Rapid and robust monitoring of flood events using Sentinel-1 and Landsat data on the Google Earth Engine

Rapid and robust monitoring of flood events using Sentinel 1 and Landsat data on the Google Earth Engine Contents lists available at ScienceDirect Remote Sensing of Environment journal homepage www el[.]

Contents lists available atScienceDirect Remote Sensing of Environment journal homepage:www.elsevier.com/locate/rse

Rapid and robust monitoring of flood events using Sentinel-1 and Landsat

data on the Google Earth Engine

Ben DeVriesa,b,⁎, Chengquan Huangb, John Armstonb, Wenli Huangb,c, John W Jonesd,

aDepartment of Geography, Environment and Geomatics, University of Guelph, Guelph, ON, Canada

bDepartment of Geographical Sciences, University of Maryland, College Park, MD, USA

cSchool of Resource and Environmental Science, Wuhan University, Wuhan, China

dU.S Geological Survey, Hydrologic Remote Sensing Branch, Reston, VA, USA

eU.S Fish and Wildlife Service, National Wetlands Inventory, Falls Church, VA, USA

A R T I C L E I N F O

Edited by: Jing M Chen

Keywords:

Sentinel-1

SAR

Flood disasters

Cloud computing

A B S T R A C T Synthetic aperture radar (SAR) sensors represent an indispensable data source for flood disaster planners and responders, given their ability to image the Earth's surface nearly independently of weather conditions and time

of day The decision by the European Space Agency (ESA) Copernicus program to open data from its Sentinel-1 SAR satellites to the public marks the first time global, operational SAR data have been made freely available Combined with the emergence of cloud computing platforms like the Google Earth Engine (GEE), this devel-opment presents a tremendous opportunity to the disaster response community, for whom rapid access to analysis-ready data is needed to inform effective flood disaster response interventions and management plans Here, we present an algorithm that exploits all available Sentinel-1 SAR images in combination with historical Landsat and other auxiliary data sources hosted on the GEE to rapidly map surface inundation during flood events Our algorithm relies on multi-temporal SAR statistics to identify unexpected floods in near real-time Additionally, historical Landsat-based surface water class probabilities are used to distinguish unexpected floods from permanent or seasonally occurring surface water We assessed our algorithm over three recent flood events using coincident very high- spatial resolution imagery and operational flood maps Using very high resolution optical imagery, we estimated an area-normalized accuracy of 89.8 ± 2.8% (95% c.i.) over Houston, Texas following Hurricane Harvey in late August 2017, representing an improvement of between 1.6% and 9.8% over flood maps derived from a simple backscatter threshold Additionally, comparison of our results with SAR-derived Copernicus Emergency Management Service (EMS) maps following devastating floods in Thessaly, Greece and Eastern Madagascar in January and March 2018, respectively, yielded overall agreement rates of 98.5% in both cases Importantly, our algorithm was able to ingest hundreds of SAR and optical images served on the GEE to produce flood maps over affected areas within minutes, circumventing the need for time-consuming data download and pre-processing steps The flexibility of our algorithm will allow for the rapid processing of future open-access SAR data, including data from future Sentinel-1 missions

1 Introduction

Floods are among the costliest of natural disasters and inflict loss of

life and property on millions of people worldwide While climate

change is increasing flood risk in already vulnerable areas (Hirabayashi

et al., 2013), commonly used flood risk models have been shown to

significantly under-predict flood risk (Wing et al., 2018), underscoring

the need for observation-driven flood monitoring methods

Satellite-based Earth Observation (EO) data provide synoptic, repeated views of

potentially flooded regions, and are increasingly used in operational disaster monitoring systems (Voigt et al., 2016)

Two recent developments in the EO sector have the potential to significantly improve the efficacy of flood monitoring systems across the globe First, the opening of data from operational EO satellites, such

as Landsat, has enabled land change monitoring at relatively high spatial and temporal resolutions (Wulder et al., 2012) The European Sentinel-1A and e1B satellites comprise the first ever global, opera-tional synthetic aperture radar (SAR) mission whose data are open to

https://doi.org/10.1016/j.rse.2020.111664

Received 2 October 2018; Received in revised form 4 August 2019; Accepted 14 January 2020

⁎Corresponding author at: Department of Geography, Environment and Geomatics, University of Guelph, Guelph, ON, Canada

E-mail address:bdv@uoguelph.ca(B DeVries)

