deep learning competence center & smart data and knowledge … · 2018. 4. 19. ·...

52
Patrick Helber, Benjamin Bischke, Damian Borth, Andreas Dengel Deep Learning Competence Center & Smart Data and Knowledge Services German Research Center for Artificial Intelligence (DFKI) Earth Observation From Space Deep Learning Based Satellite Image Analysis

Upload: others

Post on 13-Oct-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Patrick Helber, Benjamin Bischke, Damian Borth, Andreas Dengel

Deep Learning Competence Center & Smart Data and Knowledge Services

German Research Center for Artificial Intelligence (DFKI)

Earth Observation From SpaceDeep Learning Based Satellite Image Analysis

Page 2: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Satellite Data Becomes Publicly Available

2

© ESA / Astrium

Page 3: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Global Challenges

3

Page 4: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Satellite Analysis -> Global Challenges

4

Infrastructure & Traffic MonitoringPollution Monitoring: Oil Spills and Pipelines

Land Change Detection, DeforestationAgriculture

Page 5: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Agriculture: Monitoring Farm Land

5

© ESA

Page 6: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Climate Change: Deforestation

6

Page 7: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Monitoring of Economic Factors

7

• Economic Growing and Shrinking

© DigitalGlobe

Page 8: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Natural Disaster: Flooding, Wildfire, Earthquake …

8

Page 9: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Earth Observing SatellitesOptical and Radar

9

Page 10: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Satellites – Optical View

• Sentinel-2

– 10 meters per pixel

– Two-satellite constellation

– 5 days revisit time

– Global land surface coverage• Onshore, large islands, inland and coastal waters

• Worldview, GeoEye, SkySat

– Up to ca. 0.30 meters per pixel

– High-resolution images

10

Author: Raman

Page 11: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Optical View: 10 meters/pixel

11

Page 12: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Optical View: High-resolution up to ca. 30cm/pixel

12

Page 13: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Radar Satellites – Sentinel 1

• Synthetic Aperture Radar (SAR)

• Sentinel-1

– 6 day repeat cycle

– Two satellite constellation

• For the first time continuous

radar scanning

13

© ESA

Page 14: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Radar-Based Earth Observation

14

Page 15: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Satellite Image ClassificationOptical Images

15

Page 16: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

EuroSAT Publicly Released by DFKI

16

Page 17: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Ground Truth Data

• European Urban Atlas

– Detailed mapping for 695 cities distributed over 30 Europeancountries

– Released August 2016

– Covered time period: 2011-2013

17

P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,” arXiv preprint arXiv:1709.00029, 2017.

Page 18: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

EuroSAT - Distributed Over 30 Countries

• 10 classes

• 27,000 images

• 64 x 64 images

• 2,000 – 3,000 images per class

• 13 spectral bands• Spatial resolution: 10 m per pixel

18

P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification,” arXiv preprint arXiv:1709.00029, 2017.

Page 19: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Land-Use and Land-Cover Classification

19

Industrial Residential Highway Annual Crop Permanent Crop Pasture

Sea & Lake RiverForestHerbaceous vegetation

Page 20: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Classification with ResNet and GoogLeNet

• Dataset split

– 80% Training and 20% Testing

• Convolutional Neural Networks

– ResNet-50

– GoogLeNet

20

C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1–9, 2015.

K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.

Inception Module

ResNetGoogLeNet

Residual Building Block

Page 21: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Models Pre-trained on ImageNet

21

• Transfer learning

– Fine-tuning of pre-trained networks

• Pre-trained on ImageNet

– ILSVRC-2012

Band combination

RGB SWIR CI

Accuracy 0.9857 0.9705 0.9830

SWIR = Short-Wave-Infrared

CI = Color-Infrared

Page 22: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Classification Results

22

• Transfer learning

– Fine-tuning of pre-trained networks

• Pre-trained on ImageNet

– ILSVRC-2012

Band combination

RGB SWIR CI

Accuracy 0.9857 0.9705 0.9830

SWIR = Short-Wave-Infrared

CI = Color-Infrared

Page 23: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Spotting Land Use Changes

23

Residential Area Built Upin Dallas, USA

August 2015 March 2017

• Large-scale scanning and monitoring

• Time component• High frequency

• Near-real-time

• Future availability

• Innovative applications

• Building systems for future

real-time applications

Page 24: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Spotting Land Use Changes

October 2015 September 2016

Deforestation (Forest Clearing)in Villamontes, Bolivia

24

Industrial Buildings Demolished in Shanghai, China

December 2015 December 2016

Page 25: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Support Mapping Services

25

Melbourne, Australia Shanghai, China

Page 26: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Look More CloselyClassification with High-Resolution Optical Images

26

Page 27: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Functional Map of the World

• Four different satellites

• Different spatial and spectralresolutions

– 4-band and 8-band

• Temporal image sequences

• 63 land-use/-cover and building types

– Security & Defense

– Emergency Response

– Infrastructure

27

Christie, Gordon and Fendley, Neil and Wilson, James and Mukherjee, Ryan, “Functional Map of the World” CVPR, 2018.

