[PPT] - Jungho Im, PhD (ersgis@unist.ac.kr) School of Urban and PowerPoint Presentation

SLIDE 1

Jungho Im, PhD (ersgis@unist.ac.kr)

School of Urban and Environmental Engineering Ulsan National Institute of Science and Technology Ulsan, S. Korea November 7, 2019

SLIDE 2

Intelligent Remote sensing and geospatial Information Science

SLIDE 3

High spatio-temporal resolution Cloud contamination (optical sensor data) Point-based data Spatial discontinuity Reference data High temporal resolution Difficult to parameterize High uncertainty Uniform spatial and temporal coverage

SLIDE 4

Disaster Monitoring and Prediction Multi-source Data fusion

SLIDE 5

Image source: NVIDIA’s blog

SLIDE 6

Image source: Difference Between

Image by Théophile Gonos

SLIDE 7

Research Examples

01 Estimation of Fugacity of Carbon Dioxide (fCO2) 02 Prediction of Monthly Arctic Sea Ice Concentrations 03 Detection of Convective Initiation 04 Overshooting Tops Detection 05 Estimation of Tropical Cyclone Intensity 06 Estimation of Ground Particulate Matter Concentrations 07 Heatwave Monitoring and Prediction

SLIDE 8

Estimation of Fugacity of Carbon Dioxide (fCO2)

Jang, E., Im, J.*, Park, G., Park, Y. (2017). Estimation of fugacity of carbon dioxide in the East Sea using in situ measurements and Geostationary Ocean Color Imager satellite data. Remote Sensing, 9, 821

SLIDE 9

Source : NOAA PMEL Carbon Program

Ocean control the climate of Earth by regulating

the concentration of Carbon Dioxide

Ocean acidification by increasing carbon dioxide
Destroy the ocean ecosystem
Monitoring carbon dioxide is important to

determine ocean acidification and climate change

SLIDE 10

The East Sea of Korea
20,903 in situ samples
Dynamic phenomena occur
Active deep convection occurs because
f the deep water formation and weak

vertical stability

SLIDE 11

Machine Learning

Random Forest Support Vector Regression

Estimate Ocean 𝐠𝑫𝑷𝟑 Based on Satellite Data

GOCI, MODIS Chlorophyll CDOM Band Reflectance HYCOM Mixed layer depth Sea surface temperature Sea surface salinity In situ Fugacity of CO2 GOCI Chlorophyll CDOM Band Reflectance HYCOM Mixed layer depth Sea surface temperature Sea surface Salinity

Calculate Sea-Air 𝑫𝑷𝟑 Flux

SLIDE 12

SLIDE 13

2015 (RF model)

300 400 350

fCO2 (µatm)

Seasonal variability of surface

seawater fCO2 averaged by month in 2015 (RF model)

Similar spatial patterns with SST,

SSS, and MLD

fCO2 in summer might be affected

by an inflow of warm current from South

Low fCO2 in coastal areas appears

to be related to biological activity

SLIDE 14

Flux(mol m-2 yr-1)

1

5
2
Sea-air CO2 flux using estimated

surface seawater fCO2 based on the RF model

The East Sea absorbs CO2 from the

atmosphere throughout the whole region, acts as a sink for atmospheric CO2

The annual mean CO2 flux value

was -1.53 mol·m-2·year-1

The largest CO2 flux to the ocean

was estimated in winter and the lowest flux in summer

SLIDE 15

Prediction of monthly Arctic sea ice concentrations

Kim, Y., Kim, H., Han, D., Im, J.*, Lee, S. (2019). Prediction of monthly Arctic sea ice concentrations using satellite and reanalysis data based on convolutional neural network deep learning. The Cryosphere, in revision

SLIDE 16

Research objectives

✓ Development of a novel monthly SIC prediction model using a deep learning approach (CNN) ✓ Examination of the prediction performance by comparing Random Forests model ✓ Analysis of the sensitivity of variables affecting SICs

SLIDE 17

SLIDE 18

Variable description

✓ Eight predictors were used to predict the monthly SIC ✓ i.e., the SICs in September is affected by the SICs in August

Variable Source Unit Temporal resolution Spatial resolution Normaliza tion SIC one-year before (sic_1y) NSIDC, The Nimbus 7 SMMR and the DMSP SSM/I and SSMIS % Daily 25km 0 - 1 SIC one-month before (sic_1m) 0 - 1 SIC anomaly one-year before (ano_1y)

1 - 1

SIC anomaly one-month before (ano_1m)

