Deep-Learning Based Super-Resolution of Aeolianite Images on the

Purpose of Edge Detection and Pattern Extraction

Antigoni Panagiotopoulou

, Lemonia Ragia

and Niki Evelpidou

Department of Surveying and Geoinformatics Engineering, University of West Attica, Egaleo Campus,

Egaleo 12244, Greece

School of Applied Arts and Sustainable Design, Hellenic Open University, Patras 26335, Greece

National and Kapodistrian University of Athens, Faculty of Geology and Geoenvironment, Athens 15772, Greece

Keywords: Aeolianite Image, Naxos, Deep-Learning, Densely Residual Laplacian Super-Resolution, Sobel Edge

Detection.

Abstract: In the current work processing of Aeolianite images, from a quarry in the island of Naxos in Greece, is

presented. The deep-learning based technique called Densely Residual Laplacian Super-Resolution (DRLN)

is applied on the original images of size 3000×4000 pixels to increase their spatial resolution per the factor

of 4. Edge detection is applied on the initial images as well as on the super-resolved images of

12000×16000 pixels. Visual and numerical comparisons on several Aeolianite scenes prove that the super-

resolved images are advantageous in relation to the initial images of lower spatial resolution, as far as edge

detection and pattern delineation are concerned. The improvement in edge detected components reaches

83%. Classification or pattern extraction could significantly benefit from encompassing the proposed

methodology for Aeolianite images as a preprocessing step.

1 INTRODUCTION

Geology and in particular geomorphology could

greatly benefit from techniques of pattern

recognition and pattern extraction (Erginal et al..,

2022; Helm et al., 2021; Liu et al., 2019; Toulia et

al., 2018). Facies analysis and the optical age of

coastal carbonate Aeolianites serve for the

investigation of the imprints of multiple

Mediterranean transgressions during Middle

Pleistocene in the Black Sea in (Erginal et al., 2022).

In opposition to the contemporary hydro-climate of

the Black Sea, the Aeolianites demonstrate the

transformation of the Black Sea into a warm inland

sea during successive Mediterranean invasions.

Before the inception of aeolian deposition, paleosols

were formed on the Eocene-aged hardened sandy

silts, which indicates strongly washed soil. The

particular study suggests that the carbonate-rich and

ooid-containing Aeolianites were time and time

again formed in the many Mediterranean

transgression stages. In fact, there was a start with an

increasingly severe dry phase coming after the

Brunhes-Matuyama magnetic reversal.

With regard to geological and geomorphology

monitoring, due to the requirement for automatic

processing of the numerous remotely sensed

imagery, numerous techniques of pattern extraction

have been developed. Segmentation, mapping,

recognition and monitoring of rocks, terrain

morphology, ebipenthic/benthic sea communities,

salt marsh and corals have been presented in the

literature (Fan et al., 2020; Goode et al., 2021;

Janowski et al., 2021; Juniani, 2021; Lou et al.,

2022; Song et al., 2021; Untiedt et al., 2021). Deep

seafloor exploration is presented in (Juliani, 2021).

The interpretation of the nature of the geological

phenomena as well as of their complicated

interactivities requires the investigation of seafloor

processes and spatial patterns or motives. Indeed,

huge zones of undersea eruptions represent crucial

territory candidates for new mineral findings. In

specific, seafloor mounds can be greatly informative

about surface changes that are occasionally caused

by seafloor mineral accumulations. The study in

(Juliani, 2021) investigates seafloor mounds by a 2-

folded methodology namely a) semantic

segmentation b) morphological similarity analysis

and clustering of segmented features. Overall, the

244

Panagiotopoulou, A., Ragia, L. and Evelpidou, N.

Deep-Learning Based Super-Resolution of Aeolianite Images on the Purpose of Edge Detection and Pattern Extraction.

DOI: 10.5220/0012038600003473

In Proceedings of the 9th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2023), pages 244-250

ISBN: 978-989-758-649-1; ISSN: 2184-500X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

model developed in the particular study segmented

1,659 features and achieved accuracy up to 84%

pixel-wise, and 80% object-wise, using data

combination of bathymetry and terrain attributes as

input. Morphological patterns that are immediate

after-effects of diversified eruption mechanisms

have been discovered.

