Call Girls in Aiims Metro Delhi ๐ฏCall Us ๐9953322196๐ ๐ฏEscort.
ย
Extracting Surface Water Bodies from Sentinel-2 Imagery
1. al393628@uji.es
janak.parajuli1@gmail.com
05/03/2021
www.mastergeotech.info
Dissertation submitted in partial fulfillment of the requirements for
the Degree of Master of Science in Geospatial Technologies
EXTRACTING SURFACE WATER BODIES FROM SENTINEL-2 IMAGERY USING
CONVOLUTIONAL NEURAL NETWORKS
Supervised by:
Prof. Filiberto Pla Baรฑรณn, PhD
Prof. Marco Painho, PhD
Rubรฉn Fernรกndez-Beltrรกn, PhD
Janak Parajuli
4. PROBLEM STATEMENT
Limitations
of
Indices
โข Not suitable to global
scale
โข Necessity of auxiliary
data
โข Complex band
equations
โข Suitable threshold
values
Problems
with
larger
depths
in
CNN
โข Loss of information
โข Vanishing Gradient
โข Increase in number of
parameters
โข High computational
cost
Applicability
of
state-of-
the-art
CNN
approaches
โข Applied only on
standard datasets
โข Need for thorough
exploration
โข Implementation on
satellite images
5. PROBLEM STATEMENT
Limitations
of
Indices
โข Not suitable to global
scale
โข Necessity of auxiliary
data
โข Complex band
equations
โข Suitable threshold
values
Problems
with
larger
depths
in
CNN
โข Loss of information
โข Vanishing Gradient
โข Increase in number of
parameters
โข High computational
cost
Applicability
of
state-of-
the-art
CNN
approaches
โข Applied only on
standard datasets
โข Need for thorough
exploration
โข Implementation on
satellite images
6. PROBLEM STATEMENT
Limitations
of
Indices
โข Not suitable to global
scale
โข Necessity of auxiliary
data
โข Complex band
equations
โข Suitable threshold
values
Problems
with
larger
depths
in
CNN
โข Loss of information
โข Vanishing Gradient
โข Increase in number of
parameters
โข High computational
cost
Applicability
of
state-of-
the-art
CNN
approaches
โข Applied only on
standard datasets
โข Need for thorough
exploration
โข Implementation on
satellite images
7. AIM & OBJECTIVES
Extracting Surface Water Bodies from
Sentinel-2 Imagery using CNNs
โขExplore state-of-the-art
approaches in CNN
โขImplement & Compare
their performance
โขDesign and propose a
novel approach in CNN
8. WORKFLOW
1
โข Data Download and Pre-Processing
2
โข Preparation of Image-Label Tiles
3
โข Patch Extraction
โข Preparation of Patch-Label Pairs
4
โข Selection & Configuration of Neural
Networks
5
โข Experimentation
6
โข Integration of Models
โข Evolution of New Architect
7
โข Choice of Indices
8
โข Development of EWI
9
โข Experimentation
10
โข Performance Assessment
Pre-Processing
Processing
Post-Processing
15. SELECTION OF NEURAL NETWORKS
Convolution
Batch
Normalization
Blocks ReLU
Global
Average
Pooling
Output
Input
DenseNet proposed by Huang (2017) et al.
16. SELECTION OF NEURAL NETWORKS
Convolution
Batch
Normalization
Blocks ReLU
Global
Average
Pooling
Output
Input
Convolutional Block
DENSE BLOCK 1
TRANSITION BLOCK 1
Convolutional Block
DENSE BLOCK 2
TRANSITION BLOCK 2
Convolutional Block
DENSE BLOCK 3
DenseNet proposed by Huang (2017) et al.
