1 of 13

1

Mapping Urban floods using multi-sensor satellite imagery and deep learning

Rohit Mukherjee

Post Doctoral Researcher

rohitmukherjee.space

rohitmukherjee@arizona.edu

@RohitMukherjee9 (Twitter)

+ social [pixel] lab with Dr. Beth Tellman

Dhaka, Bangladesh 25th August 2020

Hannah Friedrich, Zhijie Zhang, Jonathan Giezendanner

BEST FOR You�ORGANICS COMPANY

2 of 13

Urban Floods: More challenging

Complex Infrastructure, Drainage Systems, Impervious Surfaces

*Existing flood models are not as accurate in urban regions

PlanetScope 3 meters

2020-09-26, pre-flood

PlanetScope 3 meters

2020-11-22, post flood

Sentinel-2 10 meters

2020-11-27, post flood

MODIS 250 meters

2020-11-24, flood

San Pedro Sula, Honduras

BEST FOR You�ORGANICS COMPANY

3 of 13

Satellite Data: No single product is sufficient

Trade-offs in spatial, spectral, and temporal resolutions, optical/radar

No clouds, but distortions and speckle noise

Sentinel-1, VV-VH, 2020-11-25

Multispectral, but occluded by clouds

Sentinel-2, FCC, 2020-11-27

High resolution, but commercial

PlanetScope, FCC, 2020-11-25

Khartoum, Sudan

BEST FOR You�ORGANICS COMPANY

4 of 13

We covered 14 urban flood events

250+ 1024 by 1024 images labeled using PlanetScope imagery

Over 250 sq km of flooded regions labeled

Urban Flood Dataset

~1 hour total per chip (50 min labeling, 10 min reviewing)

Total of ~350 hours spent

Acknowledgements: Patrick Hellmann, Simone Holladay, Aaron Krupp, Lin Ji,

and Natasha Rapp

BEST FOR You�ORGANICS COMPANY

5 of 13

Hand-Labels are not ground truth

Labelers do not always agree with each other

It’s difficult to define high vs low confidence classes

Hand-labeled datasets are valuable (also expensive) especially for independent evaluations

Higher quality labels are more important than larger deep learning architectures architectures

PlanetScope NIR, Red, Green at 3 meters

BEST FOR You�ORGANICS COMPANY

6 of 13

PlanetScope NIR, Red, Green at 3 meters

Hand Labels

Inference on PlanetScope at

3 meters

Inference on

Sentinel-1 at 10 meters

Floods in Khartoum (Sudan) on the 25th of August, 2020

Trained U-Nets on Urban Flood Dataset

Sentinel-1 missing finer features

BEST FOR You�ORGANICS COMPANY

7 of 13

Evaluated our PlanetScope model for spatial generalizability

PlanetScope NIR, Red,

Green at 3 meters

Urban Flood Dataset

Predictions on unseen

Dhaka Floods in 2020

Mean F1

BEST FOR You�ORGANICS COMPANY

8 of 13

Evaluated across public and commercial sensors

Sentinel-2

Sentinel-1

PlanetScope

Urban Flood Dataset

Bing Imagery

Khartoum, Sudan

    • Which accuracy metrics are more suited?

    • Is it better to underpredict or overpredict?

    • Does 90% accuracy using 3 meter PlanetScope mean the same thing as a 90% using 10 meter Sentinel-2?

Z. Zhang

BEST FOR You�ORGANICS COMPANY

9 of 13

Higher spatial resolution (e.g., Planet) = more accuracy, but

PlanetScope

MODIS

Bing Imagery

  • Sentinel-2’s short wave infrared bands are extremely effective in detecting water
  • MODIS/VIIRS is too coarse, but daily
  • Sentinel-1 is cloud-free, but distortions

Can Tho, Vietnam

10th October 2020

BEST FOR You�ORGANICS COMPANY

10 of 13

PlanetScope FCC

Urban Flood Dataset Label

Trained model

Trained on larger FloodPlanet

Label data quality is more important than training dataset size*

*pre-train with lower quality labels and then fine-tune with stronger labels

When trained on non-urban focused and larger FloodPlanet dataset, accuracy did not improve

FloodPlanet contains 19 non-urban focused global flood events with temporally collocated with (at least one of) Sentinel-1, Sentinel-2, and Landsat 8 images for each label.

BEST FOR You�ORGANICS COMPANY

11 of 13

Differences in mean surface reflectance values per spectral band

across three PlanetScope’s SuperDove sensors

that captured images within 45 minutes of each other

Mapping Floods in the desert: A golf course in Pima County, Arizona

Frazier, A. E., & Hemingway, B. L. (2021). A technical review of planet smallsat data: Practical considerations for processing and using planetscope imagery. Remote Sensing13(19), 3930.

BEST FOR You�ORGANICS COMPANY

12 of 13

What we learned

Reliable flood monitoring requires multi-source cloud free high spatial and temporal data

We need application specific robust validation strategies

Higher spatial resolution data = better accuracy, but short wave infrared bands are crucial

Commercial data will not solve all problems, but deep learning could mitigate some

Flooding due to dam failure

May 2020

Midlands, MI

Bing Imagery

PlanetScope

BEST FOR You�ORGANICS COMPANY

13 of 13

PlanetScope Model

Thank you for your attention!

Rohit Mukherjee

rohitmukherjee@arizona.edu

rohitmukherjee.space

@RohitMukherjee9 (Twitter)

Hannah Friedrich, Jonathan Giezendanner, Zhijie Zhang, Ruixue Wang, Elise Arellano-Thompson, Alex Saunders, Lucas Belury, Prashanti Sharma, Ariful Islam, Beth Tellman

This work is funded by NASA THP (Award #80NSSC21K1341).

BEST FOR You�ORGANICS COMPANY