
Disaster Damage Mapping Using Deep-Learning semantic segmentation of remotely sensed data,”
Techniques for Change Detection: Case Study of the Apr. 2019, doi: 10.1016/j.isprsjprs.2020.01.013.
Tohoku Tsunami,” Remote Sens (Basel), vol. 11, no. [27] J. Long, E. Shelhamer, and T. Darrell, “Fully
9, p. 1123, May 2019, doi: 10.3390/rs11091123. Convolutional Networks for Semantic
[15] F. Chen and B. Yu, “Earthquake-Induced Building Segmentation,” Nov. 2014, doi:
Damage Mapping Based on Multi-Task Deep 10.48550/arXiv.1411.4038.
Learning Framework,” IEEE Access, vol. 7, pp. [28] G. Xu, X. Zhang, X. He, and X. Wu, “LeViT-UNet:
181396–181404, 2019, doi: Make Faster Encoders with Transformer for Medical
10.1109/ACCESS.2019.2958983. Image Segmentation,” in Pattern Recognition and
[16] S. Mangalathu and H. V. Burton, “Deep learning- Computer Vision. PRCV 2023. Lecture Notes in
based classification of earthquake-impacted Computer Science, Vol 14432., Q. Liu, H. Wang, Z.
buildings using textual damage descriptions,” Ma, W. Zheng, H. Zha, X. Chen, L. Wang, and R. Ji,
International Journal of Disaster Risk Reduction, Eds., Xiamen, China: Springer, Singapore, 2024, pp.
vol. 36, p. 101111, May 2019, doi: 42–53. doi: 10.1007/978-981-99-8543-2_4.
10.1016/j.ijdrr.2019.101111. [29] S. Qiu, C. Li, Y. Feng, S. Zuo, H. Liang, and A. Xu,
[17] K. Abraham, M. Abo-Zahhad, and M. Abdelwahab, “GFANet: Gated Fusion Attention Network for skin
“Natural disaster damage analysis using lightweight lesion segmentation,” Comput Biol Med, vol. 155, p.
spatial feature aggregated deep learning model,” 106462, Mar. 2023, doi:
Earth Sci Inform, May 2024, doi: 10.1007/s12145- 10.1016/j.compbiomed.2022.106462.
024-01325-3. [30] L. Xiang, Y. Li, W. Lin, Q. Wang, and D. Shen,
[18] S. Mangalathu, H. Sun, C. C. Nweke, Z. Yi, and H. “Unpaired Deep Cross-Modality Synthesis with Fast
V. Burton, “Classifying earthquake damage to Training,” in Deep Learning in Medical Image
buildings using machine learning,” Earthquake Analysis and Multimodal Learning for Clinical
Spectra, vol. 36, no. 1, pp. 183–208, Feb. 2020, doi: Decision Support. DLMIA ML-CDS 2018 2018.
10.1177/8755293019878137. Lecture Notes in Computer Science(), S. Danail, T.
[19] F. Liu, L. Jiao, B. Hou, and S. Yang, “POL-SAR Zeike, C. Gustavo, S.-M. Tanveer, M. Anne, M.-H.
Image Classification Based on Wishart DBN and Lena, M. R. S. T. João, B. Vasileios, C. N. Jacinto,
Local Spatial Information,” IEEE Transactions on L. Zhi, C. Sailesh, M. Mehdi, G. Hayit, and M.
Geoscience and Remote Sensing, vol. 54, no. 6, pp. Anant, Eds., Granada, Spain: Springer, Cham, Sep.
3292–3308, Jun. 2016, doi: 2018, pp. 155–164. doi: https://doi.org/10.1007/978-
10.1109/TGRS.2016.2514504. 3-030-00889-5_18.

Authorized licensed use limited to: Makerere University Library. Downloaded on July 31,2024 at 21:25:10 UTC from IEEE Xplore. Restrictions apply.