Additionally, machine-learning based

classification of rocks, marine coralline algae and

geological structure are met in (Chen et al., 2021; De

Lima et al., 2019; Dos Anjos et al., 2021; Janke et

al., 2015; Piazza et al., 2021; Zhang et al., 2018).

Geological structures being exposed on the Earth

surface namely Anticline, Ripple marks, Xenolith,

Scratch, Ptygmatic folds, Fault, Concretion,

Mudcracks, Gneissose structure, Boodin, Basalt

columns and Dike are under study in (Zhang et al.,

2018). In the particular study there were utilized

2,206 images with 12 labels to identify geological

structures relying on the Inception-v3 model.

Classification of the geological structures was

performed by applying K-nearest neighbors,

artificial neural network and extreme gradient

boosting, relying on features extracted by the Open

Source Computer Vision Library. Experimentation

indicates that model overfitting often leads to poor

performance while accuracy may be smaller than

40%. However, deep-learning based transfer

learning proves robust in the classification of images

with geological structures, where model accuracy

equal to 83.3% and 90% was achieved. Image

texture turns out a crucial feature throughout the

specific study.

In the present work deep-learning based SR is

applied on Aeolianite images having been captured

in a qaurry in the island Naxos, Greece. The

increased image spatial resolution offers the

advantage of more accurate and successfull pattern

delineation on the Aeolianites after edge detection

has been performed. Visual and numerical

comparison are carried out. Section 2 presents the

area of deep-learning based SR and in particular the

Densely Residual Laplacian Super-Resolution

(DRLN) (Anwar and Barnes, 2022). The

experimental procedure along with the results are

given in Section 3. The conclusions are drawn in

Section 4.

2 DEEP-LEARNING BASED

SUPER-RESOLUTION:

DENSELY RESIDUAL

LAPLACIAN

The spatial resolution of images can be increased

beyond the resolution of the image acquisition

sensor through the application of Super-Resolution

techniques (Bratsolis et al., 2018; Stefouli et al.,

2019). SR methodologies stemming from the area of

machine learning and in particular deep learning

have gained tremendous usage increase in the last

years up to now (Anwar and Barnes, 2022; Niu et

al., 2020; Panagiotopoulou et al., 2021;

Panagiotopoulpou et al., 2022; Wenlong et al.,

2021). In the current study the deep-learning based

SR technique called DRLN (Anwar and Barnes,

2022) serves for super-resolution by a factor 4. The

particular technique presents as basis a modular

convolutional neural network that consists of various

components for performance boosting. There is a

cascading residual on the residual network

architecture facilitating the circulation of low-

frequency information.. Additionally, the densely

linked residual blocks end up in “deep supervision”

and learning from high-level complex features.

Furthermore, the DRLN super-resolution technique

presents the Laplacian attention characteristic. Due

to Laplacian attention, the modelling of crucial

features is done per multiple scales whilst the

network catches the inter- and intra-level

dependencies among the maps of features.

3 EXPERIMENTAL PROCEDURE

AND RESULTS

Images of Aeolianites that are situated in a quarry in

Naxos island, Greece are used for the demonstration

of the proposed methodology. The utilized images

can be found in Figure 1.

The initial images have resolution in pixels equal

to 3000 × 4000 pixels. Their spatial resolution gets

increased per the factor of 4 by means of the DRLN

super-resolution technique. The resulting SR images

have the resolution in pixels of 12000 × 16000.

Selected Aeolianite parts from the initial images as

well as from the SR images are depicted in Figures

2-5. Sobel edge detection (Jiang and Scott, 2020) is

applied on the images for the delineation of any

inherent patterns. The corresponding selected edge

detected parts are also shown in Figures 2-5. The

Deep-Learning Based Super-Resolution of Aeolianite Images on the Purpose of Edge Detection and Pattern Extraction

245

number of connected components in the edge

detected Aeolianites can be found in Table 1.

P1050360 P1050396

P1050363 P1050404

P1050372 P1050411

P1050377 P1050422

P1050388 P1050426

Figure 1: The Aeolianite images under consideration, from

quarry in Naxos island, Greece.

The kernel size of the Sobel algorithm is equal to

3 × 3. Automatic count by means of Matlab has

given the number of connected components in Table

1. The enhanced image resolution allows to detect

patterns in the Aeolianites much more successfully

than the lower spatial resolution of the initial

images. Apart from visually, the proposed

methodology is also numerically validated. The

percentage differences of the connected components

or objects in the initial and in the DRLN super-

resolved images in Table 1 prove that super-

resolution contributes to the disclosure of patterns in

Aeolianites. In all 10 Aeolianite image scenes under

consideration, the SR enables the detection of

connected components per 63.30% to 83%.