23. First Output
Convolution
Batch
Normalization
Attention Block
Feature
Extraction by
TB
Feature
Selection by
SMB
Dense Block
Feature
Propagation
Re-use
Checks
Vanishing
Gradient
Transition
Block
Volume and
Feature Maps
Halved
Computational
Cost
Reduction
Global
Average
Pooling
Remove
parameters to
optimize
Avoids
Overfitting
Output
Softmax
Classifier
Probability
Calculation
MECHANISM of AttDenseNet
24. 1
โข Only RGB Channels as Input
2
โข Selected S2 Channels as Input
3
โข Contribution of DEM on Accuracy
4 โข S2 Channels Integrated with DEM
EXPERIMENTS
On Patch Sizes
8, 12, 16 & 20
25. SELECTION OF INDICES
โข ๐๐ท๐๐ผ =
๐บ๐๐๐๐โ๐๐ผ๐
Grโ โ ๐+๐๐ผ๐
โข ๐๐ท๐๐ผ =
๐๐ผ๐ โ๐ ๐๐
๐๐ผ๐ +๐ ๐๐
โข ๐๐ท๐๐ผ_๐๐ท๐๐ผ
โข ๐ธ๐๐ผ =
๐๐ท๐๐ผโ(๐๐ถ1+๐๐ถ2)
๐๐ท๐๐ผ+(๐๐ถ1+๐๐ถ2)
Yang (2017) et al.
26. SELECTION OF INDICES
โข ๐๐ท๐๐ผ =
๐บ๐๐๐๐โ๐๐ผ๐
Grโ โ ๐+๐๐ผ๐
โข ๐๐ท๐๐ผ =
๐๐ผ๐ โ๐ ๐๐
๐๐ผ๐ +๐ ๐๐
โข ๐๐ท๐๐ผ_๐๐ท๐๐ผ
โข ๐ธ๐๐ผ =
๐๐ท๐๐ผโ(๐๐ถ1+๐๐ถ2)
๐๐ท๐๐ผ+(๐๐ถ1+๐๐ถ2)
NDWI
Reflectance
Spectral
Graph
PC1
Reflectance
Spectral
Graph
PC2
Reflectance
Spectral
Graph
Input
Derive
Equation
Extract Water
Yang (2017) et al.
27. 79
81
83
85
87
8 12 16 20
Test
Accuracy
Patch Size
TEST ACCURACY VS PATCH SIZE
Baseline
CNNWQC
CNNCWC
SAPCNN
DenseNet
AttResNet
RESULTS & DISCUSSION
62
64
66
68
70
72
74
76
78
80
8 12 16 20
Recall
of
Water
Patch Size
RECALL OF WATER VS PATCH SIZE
Baseline
CNNWQC
CNNCWC
SAPCNN
DenseNet
AttResNet
Quantitative Analysis: All Models
Performance of neural networks for RGB channels as input
28. 84
86
88
90
8 12 16 20
Test
Accuracy
Patch Size
TEST ACCURACY VS PATCH SIZE
Baseline
CNNWQC
CNNCWC
SAPCNN
DenseNet
AttResNet
RESULTS & DISCUSSION
70
72
74
76
78
80
82
84
8 12 16 20
Recall
of
Water
Patch Size
RECALL OF WATER VS PATCH SIZE
Baseline
CNNWQC
CNNCWC
SAPCNN
DenseNet
AttResNet
Quantitative Analysis: All Models
Performance of neural networks for selected S2 channels as input
29. 80
82
84
86
8 12 16 20
Test
Accuracy
Patch Size
TEST ACCURACY VS PATCH SIZE
With DEM
Without DEM
RESULTS & DISCUSSION
64
66
68
70
72
74
76
8 12 16 20
Recall
of
Water
Patch Size
RECALL OF WATER VS PATCH SIZE
With DEM
Without DEM
Quantitative Analysis: Baseline Model
Contribution of DEM with RGB on Baseline model
30. 86
88
90
92
8 12 16 20
Test
Accuracy
Patch Size
TEST ACCURACY VS PATCH SIZE
Baseline
CNNWQC
CNNCWC
SAPCNN
DenseNet
AttResNet
RESULTS & DISCUSSION
76
78
80
82
84
86
8 12 16 20
Recall
of
Water
Patch Size
RECALL OF WATER VS PATCH SIZE
Baseline
CNNWQC
CNNCWC
SAPCNN
DenseNet
AttResNet
Quantitative Analysis: All Models
Performance of neural networks for selected
S2 channels integrated with DEM
31. 88
90
92
8 12 16 20
Test
Accuracy
Patch Size
TEST ACCURACY VS PATCH SIZE
AttDenseNet
DenseNet
RESULTS & DISCUSSION
80
82
84
86
88
8 12 16 20
Recall
of
Water
Patch Size
RECALL OF WATER VS PATCH SIZE
