To Understand the Earth and Us? GPU Tech Conference 2019 (S9495) - - PowerPoint PPT Presentation

▶
to understand the earth and us
SMART_READER_LITE
LIVE PREVIEW

To Understand the Earth and Us? GPU Tech Conference 2019 (S9495) - - PowerPoint PPT Presentation

Jun 15, 2023 239 likes •659 views

How AI is Changing the Way To Understand the Earth and Us? GPU Tech Conference 2019 (S9495) Taegyun Jeon Founder and CEO SI Analytics Contents Earth Observation with Artificial Intelligence Case #1: Object Detection and Classification

slide-1
SLIDE 1

GPU Tech Conference 2019 (S9495)

How AI is Changing the Way To Understand the Earth and Us?

Taegyun Jeon

Founder and CEO SI Analytics

slide-2
SLIDE 2

Contents

  • Earth Observation with Artificial Intelligence
  • Case #1: Object Detection and Classification with TensorRT
  • Case #2: Road Extraction (DeepGlobe Challenge)
  • Conclusions
slide-3
SLIDE 3

Earth Observation

01

slide-4
SLIDE 4
slide-5
SLIDE 5

Earth Observation

✓ defense & Intelligence ✓ infrastructure monitoring ✓ forecasting weather ✓ biodiversity and wildlife trends ✓ land-use change ✓ natural disasters ✓ natural resources ✓ agriculture ✓ emerging diseases ✓ mitigating climate change ✓ maritime monitoring

slide-6
SLIDE 6

KITSAT-1 (1992) GSD: 400m KITSAT-2 (1993) GSD: 200m KITSAT-3 (1999) GSD: 13m

slide-7
SLIDE 7

KOMPSAT-3A (2015) GSD: 0.55m

slide-8
SLIDE 8

This image of New York City, taken Nov. 4, 2015, by South Korea's Kompsat-3A satellite, is an example of the products that SI Imaging Services of Korea has begun selling on the market.

slide-9
SLIDE 9

9

Earth Observation with Artificial Intelligence

EO with AI Traditional EO

ORDERING On-demand data On-demand analysis Reactive tasking based on single satellites Reactive tasking based on constellations Data cost is driven by the data source (higher CAPEX system equates to higher data prices); lower-cost systems would imply lower data prices and services development PROCESSING Owned data analysis Cloud approach + Owned data analysis Manual/automated operations

  • n desktop or internal network

Deep Learning based on Big Data DELIVERING Ad hoc services, ordering through reseller or web-portal tasking Service subscription basis Reselling network, privileged distributors (government user focused) Platform deliveries (private sector focused) and reselling network for governments

slide-10
SLIDE 10

KOMPSAT Archive KOMPSAT-2 (EO) KOMPSAT-3 (EO) KOMPSAT-3A (EO) KOMPSAT-5 (SAR) Scenes (Dec 15, 2016) 2,645,022 781,389 80,340 52,245 Data volume (TB) 743 TB 700 TB 59 TB 104 TB Coverage per day (km2) 1,700,000 300,000 240,000 Up to 1,000,000

South Korea (100,210 km2) England (243,610 km2) USA (9,834,000 km2)

KOMPSAT-2 Coverage

Coverage

slide-11
SLIDE 11

MNIST (28,28,1) ImageNet (224,224,3) SpaceNet (3K,3K,8)

Satellite Scene (25K, 25K, 4)

0.7KB 150KB 87MB

2.5GB

Volume

slide-12
SLIDE 12

$60M $835M++ (Satrec Initiative) SpaceEye-X ~0.5m resolution (DigitalGlobe) WorldView-4 ~0.3m resolution

Resolution

slide-13
SLIDE 13

Reusable rocket and Constellation space program

✓ Low launch cost ✓ Low manufacturing cost ✓ Huge daily data

slide-14
SLIDE 14

Object Detection and Classification

02

slide-15
SLIDE 15

Detection and Classification

  • Aircraft Detection & Classification

▪ Task: Detect and classify all aircraft

  • n North Korea Airforce bases

▪ Construct Own Dataset for civil aircraft and military fighters ▪ Compatibility: Transfer Learning (GoogleEarth & KOMPSAT 2, 3, 3A) ▪ Detection Accuracy: 89% ▪ Classification Accuracy: 95.2% ▪ Target Area: All NK Airforce bases ▪ Fill the gap for rare observation: Combine synthetic data from GAN

