Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Accurate Segmentation of Hyperspectral Images UsingDeep Neural Networks—Are We There Yet?
Jakub Nalepa, Michal Marcinkiewicz, Pablo Ribalta Lorenzo, Krzysztof Czyz,Michal Kawulok & HYPERNET Team
KP Labs, Gliwice, PolandSilesian University of Technology, Gliwice, Poland
The ESA Earth Observation �-Week,Frascati, Italy. November 14, 2018
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Introduction
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 1 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
About us
HYPERspectral image segmentation using deep neural NETworksHYPERNET
May 2018—May 2019Polish Industry Incentive Scheme
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 2 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 3 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Segmentation of HSI—conventional machine learning
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 4 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Segmentation of HSI—deep learning
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 5 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Deep networks in the wild. . .
Hyperspectral image segmentationMedical imagingObject detectionText classificationTemporal and time-series analysisSelf-driving carsMusic compositionReal-time analysis of behaviorsTranslationSpeech recognitionLanguage modelingDocument summarization. . .
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 6 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Deploying a deep neural network in the wild
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 7 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Deploying a deep neural network in the wild
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 8 / 30
The roadmapDesign a topologySelect hyper-parametersTrain the network
- Topology: P. Ribalta and J. Nalepa, pp 505-512, Proc. GECCO, ACM 2018.- Hyper-params: P. Ribalta and J. Nalepa, et al., pp 481-488, Proc. GECCO, ACM 2017.- Hyper-params: P. Ribalta and J. Nalepa, et al., pp 1864-1871, Proc. GECCO, ACM 2017.
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Deploying a deep neural network in the wild
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 8 / 30
The roadmapDesign a topologySelect hyper-parametersTrain the network
- Topology: P. Ribalta and J. Nalepa, pp 505-512, Proc. GECCO, ACM 2018.- Hyper-params: P. Ribalta and J. Nalepa, et al., pp 481-488, Proc. GECCO, ACM 2017.- Hyper-params: P. Ribalta and J. Nalepa, et al., pp 1864-1871, Proc. GECCO, ACM 2017.
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Deep networks for hyperspectral image segmentation
Spectral Spatial-spectral
(Selected) Open issues:Validation of hyperspectral image segmentation algorithmsResouce-frugality of deep neural networksRobustness of deep neural networks
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 9 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
About usHyperspectral imagingDeep networks for hyperspectral image segmentation
Deep networks for hyperspectral image segmentation
Spectral Spatial-spectral
(Selected) Open issues:Validation of hyperspectral image segmentation algorithmsResouce-frugality of deep neural networksRobustness of deep neural networks
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 9 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Validation of hyperspectral image
segmentation algorithms
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 10 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Selection of training, validation, and test sets
How to select training, validation, and test sets?What are the benchmark (ground-truth) datasets?
