Artificial Intelligence and Data Science in Earth Observation
Transcript of Artificial Intelligence and Data Science in Earth Observation
![Page 1: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/1.jpg)
Artificial Intelligence and Data Science in Earth Observation
Xiaoxiang Zhu
Xiaoxiang Zhu
Remote Sensing Technology Institute (IMF), DLR
Signal Processing in Earth Observation (SiPEO), TUM
@xiaoxiang_zhu
![Page 2: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/2.jpg)
DLR’s Mission in Earth Observation
We research and develop solutions for major challenges in the following areas …
Meteorology Sustainable Development
Security
Global Change Research
Resource Management City Planning
Earth System Research and Environmental Sciences
Mobility
![Page 3: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/3.jpg)
The Golden Era of Big Earth Observation Data
– Sentinels and future national satellites provide
– continuous, reliable and quality controlled acquisition of big EO data
– free and open data
– long-term perspective
– Complementary NewSpace approaches, e.g. Planet
– Internet giants and Start-Ups (Descartes Lab, Orbital Insight,…) enter EO
Classical evaluation methods no longer sufficient AI4EO
But:
High EO quality requirements and wide application diversity call for EO-specific AI research and innovative AI4EO methods
2010 2012 2014 2016 2018 2020
50
40
30
20
10
PB
![Page 4: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/4.jpg)
Date Fusion
Data Mining
Machine Learning/Deep Learning
Big Data Management and HPC
Data Science and AI in Earth Observation
![Page 5: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/5.jpg)
AI4EO Deep Learning in Remote Sensing
Main Course – Two EO Data Science Stories
Geoscientific applications Global Urban Mapping
![Page 6: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/6.jpg)
Machine Learning/Deep Learning
Classical Neural Net mid 1980s
Deep Neural Net since 2006/2012
![Page 7: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/7.jpg)
We need more AI4EO
Papers related to Deep Learning in RS Web of Science
IEEE Geoscience and Remote Sensing Magazine, Dec. 2017
![Page 8: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/8.jpg)
Deep Learning in EO – Hot Topic or Hype?
– Phase 1: Quick wins and quick papers
“we can also do it with DL” “ e.g. 86.7 % 89.3 %”
– Phase 2: Understand that EO is different from internet image labelling
– Design new architectures for specific problems, and train from scratch
– Phase 3: Remember your EO expert knowledge and find how to integrate it into DL
– “Opening the black box”, “turn the black box gray” – Re-implant physics, Bayes and domain expertise into the learning process
![Page 9: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/9.jpg)
One of Our Phase 1 Successes
Spatiotemporal Scene Interpretation of Space Videos via Deep Neural Network and Tracklet Analysis
Workflow
Results
Data
“Spatiotemporal Scene Interpretation of Space Videos via Deep Neural Network and Tracklet Analysis”, L. Mou, X. Zhu
Winner of
Data Fusion Contest 2016
![Page 10: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/10.jpg)
What makes Deep Learning in Earth Observation Special?
– Classification and detection are only small fractions of EO problems
– Focus on retrieval of physical or bio-chemical variables High accuracy , traceability and reproducibility of results, Quality measures
– Decadal expert domain knowledge available
– Well-controlled data acquisition (radiometric, geometry, spectrometric, statistical, SNR,…)
– Data can be 5-dimensional (x-y-z-t-λ), complex-valued and multi-modal : SAR, Lidar, multi-/super-/hyperspectral, GIS, OSM, citizen science, social media,…
– Often: lack of sufficient training data
![Page 11: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/11.jpg)
Deep Learning@EOC
– Detection, segmentation and classification of buildings, ships, vehicles, persons...
– Classification of Land Use/Land Cover, Settlement Types and LCZs
– Change Detection and Time Series Analysis
– SAR/Optical Matching
– 2D/3D optical/SAR/PolSAR/LiDAR fusion
– Synthesizing optical images from SAR data and vice versa
– Sentinel-2 cloud removal
– IM2Height and IM2Building Footprint
– Fusion of EO and social media data (image and text)
– Solving non-linear inverse problems in atmospheric sensing
– Merging multi-decadal satellite data for climate studies
![Page 12: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/12.jpg)
Our Deep Nets Zoo
![Page 13: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/13.jpg)
Hyperspectral Image Analysis
![Page 14: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/14.jpg)
Unsupervised Spectral-Spatial Feature Learning via Deep Residual Conv-Deconv Net
Mou, Ghamisi, and Zhu, IEEE TGRS 56 (1), pp. 391-406, 2018.
