1 of 31

Gap filling in flood detection

feat. GEE and Python

Ian Davies

School of Environmental and Forest Sciences, UW

2 of 31

Motivation

I do a little bit of work with a small company called Cloud to Street.

Some things you probably already know - floods are the costliest and most common natural disaster worldwide. The most flood prone places are also the poorest, and many don’t have the capability in terms of local data or resources to assess flood risk and assess damaged areas during and directly after a flood. Traditional flood risk assessment is expensive and even in the US traditional surveying methods are reserved for highly populated areas, leaving many at-risk rural areas undersurveyed.

We analyze flood risk and flood event data using remote, low-cost methods to make timely information available to governments and aid organizations.

The work I’m concerned with falls into two categories - flood detection of historical imagery to build our database of flood recurrence. This is important for assessing flood risk - areas that have high flood recurrence are more likely to flood again.

We also do rapid and soon NRT flood mapping which is where a flood occurs, and we provide flood maps ASAP from whatever sensors are available.

So inexpensive, rapid, and accurate flood mapping is important.

3 of 31

Flood detection

Ideal:

Very few clouds
No haze
Image (L8) taken 1 day after peak precip

4 of 31

Flood detection

Less ideal:

Clouds and shadows obscure possible flooded areas
Image taken 5 days after peak precip, possibly capturing <max extent

5 of 31

Flood detection

Much less ideal:

Clouds and shadows obscure possible flooded areas

6 of 31

7 of 31

Radar

SAR (microwaves) can penetrate clouds �
But there’s really just one public SAR satellite system (Sentinel-1A/B), combined repeat time of ~6 days�
Many more optical sensors at our disposal (Landsat 7/8, Sentinel-2, MODIS)

8 of 31

Sentinel-1

9 of 31

Landsat 8

10 of 31

Question: What can we infer from cloud-covered floods (and with what certainty) when all we have is optical sensors?

11 of 31

Idea: Use auxiliary data

Utilize all available data to try and predict whether cloud-covered pixels are inundated or not

Spectral data to detect floods where we can
Topography/land cover data to estimate Pr(Flooded) where we can’t

12 of 31

Idea: Use auxiliary data

13 of 31

Goals

Rule out flooding

(High specificity)

Accurately predict flooding (High recall)

Really two goals here:

Low hanging fruit: rule out the possibility of flooding beneath clouds. If we can eliminate areas that are unlikely to be flooded because they are sloped or heavily vegetated, then that reduces the uncertainty in our flood maps. This is more feasible and would amount to, basically, a context-dependent flood risk map.

Areas can have characteristics that prevent them from getting flooded (or make it very unlikely).

Trickier is the fact that areas can have characteristics that make them more prone to flooding, but that doesn’t mean they are flooded. A flood risk map doesn’t mean that every pixel with high risk is always flooded all the time. There has to be some trigger.

So the trick of this classifier is that it has to do more than reproduce flood-risk. It has to be context dependent.

High hanging fruit: Accurately predict where there is flooding. High true positive rate. This is the ultimate goal and what I’m investigating. But may not be possible due to, among other things, lack of high-resolution data.

14 of 31

Workflow (in progress!)

Use cloudfree imagery and aux datasets to train a classifier
Test on “cloudy” images with mapped floods

15 of 31

Workflow

Find known flood events

Find imagery that overlaps flood event date/area

Calculate auxiliary features underneath imagery

Generate fake cloud cover, mask images

Sample pixels in clear imagery

Train classifier on sample pixels

Test on “cloudy” pixels

GEE

Python

16 of 31

Find flood events from DFO

Dartmouth Flood Observatory
Flood events mapped from space, mostly MODIS
Google Fusion Table contains flood date and (rough) bounds
For now, looking at just US floods for data availability

Event 4591, �April 4-11, 2018

17 of 31

Find clear imagery of flood events

Not interested in clouds yet! Only clear images <10% cloud cover�
Just L8 imagery for now

18 of 31

Calculate auxiliary features

Land cover (NLCD)
Distance from perm water
DEM-derived features

TWI, SPI, slope, curve, HAND, ...

19 of 31

Map floods and mask out any clouds, shadows, snow

Flood detection done with Automated Water Extraction Index (AWEI) (Feyisa et al. 2014)
Fewer urban misclassifications than MNDWI, thresholded at zero

20 of 31

Map floods and mask out any clouds, shadows, snow

From Feyisa et al. (2014)

21 of 31

Stack, generate cloud cover, and mask

Clear pixels

Cloudy pixels

Slope

Aspect

Distance from river

Etc ...

Clear flood event image

Cloudy image over same area

+

22 of 31

Sample from clear image

23 of 31

Export to Python, clean, convert to sparse matrix

24 of 31

Export to Python, clean, convert to sparse matrix

25 of 31

Export to Python, clean, convert to sparse matrix

26 of 31

Train exhaustively or extensively?

Two possibilities for training:�
Train on 1000s of points from one image to predict cloudy floods in that same image

27 of 31

Train exhaustively or extensively?

Two possibilities for training:�
Train on 1000s of points from one image to predict cloudy floods in that same image�
Train on 1000s of points from many different images/flood events to predict cloudy floods in any image

28 of 31

SVM

First attempt using SVM (linear and with kernel)
Not amazing
Better at predicting dry areas than predicting floods - low hanging fruit

29 of 31

Lots of bugs still, but how much can SVM improve?

30 of 31

TensorFlow CNNs

Seems very useful but I am not sure what sort of CNN would be appropriate - segmentation?

31 of 31

Thanks!