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 11 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Selection of training, validation, and test setsMethod Datasetsú SettingsMultiscale superpixels (Dundar & Ince, 2018) IP, PU RandomWatershed + SVM (Tarabalka et al., 2010) PU ArbitraryClustering (SVM) (Bilgin et al., 2011) Washington DC, PU Full imageMultiresolution segm. (Amini et al., 2018) Three in-house datasets RandomRegion expansion (Li et al., 2018) PU, Sa, KSC Full imageDBN (spatial-spectral) (Li et al., 2014) HU RandomDBN (spatial-spectral) (Chen et al., 2015) IP, PU RandomDeep autoencoder (Chen et al., 2014) KSC, PU Monte CarloCNN (Zhao & Du, 2016) PC, PU, RandomCNN (Chen et al., 2016) IP, PU, KSC Monte CarloActive learning + DBN (Liu et al., 2017) PC, PU, Bo RandomDBN (spectral) (Zhong et al., 2017) IP, PU RandomRNN (spectral) (Mou et al., 2017) PU, HU, IP RandomCNN (Santara et al., 2017) IP, Sa, PU RandomCNN (Lee & Kwon, 2017) IP, Sa, PU Monte CarloCNN (Gao et al., 2018) IP, Sa, PU Monte CarloCNN (Ribalta et al., 2018) Sa, PU Randomú IP—Indian Pines; PU—Pavia University; Sa—Salinas; KSC—Kennedy SpaceCenter; HU—Houston University; PC—Pavia Centre; Bo—Botswana
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 12 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Benchmark hyperspectral images—examples
Broccoli 1 Broccoli 2 Fallow 1 Fallow 2 Fallow 3 Stubble Celery Grapes Soil Corn Lettuce 1 Lettuce 2 Lettuce 3 Lettuce 4 Vineyard 1 Vineyard 2
Gravel
Trees
Bricks
Bitumen
Bare Soil
Metal
Ashpalt
Meadows
Shadow
Salinas Valley (512x217px, AVIRIS, 3.7m, 224 bands) Pavia University (610x340px, ROSIS, 1.3m, 103 bands)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 13 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Selection of training, validation, and test setsMethod Datasetsú SettingsMultiscale superpixels (Dundar & Ince, 2018) IP, PU RandomWatershed + SVM (Tarabalka et al., 2010) PU ArbitraryClustering (SVM) (Bilgin et al., 2011) Washington DC, PU Full imageMultiresolution segm. (Amini et al., 2018) Three in-house datasets RandomRegion expansion (Li et al., 2018) PU, Sa, KSC Full imageDBN (spatial-spectral) (Li et al., 2014) HU RandomDBN (spatial-spectral) (Chen et al., 2015) IP, PU RandomDeep autoencoder (Chen et al., 2014) KSC, PU Monte CarloCNN (Zhao & Du, 2016) PC, PU, RandomCNN (Chen et al., 2016) IP, PU, KSC Monte CarloActive learning + DBN (Liu et al., 2017) PC, PU, Bo RandomDBN (spectral) (Zhong et al., 2017) IP, PU RandomRNN (spectral) (Mou et al., 2017) PU, HU, IP RandomCNN (Santara et al., 2017) IP, Sa, PU RandomCNN (Lee & Kwon, 2017) IP, Sa, PU Monte CarloCNN (Gao et al., 2018) IP, Sa, PU Monte CarloCNN (Ribalta et al., 2018) Sa, PU Randomú IP—Indian Pines; PU—Pavia University; Sa—Salinas; KSC—Kennedy SpaceCenter; HU—Houston University; PC—Pavia Centre; Bo—Botswana
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 14 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Selection of training, validation, and test setsMethod Datasetsú SettingsMultiscale superpixels (Dundar & Ince, 2018) IP, PU RandomWatershed + SVM (Tarabalka et al., 2010) PU ArbitraryClustering (SVM) (Bilgin et al., 2011) Washington DC, PU Full imageMultiresolution segm. (Amini et al., 2018) Three in-house datasets RandomRegion expansion (Li et al., 2018) PU, Sa, KSC Full imageDBN (spatial-spectral) (Li et al., 2014) HU RandomDBN (spatial-spectral) (Chen et al., 2015) IP, PU RandomDeep autoencoder (Chen et al., 2014) KSC, PU Monte CarloCNN (Zhao & Du, 2016) PC, PU, RandomCNN (Chen et al., 2016) IP, PU, KSC Monte CarloActive learning + DBN (Liu et al., 2017) PC, PU, Bo RandomDBN (spectral) (Zhong et al., 2017) IP, PU RandomRNN (spectral) (Mou et al., 2017) PU, HU, IP RandomCNN (Santara et al., 2017) IP, Sa, PU RandomCNN (Lee & Kwon, 2017) IP, Sa, PU Monte CarloCNN (Gao et al., 2018) IP, Sa, PU Monte CarloCNN (Ribalta et al., 2018) Sa, PU Randomú IP—Indian Pines; PU—Pavia University; Sa—Salinas; KSC—Kennedy SpaceCenter; HU—Houston University; PC—Pavia Centre; Bo—Botswana
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 15 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Training-test information leak
t1
t2
t3
1 2
3 4
Training (ti) and test (Âi) pixels with their spatial neighborhoods.ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 16 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Patch-based training-test splitsa) b) c) d) e) f) g)
a) b) c) d) e) f) g)
a) True-color composite, b) ground-truth, c)–g) 5 folds (black/white patches are for training).ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 17 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Do train-test splits matter (Monte-Carlo vs. Patch-based)?