1 - 1

Sea surface temperature one- month before (sst) NOAA OISST ver.2 K Daily 0.25° 0 - 1 2-meter air temperature one- month before (t2m) ECMWF ERA Interim K Monthly 0.125° 0 - 1 forecast albedo one-month before (fal) % 0 - 1 the amount of v-wind one-month before (v-wind) m/s 0 - 1

SLIDE 19

MAE RMSE nRMSE NSE All range of SICs (0-100%) Simple prediction 9.36% 21.93% 61.94% 0.83 RF 2.45% 6.61% 18.64% 0.96 CNN 2.28% 5.76% 16.15% 0.97 Low SICs (0-40%) Simple prediction 3.88% 11.96% 33.22% 0.60 RF 2.38% 7.23% 19.87% 0.90 CNN 2.13% 6.18% 16.87% 0.93

(a) mean absolute SIC anomaly (%) (b) MAE of simple prediction model (predicted-NSIDC%) (c) MAE of RF (RF- NSIDC, %) (d) MAE of CNN (CNN-NSIDC, %)

SLIDE 20

(a) SIC (NSIDC, %) (b) SIC (RF, %) (c) SIC (CNN, %) (d) SIC anomaly (NSIDC, %) (e) Error (RF-NSIDC, RMSE = 7.47%; nRMSE = 32.71%) (f) Error (CNN-NSIDC, RMSE = 5.00%; nRMSE = 21.93%)

SLIDE 21

Detection of Convective Initiation

Han, H., Lee, S., Im, J.*, Kim, M., Lee, M., Ahn, M., Chung, S. (2015). Detection of convective initiation using Meteorological Imager on board Communication, Ocean and Meteorological Satellite based on machine learning approaches. Remote Sensing, 7, 9184-9204 . Lee, S., Han, H., Im, J.*, Jang, E. (2017). Detection of deterministic and probabilistic convective initiation using Himawari-8 Advanced Himawari Imager data. Atmospheric Measurement Techniques, 10, 1859–1874 Han, D., Lee, J., Im, J.*, Sim, S., Lee, S., Han, H. (2019). A novel framework of detecting convective initiation combining automated sampling, machine learning, and repeated model tuning using geostationary satellite data. Remote Sensing, 11, 1454.

SLIDE 22

Numerical weather prediction (NWP) Weather radar

▪ High uncertainty for nowcasting ▪ Low spatial resolution ▪ Limited spatial coverage Heavy Rainfall & Severe Storm

SLIDE 23

Data and methods

GOES-R Interest fields

Himawari-8 AHI

Machine learning approach & logistic regression

Developing optimized interest field of CI over Korean peninsula

Machine learning approaches

Interest fields optimized for identifying CIs over Korean Peninsula

SLIDE 24

SLIDE 25

An example
CI prediction using RF : 20150801 0610-0800 (for 120 mins)
Radar CAPPI reference: 20150801 0810 UTC

CI

SLIDE 26

Overshooting Tops Detection

Kim, M., Im, J.*, Park, H., Park, S., Lee, M., Ahn, M. (2017). Detection of tropical overshooting cloud tops using Himawari-8 imagery. Remote Sensing, 9, 685. Kim, M., Lee, J., Im, J.* (2018). Deep learning-based monitoring of overshooting cloud tops from geostationary satellite data. GIScience and Remote Sensing, 55, 763-792.

SLIDE 27

Overshooting cloud tops (OT)
“A domelike protrusion above a cumulonimbus

anvil, representing the intrusion of an updraft through its equilibrium level” (AMS)

Cumulonimbus clouds with OTs can cause

severe weather conditions, influencing in-flight and ground aviation operations.

We try to develop a system to mimic the “human

being’s system” which is more intuitive and

generalizable. → “Convolutional Neural Network

(CNN)”.

Rather than identifying individual pixels as OTs or

non-OTs, human beings typically find OTs through visual inspection of contextual information of pixels (i.e. dome-like shape).