![Page 15: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/15.jpg)
Unsupervised Spectral-Spatial Feature Learning via Deep Residual Conv-Deconv Net
University of Pavia, Italy
Application I: Classification Application II: “Free” Object Localization
• We fo
un
d so
me n
euro
ns in
ou
r netw
ork o
wn
go
od
d
escriptio
n p
ow
er for sem
antic visu
al pattern
s in th
e o
bject level. Fo
r examp
le, the n
euro
ns #
52 and
#03
can b
e used
to p
recisely captu
re me
tal sh
ee
ts (left) an
d ve
ge
tative
cove
rs (righ
t).
Mou, Ghamisi, and Zhu, IEEE TGRS 56 (1), pp. 391-406, 2018.
![Page 16: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/16.jpg)
Time Series Data Analysis
![Page 17: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/17.jpg)
Recurrent Convolutional Neural Network for Change Detection
Mou , Bruzzone, Zhu, IEEE TGRS 57 (2), pp. 924-935, 2019
![Page 18: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/18.jpg)
Recurrent Convolutional Neural Network for Change Detection
Location: Taizhou City, China
Legend: Changed areas (in binary change detection); city expansion; soil change; water change
Mou , Bruzzone, Zhu, IEEE TGRS 57 (2), pp. 924-935, 2019
![Page 19: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/19.jpg)
Example – Urban Growth of Beijing (1984 - 2016)
![Page 20: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/20.jpg)
Munich Airport
![Page 21: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/21.jpg)
Global Applications with Sentinels
![Page 22: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/22.jpg)
Global Cloud Cover – 67%
Image: ESA/Cloud-CCI
![Page 23: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/23.jpg)
cGAN for Removing Clouds from Sentinel-2 Data using Cloud-free Radar Data
Motivation: Optical sensors cannot penetrate clouds, but microwaves do.
Objective: Train generative adversarial network to produce cloud-free optical imagery
B3 B2 B4 B5 B6 B7 B8 B8A B11 B12
0.4 2.5 𝜇m
Sentinel-2 Sentinel-1
5.5 cm …
Cloud-free Sentinel-2 Cloudy Sentinel-2 Sentinel-1
+
Grohnfeldt, Schmitt, Zhu (2018), Proceeding of the ISPRS TC II Symposium 2018, Riva del Garda, Italy.
![Page 24: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/24.jpg)
High Resolution Remote Sensing Imagery Analysis
![Page 25: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/25.jpg)
Shi , Li, Zhu, IEEE GRSL, in press, 2019
![Page 26: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/26.jpg)
Multi-task CNNs for Car Instance Segmentation
Mou & Zhu, IEEE TGRS 56(11), pp. 6699-6711, 2018.
![Page 27: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/27.jpg)
Vehicle Detection and Tracking
Li et al., IEEE TGRS, in press, 2019
![Page 28: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/28.jpg)
Open Issues
– novel applications, other than classification and detection related tasks
– transferability of deep nets
– automated deep topology learning
– very limited annotated data in remote sensing
– how to benchmark the fast growing deep-learning algorithms in remote sensing?
– how to combine physics-based modeling and deep neural network?
– and many more…
![Page 29: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/29.jpg)
Munich School for Data Science @ Helmholtz, TUM & LMU (MuDS) Speakers: Fabian Theis (HMGU), Frank Jenko (IPP), Xiaoxiang Zhu (DLR)
Scale: 12M€, 38 Doctoral candidates
www.mu-ds.de open for application by Feb 28th 2019 @MunichDS
![Page 30: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/30.jpg)
![Page 31: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/31.jpg)
Global Urban Mapping
![Page 32: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/32.jpg)
Urban Planet
[UN, 2014]
![Page 33: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/33.jpg)
Urban Growth Happens Mostly in Developing Areas
Lagos, 21 Million Population
![Page 34: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/34.jpg)
So2Sat: Understanding Global Urbanization from Space to Social Networks www.so2sat.eu
So2Sat: 3D/4D urban models infrastructure type classification high resolution population density map
![Page 35: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/35.jpg)
Red Tile Roof
Hyperspectral Sensor
Social Media Images
Text Messages
Xiaoxiang Zhu @xiaoxiang_dlr
I’m in the rooftop bar on 10th floor. Last day in Rio de Janeiro! @Helmholtz @DLR_de
Radar Sensor
So2Sat in a Nutshell
![Page 36: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/36.jpg)
10 Petabytes = half of the archive at DFD
![Page 37: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/37.jpg)
TerraSAR-X/TanDEM-X
Global 3D/4D Urban Mapping
![Page 38: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/38.jpg)
![Page 39: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/39.jpg)
Radar Geometry in Range-Elevation Plane
r
s
![Page 40: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/40.jpg)
Radar Tomography – “X-Ray” of the Earth
r
s
![Page 41: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/41.jpg)
4D City@ LRZ Calculation for every single pixel = solving optimization problem with a matrix dimension of ca. 102× 106 since 2012, 26mio CPU hours granted
Why HPC?