Deep neural network architecturesSpatial-spectral (3D) CNN (Gao et al., Remote Sens., 2018)Spectral (1D) CNN
Training-test subset cardinalities: as in Gao et al., 2018 (balanced andimbalanced)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 18 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Do train-test splits matter (Monte-Carlo vs. Patch-based)?
Deep neural network architecturesSpatial-spectral (3D) CNN (Gao et al., Remote Sens., 2018)Spectral (1D) CNN
Convolution
Batch norm
.
Max pooling
Fully connected …
f = 200 s = 1×1×5
s=2×2
Softmax
l1=512, l2=128
1×1×b
Training-test subset cardinalities: as in Gao et al., 2018 (balanced andimbalanced)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 18 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Do train-test splits matter (Monte-Carlo vs. Patch-based)?
Deep neural network architecturesSpatial-spectral (3D) CNN (Gao et al., Remote Sens., 2018)Spectral (1D) CNN
Convolution
Batch norm
.
Max pooling
Fully connected …
f = 200 s = 1×1×5
s=2×2
Softmax
l1=512, l2=128
1×1×b
Training-test subset cardinalities: as in Gao et al., 2018 (balanced andimbalanced)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 18 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Sneak-peek from the results (Salinas Valley). . .
OA—overall accuracy, AA—average accuracy
Conclusion: Monte-Carlo cross-validation renders over-optimistic insights into classification performance.
For more results, see J. Nalepa et al., https://arxiv.org/abs/1811.03707, 2018 (submitted to IEEEGeoscience and Remote Sensing Letters)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 19 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Sneak-peek from the results (Salinas Valley). . .
OA—overall accuracy, AA—average accuracy
Conclusion: Monte-Carlo cross-validation renders over-optimistic insights into classification performance.
For more results, see J. Nalepa et al., https://arxiv.org/abs/1811.03707, 2018 (submitted to IEEEGeoscience and Remote Sensing Letters)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 19 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Resource-frugality of deep neural
networks
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 20 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Hardware-constrained environment
Reduction
ú
of:Memory footprintInference time
*Without adversely a�ecting classification performance. . .
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 21 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Hardware-constrained environment
Reductionú of:Memory footprintInference time
*Without adversely a�ecting classification performance. . .
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 21 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Quantization of deep neural networksQuantization of deep neural networks
During DNNtraining
Post-training
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 22 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Quantization of deep neural networksQuantization of deep neural networks
During DNNtraining
Post-training
Prune redundant network nodesCompress constantsMap weights/parameters to 8-bit precision
Min Max
-127 0 127
0.0
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 23 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Sneak-peek from the results. . .
Con
volu
tion
Batch
norm
.
Max p
oolin
g
Drop
out
Fully
connected
…
Con
volu
tion
Batch
norm
.
f = 96s = 5×1
f=32s=1×1 s=128s=2×2
Softmax
OA: 84.62% æ 83.45% (drop of 1.17%) OA: 78.30% æ 77.02% (drop of 1.28%)
For more results, see: P. Ribalta, M. Marcinkiewicz, J. Nalepa, Proc. IEEE DSD,pp 260-267, 2018.
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 24 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Sneak-peek from the results. . .
Con
volu
tion
Batch
norm
.
Max p
oolin
g
Drop
out
Fully
connected
…
Con
volu
tion
Batch
norm
.
f = 96s = 5×1
f=32s=1×1 s=128s=2×2
Softmax
0 100 200 300 400 500
Size in KBs
Model sizes for Salinas Valley dataset
Non-quantized Quantized
0 100 200 300
Size in KBs
Model sizes for Pavia University dataset
Non-quantized Quantized
OA: 84.62% æ 83.45% (drop of 1.17%) OA: 78.30% æ 77.02% (drop of 1.28%)
For more results, see: P. Ribalta, M. Marcinkiewicz, J. Nalepa, Proc. IEEE DSD,pp 260-267, 2018.