AttDenseNet
DenseNet
Quantitative Analysis: AttDenseNet
Proposed Network (AttDenseNet) Vs DenseNet for
selected S2 channels integrated with DEM as input
32. 72
74
76
78
80
82
84
86
8 12 16 20
Test
Accuracy
Patch Size
TEST ACCURACY VS PATCH SIZE
Baseline
NDWI
NDVI
NDVI_NDWI
RESULTS & DISCUSSION
Quantitative Analysis: Comparison with Indices
Performance comparison of Baseline for only RGB
channels as input with index-based approaches
48
56
64
72
8 12 16 20
Recall
of
Water
Patch Size
RECALL OF WATER VS PATCH SIZE
Baseline
NDWI
NDVI
NDVI_NDWI
*EWI done only for qualitative analysis
due to memory issues
36. LIMITATIONS
Hardware
Memory
โข Step = 8
โข No PCs for EWI
โข Required
Reduction Factor
Spatial
Resolution
โข Maximum 10m
โข High Resolution
preferable
Image-Label
Compatibility
โข Image from 2020
โข Label from 2015
Modelโs
Configuration
โข Same for all
models
โข No superpixels
for SAPCNN
โข AttResNet was
affected much
37. RECOMMENDATION
Further Enhancement of Proposed Network
Introduction of Temporal Component
Original configuration of Models would produce better results
38. CONCLUSION
A novel CNN-based approach proposed for water extraction
problems
Neural Networks better than traditional approaches
State-of-the-art approaches successfully implemented in
satellite imageries
39. F. Pu, C. Ding, Z. Chao, Y. Yu, and X. Xu. โWater-quality classification of inland lakes using landsat8 images by
convolutional neural networks. โIn: Remote Sensing11.14 (2019), p. 1674
F. Wang, M. Jiang, C. Qian, S. Yang, C. Li, H. Zhang, X. Wang, and X.Tang. โResidual Attention Network for Image
Classification.โ In: (Apr.2017).url: http://arxiv.org/abs/1704.06904
G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger. โDensely Connected Convolutional Networks.โ In:2017
IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Vol. 2017-January. IEEE,2017, pp. 2261โ
2269.isbn: 978-1-5386-0457-1.doi: 10.1109/CVPR.2017 . 243. arXiv: 1608 . 06993.
J. Yang and X. Du. โAn enhanced water index in extracting water bodies from Landsat TM imagery.โ In: Annals of
GIS23.3 (2017), pp. 141โ148.issn: 19475691.doi: 10.1080/19475683.2017.1340339.url: https://doi.org/10.1080/194
75 683 .2017.1340339
M. Rezaee, M. Mahdianpari, Y. Zhang, and B. Salehi. โDeep convolutional neural network for complex wetland
classification using optical remote sensing imagery.โ In: IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing11.9 (2018), pp. 3030โ3039
Y. Chen, R. Fan, X. Yang, J. Wang, and A. Latif. โExtraction of urban water bodies from high-resolution remote-
sensing imagery using deep
REFERENCES
43. NOVELTY & CONTRIBUTION
Novel approach for feature extraction in the study area
Implementation of DenseNet & AttResNet to extract water
Proposition of a novel approach in CNN