Magnified view

  • Prob. (Detection)
  • Prob. (Classification)

Detection Results (Haneda Airport from KOMPSAT-2, 3, 3A) Detection and Classification (NK Airforce bases from GoogleEarth) Automatically generated deployment status report (NK Airforce) User Interface for Detection and Classification

15

slide-16
SLIDE 16

Detection and Classification

  • Objective: Detect aircraft and fighter, then classify the types of aircraft

16

Overlap:

slide-17
SLIDE 17

North Korean Air Forces (25 regions)

17

slide-18
SLIDE 18

Detection Results

  • ROI: 25 Airports (North Korea)
  • Detection results:

Precision (0.84), Recall (0.79), F1 (0.82)

  • Classification results:

Top-1 (91.5%), Top-3 (95.4%)

18

  • S. Jeon, J. Seo and T. Jeon, “Multi-task Learning for Fine-grained Visual Classification of Aircraft”, MLAIP Workshop @ ACML (2017)
slide-19
SLIDE 19

Classification with TensorRT

37.74 525.57

83.5 880.35

100 200 300 400 500 600 700 800 900 1000 DenseNet (512,512,3) VGG (128,128,3) Image/s w/o TRT w/ TRT

* Experiments on DGX-Station

2.2X 1.6X

slide-20
SLIDE 20

20

Probability (Detection) Probability (Classification) Magnified view

slide-21
SLIDE 21

21

2

slide-22
SLIDE 22

22

slide-23
SLIDE 23

Synthetic Data Generator and Refiner

Adversarial Learning to refine the synthetic images from reference images

23

  • J. Seo, S. Jeon and T. Jeon, “Domain Adaptive Generation of Aircraft on Satellite Imagery via Simulated and Unsupervised Learning”, MLAIP Workshop @ ACML (2017)
slide-24
SLIDE 24

Synthetic Data Generator and Refiner

Qualitative and Quantitative Evaluation

24

slide-25
SLIDE 25

Summary

  • Task: Detect and classify all aircraft on North Korea Airforce bases
  • Construct Own Dataset for civil aircraft and military fighters
  • Compatibility: Transfer Learning between GoogleEarth and KOMPSAT 2, 3, 3A
  • TensorRT: Speed-up to 2.2X (DenseNet) and 1.6X (VGG).
  • Fill the gap for rare observation: Combine synthetic data from GAN

25

slide-26
SLIDE 26

Road Extraction

03

slide-27
SLIDE 27
  • Automatic Mapping

from Image to Road

  • Usages

▪ Automated Map Update ▪ Urban Planning ▪ City Monitoring ▪ Road Navigation ▪ Operation of Unmanned Vehicles ▪ Attention of Safety Road

Road Extraction

slide-28
SLIDE 28

DeepGlobe Challenge (CVPR 2018)

slide-29
SLIDE 29

Challenges of Road Extraction

  • Wide-area Processing
  • Noisy Labeling & Ambiguity
  • Extraction of Road Network Topology
  • Model Efficiency
  • Intrinsic Noise of Road Image
slide-30
SLIDE 30

D-LinkNet: 1st Winner of the 2018 Challenge

slide-31
SLIDE 31

D-LinkNet: 1st Winner of the 2018 Challenge

slide-32
SLIDE 32

Our Motivation

Non-Local Operations

slide-33
SLIDE 33

Non-Local LinkNet (NL-LinkNet)

Non-Local Block (NLB)

Overall Architecture

slide-34
SLIDE 34

Quantitative Comparison

slide-35
SLIDE 35

Visual Comparison

slide-36
SLIDE 36

Model Efficiency

slide-37
SLIDE 37

Summary

  • Core Idea: Non-local Operations
  • Non-Local Block is better than traditional convolutional ops.
slide-38
SLIDE 38

Conclusions

04

slide-39
SLIDE 39

Conclusions

  • Object Detection and Classification: Use case for Defense
  • TensorRT: Speed-up to 2.2X (DenseNet) and 1.6X (VGG).
  • Fill the gap for rare observation: Combine synthetic data from GAN
  • Non-Local Block: Extraction of Road Network Topology
slide-40
SLIDE 40

Thank you for attention!

SI Analytics Co., Ltd. (Satrec Initiative Group) 441Expo-ro, Yuseong-gu, Daejeon, 34051, Korea tgjeon@si-analytics.ai www.si-analytics.ai