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 24 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Robustness of deep neural networks
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 25 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Are deep networks robust (against noise)?Simulation: Gaussian noise, SNR wavelength-dependentDeep network: Spatial-spectral CNN
OA: 90.59% OA: 31.90% OA: 48.39% 65.78%
(Thanks to Adam Popowicz for the noise model.)ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 26 / 30
IntroductionOpen issues in hyperspectral image segmentation
Conclusions
Validation of HSI segmentationResource-frugality of DNNs for HSIRobustness of DNNs for HSI
Are deep networks robust (against noise)?Simulation: Gaussian noise, SNR wavelength-dependentDeep network: Spatial-spectral CNN
OA: 90.59% OA: 31.90% OA: 48.39% 65.78%
(Thanks to Adam Popowicz for the noise model.)ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 26 / 30
IntroductionOpen issues in hyperspectral image segmentation
ConclusionsTake-home messageHYPERNET
Conclusions
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 27 / 30
IntroductionOpen issues in hyperspectral image segmentation
ConclusionsTake-home messageHYPERNET
Conclusions
We are not there yet
Open issues in hyperspectral image analysisLack of ground-truth dataResource-frugality of deep neural nets for HSI segmentationRobustness of deep neural nets for HSI segmentationBetter generalization for unseen data (æ better regularizers)E�cient pre-/post-processing of HSI data (e.g., band selection)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 28 / 30
IntroductionOpen issues in hyperspectral image segmentation
ConclusionsTake-home messageHYPERNET
Conclusions
We are not there yetOpen issues in hyperspectral image analysis
Lack of ground-truth dataResource-frugality of deep neural nets for HSI segmentationRobustness of deep neural nets for HSI segmentationBetter generalization for unseen data (æ better regularizers)E�cient pre-/post-processing of HSI data (e.g., band selection)
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 28 / 30
IntroductionOpen issues in hyperspectral image segmentation
ConclusionsTake-home messageHYPERNET
HYPERNET—code and beyond
https://github.com/ESA-PhiLab/hypernet
HYPERNET Team: Jakub Nalepa, Michal Kawulok, Michal Myller,Lukasz Tulczyjew, Marek Antoniak, Rafal Zogala
State-of-the-art spectral and spatial-spectral deep nets for HSI (and beyond)Band selection using attention-based convolutional neural nets (P. Ribalta, L. Tulczyjew, M.Marcinkiewicz, J. Nalepa, https://arxiv.org/abs/1811.02667, after the first round of reviews inIEEE Transcations on Geoscience and Remote Sensing)Data augmentation (generative adversarial nets, noise-based,. . . )Visualization of HSIPatch-based benchmark generation, and more. . .
All in Python (Keras/Pytorch) with Jupyter-notebook examples
Wanting more? Drop me a line at [email protected]
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 29 / 30
IntroductionOpen issues in hyperspectral image segmentation
ConclusionsTake-home messageHYPERNET
HYPERNET—code and beyond
https://github.com/ESA-PhiLab/hypernet
HYPERNET Team: Jakub Nalepa, Michal Kawulok, Michal Myller,Lukasz Tulczyjew, Marek Antoniak, Rafal Zogala
State-of-the-art spectral and spatial-spectral deep nets for HSI (and beyond)Band selection using attention-based convolutional neural nets (P. Ribalta, L. Tulczyjew, M.Marcinkiewicz, J. Nalepa, https://arxiv.org/abs/1811.02667, after the first round of reviews inIEEE Transcations on Geoscience and Remote Sensing)Data augmentation (generative adversarial nets, noise-based,. . . )Visualization of HSIPatch-based benchmark generation, and more. . .All in Python (Keras/Pytorch) with Jupyter-notebook examples
Wanting more? Drop me a line at [email protected]
ESA �-Week 2018 J. Nalepa et al.: Hyperspectral image segmentation—are we there yet? 29 / 30
Accurate Segmentation of Hyperspectral Images UsingDeep Neural Networks—Are We There Yet?
Jakub Nalepa, Michal Marcinkiewicz, Pablo Ribalta Lorenzo, Krzysztof Czyz,Michal Kawulok & HYPERNET Team
KP Labs, Gliwice, PolandSilesian University of Technology, Gliwice, Poland
The ESA Earth Observation �-Week,Frascati, Italy. November 14, 2018