![Page 42: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/42.jpg)
ca. 1 million Pts/km2, 4D Information
![Page 43: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/43.jpg)
[mm/y]
0 0.2 0.4 0.6 0.8 1 1.2 1.4-25
-20
-15
-10
-5
0
5
10
15
Dis
pla
cem
ent
[mm
]
Temporal Baseline [year]
![Page 44: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/44.jpg)
global?
![Page 45: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/45.jpg)
TanDEM-X has global coverage, but …
only moderate resolution
number of coverages limited
TanDEM-X for Global Coverage, But… medium resolution , small number of images
![Page 46: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/46.jpg)
Building heights
Building shapes
Signal Processing Algorithms
Deep Learning Algorithms
X-Ray of the Earth
![Page 47: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/47.jpg)
First Impression of the Global 3D Urban Models accuracy better than 2m
![Page 48: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/48.jpg)
settlement type? morphological structure first
settlement type? morphological structure first
![Page 49: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/49.jpg)
Global Local Climate Zones Classification will be global soon
1 7 C
1: Compact high-r ise
2: Compact mid-rise
3: Compact low-rise
4: Open high-r ise
5: Open mid-rise
6: Open low-rise
7: Lightweight low-rise
8: Large low-rise
9: Spar sely built
10: Heavy industry
A: Dense trees
B: Scattered trees
C: Bush, scr ub
D: Low plants
E: Bare r ock or paved
F: Bare soil or sand
G: Water
2 8 D
3 9 E
4 10 F
5 A G
6 B
![Page 50: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/50.jpg)
7. lightweight low-rise
![Page 51: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/51.jpg)
So2Sat LCZ42 Benchmark Dataset
– Hand labelled 42 cities covering 10 culture zones
– Data: – Sentinel-1 – Sentinel-2, seasonal
– 10 v0tes for each label
Labeling effort: 15 person × 1 Month/person
![Page 52: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/52.jpg)
![Page 53: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/53.jpg)
53
![Page 54: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/54.jpg)
tweeting for social good?
![Page 55: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/55.jpg)
Building Settlement Type Classification – by the Fusion of Remote Sensing and Social Media Text Messages
Text Messages
Xiaoxiang Zhu @xiaoxiang_dlr
I’m in the rooftop bar on 10th floor. Last day in Rio de Janeiro! @Helmholtz @DLR_de
![Page 56: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/56.jpg)
Tweets for Building Functions Identification re
sid
en
tia
l n
on
-re
sid
en
tia
l m
ixe
d u
sed
![Page 57: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/57.jpg)
Preliminary Results – OSM Ground Truth
residential
commercial
![Page 58: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/58.jpg)
residential
commercial
Preliminary Results – Tweets Predicted
![Page 59: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/59.jpg)
Building Settlement Type Classification – by the Fusion of Remote Sensing and Social Media Images
Social Media Images
![Page 60: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/60.jpg)
Building Instance Classification from Street View Data by CNN
apartment church garage house industrial office building retail roof Chicago Vancouver Munich
Kang et al. , ISPRS J. Photogramm. Remote Sens. 56(6), 3062-3077, 2018
![Page 61: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/61.jpg)
Flickr Random Search
– Queries Flickr API with random bounding boxes
– Up to 100,000 geotagged photos/day per bot
– Ca. 17.1 Mio geotagged Images
![Page 62: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/62.jpg)
Predict Settlement Types Using Social Media Images
Extract geospatial knowledge from social media images for building instance classification
![Page 63: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/63.jpg)
Our Vision in 2022 A first and unique global and consistent 3D/4D spatial data set on the urban morphology
![Page 64: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/64.jpg)
Mumbai
population density vs. time
3D model vs. time
Courtesy: H. Taubenböck, DFD
Dharavi
City boundary
Urban Footprint
![Page 65: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/65.jpg)
The So2Sat Data will be Open
– better understanding and boosting research on the global change process of urbanization
– unique data set for stakeholders such as the United Nations
– a helping hand to address poverty
So2Sat @ LRZ Demand: > 100 mio CPU hours; >10 PB Storage
![Page 66: Artificial Intelligence and Data Science in Earth Observation](https://reader031.fdocuments.us/reader031/viewer/2022012012/6156ffdda097e25c764fea4c/html5/thumbnails/66.jpg)
DLR/Alibaba AI4EO Challenge Global urban mapping So2Sat AI4EO research @DLR&TUM
Join us for AI4EO:
Contact: [email protected] @xiaoxiang_zhu