Table 1: The number of connected components or objects

in the initial images and the corresponding super-resolved

images.

Image ID Initial DRLN Super-

Resolved

Difference

P1050360 126 642 80.37

P1050363 11 65 83.08

P1050372 43 240 82.08

P1050377 217 938 76.87

P1050388 75 228 67.10

P1050396 76 342 77.78

P1050404 126 463 80.40

P1050411 143 699 79.54

P1050422 40 109 63.30

P1050426 144 570 74.74

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

246

P1050360-initial

P1050360-initial-

edge detected

P1050360-DRLN

P1050360-DRLN-

edge detected

P1050363-initial

P1050363-initial-

edge detected

P1050363-DRLN

P1050363-DRLN-

edge detected

P1050372-initial

P1050372-initial-

e detecte

P1050372-DRLN

P1050372-DRLN-

e detecte

Figure 2: Selected small parts from the initial images and the super-resolved images of Aeolianite. The corresponding edge

detected images are also depicted.

Deep-Learning Based Super-Resolution of Aeolianite Images on the Purpose of Edge Detection and Pattern Extraction

247

P1050377-initial P1050377-initial-

edge detected

P1050377-DRLN P1050377-DRLN-

edge detected

P1050388-initial P1050388-initial-

edge detected

P1050388-DRLN P1050388-DRLN-

edge detected

P1050396-initial P1050396-initial-

e detecte

P1050396-DRLN P1050396-DRLN-

e detecte

Figure 3: Selected small parts from the initial images and the super-resolved images of Aeolianite. The corresponding edge

detected images are also depicted.

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

248

P1050404-initial

P1050404-initial-

edge detected

P1050404-DRLN

P1050404-DRLN-

edge detected

P1050411-initial

P1050411-initial-

edge detected

P1050411-DRLN

P1050411-DRLN-

edge detected

P1050422-initial

P1050422-initial-

e detecte

P1050422-DRLN

P1050422-DRLN-

e detecte

Figure 4: Selected small parts from the initial images and the super-resolved images of Aeolianite. The corresponding edge

detected images are also depicted.

P1050426-initial

P1050426-initial-

e detecte

P1050426-DRLN

P1050426-DRLN-

e detecte

Figure 5: Selected small parts from the initial images and the super-resolved images of Aeolianite. The corresponding edge

detected ima

es are also de

icted.

Deep-Learning Based Super-Resolution of Aeolianite Images on the Purpose of Edge Detection and Pattern Extraction

249

4 CONCLUSIONS

In this work the deep-learning based super-

resolution technique called DRLN is applied on

images of Aeolianites from a quarry in Naxos,

Greece. Edge detection for the delineation of

patterns on the images is performed on both the

initial images and the super-resolved images per

factor 4. The SR Aeolianite images reveal the

inherent patterns better than the initial images up to

the percentage of 83%. The methodology that is

presented in this work could serve as an excellent

tool for Aeolianite image preprocessing before a

classification or pattern extraction task.

REFERENCES

Anwar, S., Barnes, N. (2022). Densely Residual Laplacian

SuperResolution. In IEEE Transactions on Pattern

Analysis and Machine Intelligence, vol. 44, pp. 1192-

1204.

Bratsolis E. et al. (2018). Comparison of optimized

mathematical methods in the improvement of raster data

and map display resolution of Sentinel-2 images. In

2018 IEEE ICIP.

Chen, J., Yang, T., Zhang, D., Huang, H., Tian, Y. (2021).

Deep learning based classification of rock structure of

tunnel face. In Geoscience Frontiers, vol. 12, pp. 395-

404.

De Lima, R. P., Bonar, A., Coronado, D. D., Marfurt K.,

and Nicholson, C. (2019). Deep convolutional neural

networks as a geological image classification tool. In

The Sedimentary Record,pp. 4-9.

Dos Anjos, C. E. M. et al. (2021). Deep learning for

lithological classification of carbonate rock micro-CT

images. In Computational Geosciences, vol. 25, pp.

971-983.