Airport Car DealershipBorder Checkpoint

Police Station Shopping MallFlooded Road

Lighthouse DamSolar Farm

Page 28: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Task: Sequence Classification

• Sequences determine the underlying class

– Construction sites

– Flooded road

28

Page 29: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Classification and Fusion Approaches

• DenseNet-161

– CNN Features (I)

– Metadata Features (M)

– CNN + Metadata Features (IM)

• Sequence Classification

with LSTM

29

DenseNet

Approach LSTM-M CNN-I LSTM-I CNN-IM LSTM-IM

F1-Score 0.193 0.679 0.688 0.722 0.734

Christie, Gordon and Fendley, Neil and Wilson, James and Mukherjee, Ryan, “Functional Map of the World” CVPR, 2018.

CNN (Features)

Metadata Features

Page 30: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Misclassification

Hard to classify

• Port Shipyard

• Office building Fire station, police station

• Hospital Educational institution

• Police station Educational institution

• Multi unit residential Single unit residential

30

Shipyard Port

HospitalEducational Institution

Approach Port Shipyard Police station

Office Building

Hospital

F1-Score 0.193 0.351 0.329 0.225 0.458

Page 31: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Satellite Image SegmentationHigh-Resolution Optical Images (Aerial)

31

Page 32: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Fine-Grained Segmentation With 10cm/Pixel Images

32

Page 33: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Fine-Grained Cadastral Map for Land Change Detection

33

Work in progress!

Page 34: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Netherlands: High-Resolution RGB+NIR

34

Page 35: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Leverage Learning with Altitude Measurements

35

Page 36: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Beyond the Visible SpectrumMultispectral Signatures

36

Page 37: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Multispectral Classification

37

EuroSATRGB

FMoW

Page 38: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Multispectral Segmentation

• Damage Estimation Caused by Wildfires

– Dataset• Satellite: Sentinel-2

• Mapping: EMS Coperncicus

• 13 spectral bands

• 1.564 scenes from 5 different countries

• 256px images

– Band combinations• Visible spectrum: RGB

• Infrared spectrum: Red Edge

• Near-Infrared: NIR

38

Page 39: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

RGB + Red Edge + NIR Spectral Information

39

SegNet

SegNet Input Pix. Accuracy mIOU

RGB 0.9878 0.7657

Red Edge 0.9885 0.7592

NIR 0.9815 0.4964

RGB + NIR 0.9864 0.7844

RGB + Red Edge 0.9855 0.7928 (+2,71)

RGB + Red Edge + NIR Prediction

loss

epoch

Page 40: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Qualitative Results

40

GT RGB RGB + Red Edge

RGB Red Edge10m 20m

Page 41: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Beyond the Visible SpectrumRadar Waves

41

Page 42: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Optical Satellite Images - Obstructed View

42

Page 43: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Optical Satellite Images - Obstructed View

43

Page 44: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Optical Satellite Images - Obstructed View

44

Page 45: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Optical Satellite Images - Obstructed View

45

Page 46: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Vision Independent of Clouds

46

Page 47: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Radar-Based Earth Observation Imagery

47

VVVH

© ESA

• Synthetic Aperture Radar (SAR)

• Satellite: Sentinel-1

• Dual-Polarization

– VV and VH

• Scenarios• Deforestation

• Find safest route for icebreakers

• Land monitoring

• Ship traffic monitoring

• Natural disasters like flooding and earthquake

Page 48: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Dual-Polarized SAR

48

Flooding

• Satellite: Sentinel-1

• Mapping: EMS Copernicus

• 9.426 radar image scenes

• 56 flooding events in 12 countries

• Globally spread• Australia, USA, France, UK, …

Page 49: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Segmentation with Dual-Polarized Radar Waves

• Convolutional Networks

– SegNet: Convolutional Encoder-Decoder Network

49

SegNet PredictionInput

VV + VH

Model Pix. Accuracy mIOU

FCN-6 0.9496 0.6772

VGG16 0.9413 0.6740

SegNet 0.9530 0.6541

Segmentation

Page 50: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Natural Disasters: Damage Maps

Flooding: Damage Estimation Fire: Damage Estimation

• Biomass by radar waves

50

• Government, Insurance, Investment

Copernicus Emergency Management Service (© European Union, 2012-2017)

Destroyed

Highly damaged

Slight damage

Page 51: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Multimedia Satellite Task 2018

• Emergency Response for Flooding Events

• Focus on Road Passibility (Impact of Infrastructure) (Road Access, blocked Road) with two subtasks

• Classification of Road-Access & Passability– in Social Multimedia

– in multiple Satellite Images (Radar, Optical)

• Registration open (http://www.multimediaeval.org)

51

Images in Tweets (Hurricane Harvey)

Satellite Images of Houston (US) for Hurricane Harvey

CombineResultvia Geo-Location

Page 52: Deep Learning Competence Center & Smart Data and Knowledge … · 2018. 4. 19. · Classification and Fusion Approaches • DenseNet-161 –CNN Features (I) –Metadata Features (M)

Thanks!

52

NVAIL partner

all networks trained on DGX-1