SLIDE 28

[Source: Kristopher M. Bedka at NASA Langley Research Center]

Time series infrared images of 13th May, 2015

Infrared

Overshooting cloud tops viewed by Himawari-8

Visible

SLIDE 29

Final CNN structure in this study

SLIDE 30

POD and FAR for whole images
Comparisons of model performances

Date POD FAR 0600 UTC on Sep. 25th 2015 90.59% 13.27% 0600 UTC on May 25th 2016 85.19% 8.89% 0600 UTC July 8th 2016 86.93% 31.77% Average 87.57% 17.98% Authors Technique POD FAR Bedka et al. (2010)+ Bedka and Khlopenkov (2016) IRW-texture candidates+ detection criteria 35.1% 24.9% Bedka and Khlopenkov (2016) OT probability 69.2% 18.4% OT probability with visible rating detection 51.4% 1.6% Kim et al. (2017) Random Forest 77.76% 31.73% This study CNN 79.68% 9.78%

POD (Probability Of

Detection) = # of OTs detected correctly /Total # of OT reference

FAR (False Alarm Rate)

= # of misclassified OTs /Total # of detected OTs

SLIDE 31

CNN RF CNN RF

SLIDE 32

Product: Overshooting tops (OT)
Spatial coverage: Northeast Asia
Product period : 10min

32

<지상국 산출 영상> <가시영상과 산출 결과>

SLIDE 33

Estimation of Tropical Cyclone Intensity

Lee, J., Im, J.*, Cha, D.*, Park, H., Sim, S. (2019). Tropical cyclone intensity estimation through multidimensional convolutional neural networks using geostationary satellite data, in review.

SLIDE 34

Estimation of Ground Particulate Matter Concentrations

Park, S., Shin, M., Im, J.* , Song, C., Choi, M., Kim, J., Lee, S., Park, R., Kim, J., Lee, D., Kim, S. (2019). Estimation of ground level particulate matter concentrations through the synergistic use of satellite

bservations and process-based models over South Korea, Atmospheric Chemistry and Physics, 19,

1097-1113. Park, S., Lee, J., Im, J.*, Song, C., Kim, J., Lee, S., Park, R., Kim, S., Yoon, J., Lee, D., Quackenbush, L. (2019). Estimation of spatially continuous particulate matter concentrations under all sky conditions through the synergistic use of satellite-based AOD and numerical models, in review.

SLIDE 35

mixture of solid particles and liquid droplets found in the air. It includes particulate matter with aerodynamic diameter less than 10 µm and 2.5 µm, respectively) Most particles form in the atmosphere as a result of complex reactions of chemicals such as sulfur dioxide and nitrogen oxides.

(Source: http://breathe.createareaction.com/learn// Particulate Matter. Basic Information. U.S. EPA)

SLIDE 36

Based on climate system

model and emission models

Atmospheric chemistry models

Various statistical regression

approaches, including simple and multiple linear models, geographically weighted regression, and population- weighted regression analysis

Difficult to simulate non-linear

relationships between variables

Difficult to model extreme

values

Statistical approaches

Good at modeling non-linear

relationships between variables

Examples include random

forest, extremely randomized trees, support vector machines, neural networks, and deep learning.

Hard to estimate for untrained

patterns

Overfitting

Artificial intelligence (Machine learning)

𝑸𝑵𝒕𝒃𝒖 = 𝑩𝑷𝑬𝒕𝒃𝒖 × 𝑸𝑵𝑯𝑭𝑷𝑻_𝒅𝒊𝒇𝒏 𝑩𝑷𝑬𝑯𝑭𝑷𝑻_𝒅𝒊𝒇𝒏

Distribution of estimated PM2.5 using the ratio of GEOS- Chem product and satellite-based AOD (Donkelaar et al.) Spatial distribution of MODIS AOD based PM2.5 conc. using geographically weighted regression (Bai et al.) Distribution of PM2.5 concentration based on MODIS AOD using ANN method (Gupta et al.)

SLIDE 37

SLIDE 38

SLIDE 39

SLIDE 40

SLIDE 41

GEOS-Chem & GOCI-based GEOS-Chem & RF model CMAQ & RF model PM10 PM2.5

SLIDE 42

✓ 버전 별 계절 별 분포 지도

연구기간: 2016.01 – 2016.12

SLIDE 43

Comparative assessment of various machine learning-based bias correction methods for numerical weather prediction model forecasts

f extreme air temperatures in urban areas

Cho, D., Yoo, C., Im, J.*, Cha, D.* (2019). Comparative assessment of various machine learning- based bias correction methods for numerical weather prediction model forecasts of extreme air temperatures in urban area, in re-review.

SLIDE 44

Comparison between convolutional neural networks and random forest for local climate zone classification in mega urban areas using Landsat images

Yoo, C., Han, D., Im, J., & Bechtel, B. (2019). Comparison between convolutional neural networks and random forest for local climate zone classification in mega urban areas using Landsat images. ISPRS Journal of Photogrammetry and Remote Sensing, 157, 155-170.

SLIDE 45