Erginal, A. E. et al. (2022). Aeolian imprints of multiple

Mediterannean invasions of the Black Sea during

Pleistocene. In Palaeogeography, Palaeoclimatology,

Palaeoecology, vol. 592.

Fan, G., Chen, F., Chen, D., Li, Y. and Dong, Y. (2020). A

deep-learning model for quick and accurate rock

recognition with smartphones. In Hindawi Mobile

Information Systems, vol. 2020.

Goode, S. L. et al., (2021). Fine-scale mapping of mega-

epibenthic communities and their patch characteristics

on two New Zealand Seamounts. In Frontiers in Marine

Science, vol. 8.

Helm, C. W. et al. (2021). Large geometric patterns from

the middle stone age in Aeolianites on the Cape South

Coast, South Africa. In Rock Art Research, vol. 38, pp.

10-22.

Janowski, L. et al., (2021). Offshore benthic habitat

mapping based on object-based image analysis and

geomorphometric approach. A case study from the

Slupsk Bank, Southern Baltic Sea. In Science of the

Total Environment, vol. 801.

Janke, R. J. et al., (2015). Classification of debris-covered

glaciers and rock glaciers in the Andes of central Chile.

In Geomorphology, vol. 241, pp. 98-121.

Jiang, X. J. and Scott P. J. (2020). Characterization of free-

form structured surfaces, Chapter 11-Advanced

Metrology, pp. 281-317, Academic Press.

Juliani, C. and Juliani, E. (2021). Deep learning of terrain

morphology and pattern discovery via network-based

representational similarity analysis for deep-sea mineral

exploration. In Ore Geology Reviews, vol. 129

Liu, Y., Sun, W. and Durlofsky, L., J. (2019). A deep-

learning-based geological parameterization for history

matching complex models. In Mathematical

Geosciences, vol. 51, pp. 725-766.

Lou, Y. et al. (2022). Image-based machine learning for

monitoring the dynamics of the largest salt marsh in the

Yangtze River Delta. In Journal of Hydrology, vol. 608.

Niu, B. et al. (2020). Single Image Super-Resolution via a

Holistic Attention Network. In Vedaldi A., Bischof H.,

Brox T., Frahm JM. (eds) Computer Vision – ECCV

2020. ECCV 2020. Lect. Notes Comput. Sci., vol12357.

Springer, Cham.

Panagiotopoulou, A, Grammatikopoulos, L., Kalousi, G.

and Charou, E. (2021) “Sentinel-2 and SPOT-7 Images

in Machine Learning Frameworks for Super-

Resolution”. In Lect. Notes Comput. Sci., vol. 12667.

Springer, Cham.

Panagiotopoulou A. et al. (2022). Sentinel-2 images at 2.5m

spatial resolution via deep-learning: a case study in

Zakynthos. In Proceedings of 2022 IEEE IVMSP.

Piazza, G., Valsecchi, C. and Sottocornoloa, G. (2021).

Deep learning applied to SEM images for supporting

marine coralline algae classification. In Diversity, vol.

13.

Song, W., Dong, L.-H., Xia, X.-J. and Liu, T.-M. (2021).

Deep sea nodule mineral image segmentation algorithm

based on Mask R-CNN. In ACM TURC.

Stefouli, M. et al. (2019). Lignite mine monitoring and

mapping using freely-available radar and optical

satellite imagery. In International Journal of Mining

and Mineral Engineering, vol. 10, pp. 131-157.

Toulia, E., Kokinou, E. and Panagiotakis, C. (2018). The

contribution of pattern recognition techniques in

geomorphology and geology: the case study of Tinos

island (Cyclades, Aegean, Greece). In European

Journal of Remote Sensing, vol. 51, pp. 88-99.

Untiedt, C. B. et al. (2021). Identifying Black Corals and

Octocorals from Deep-Sea Imagery for Ecological

Assessments: Trade-offs between morphology and

taxonomy. In Frontiers in Marine Science, vol. 8.

Wenlong, Z., Yihao, L., Chao, D. and Yu. Q. (2021).

RankSRGAN: Generative adversarial networks with

ranker for image super-resolution. In Proceedings of

IEEE/CVF ICCV, pp. 3096-3105.

Zhang, Y., Wang, G., Li, M. and Han, S. (2018). Automated

classification analysis of geological structures based on

images data and deep-learning models. In Applied

Sciences, vol. 8.

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

250