Available online 31 January 2020

0034-4257/ © 2020 Elsevier Inc All rights reserved

T

the global public Launched by the European Space Agency (ESA)

through its Copernicus program in March 2014 and 2016, respectively,

the Sentinel-1 satellites provide nominal 6-day repeat imagery over

Europe and nominal 12-day repeat imagery over the rest of Earth's

terrestrial surfaces, enabling regular monitoring of change processes

(Torres et al., 2012) Given the limitations of optical data for

high-re-visit monitoring of floods, expanded access to operational SAR EO data

is an important development for disaster monitoring However, despite

the continued availability of SAR data through the Sentinel-1 satellites

and the free, open-source pre-processing software distributed by ESA,

SAR pre-processing still remains a technically challenging and

compu-tationally intensive task

Second, the advent of cloud computing architectures is leading to a

shift in the way in which EO data are processed, allowing users and

developers to access entire EO data archives and circumventing the

need to download and locally store huge volumes of data By leveraging

the advantages of high performance computing (HPC) and “data cube”

architecture, EO cloud computing services allow for accelerated access

to large volumes of EO data, reducing the burden of data pre-processing

and formatting traditionally borne by users (Lewis et al., 2016) The

Google Earth Engine (GEE), a novel computing platform introduced by

Google, Inc., has enabled the development of global-scale data products

using satellite image time series, such as that from the Landsat archive

(Gorelick et al., 2017) The GEE has been used to conduct global and

regional scale investigations of land surface dynamics, including forest

cover (Hansen et al., 2013; Johansen et al., 2015), surface water

(Donchyts et al., 2016;Pekel et al., 2016;Tang et al., 2016), populated

areas (Patela et al., 2015), cropland and soils (Padarian et al., 2015;

Xiong et al., 2017) and other applications (Dong et al., 2016;Joshi

et al., 2016;Lee et al., 2016) Among its large store of geospatial

da-tasets, the GEE houses a complete and continually updated archive of

Sentinel-1 Ground Range Detected (GRD) data The provision of

ana-lysis-ready SAR data on the GEE represents a significant step forward in

applied SAR remote sensing, as the complexity of SAR preprocessing

has previously presented a barrier to its adoption Additionally, the

presence of other data sources, such as the entire global Landsat

ar-chive, allow for relatively easy integration of diverse EO data sources

The ability of SAR sensors to detect floods relies in large part on the

distinct scattering mechanisms exhibited at open water surfaces SAR

sensors transmit microwave energy to the Earth's surface at off-nadir

angles, leading to specular reflection – the near-complete reflection of

transmitted energy away from the sensor – from smooth, open water

surfaces This reflection results in very low backscatter, or signal

re-ceived by the sensor Many SAR-based flood detection algorithms

ex-ploit this scattering mechanism by applying backscatter thresholds to

SAR images to classify water pixels (Chini et al., 2017;Matgen et al.,

2011;Pulvirenti et al., 2011;Twele et al., 2016), often in concert with

object-based detection methods (Giustarini et al., 2013;Martinis et al.,

2009, 2015;Mason et al., 2014) Capillary waves, often present across

large water surfaces in the presence of high winds, give rise to Bragg

scattering, presenting a challenge for simple threshold methods,

espe-cially when vertically polarized transmitted and received energy is used

(Brisco, 2015) Change detection approaches using multi-temporal SAR

imagery are often used to identify flooded pixels compared to baseline

conditions (Amitrano et al., 2018;Badji and Dautrebande, 1997;Cian

et al., 2018;Clement et al., 2017;Hostache et al., 2012;Long et al.,

2014; Lu et al., 2015; Schlaffer et al., 2017) Interferometric or

co-herence-based change detection methods exploit both the amplitude

(intensity) and phase components of the received microwave energy,

allowing for a more detailed characterization of the complex SAR

backscatter signatures that typically result from wind-affected open

water surfaces or flooding among standing vegetation or built-up areas

(Geudtner et al., 1996;Plank et al., 2017;Refice et al., 2014)

While SAR images can provide reasonably reliable estimates of large

areal flood extents, challenges arising from limited spatial and temporal

development of approaches involving the integration of SAR observa-tions into models or other types of data streams For example, assim-ilation of SAR data into hydraulic models has been shown to improve flood extent estimates on a near real-time basis (Giustarini et al., 2011) and in the presence of built-up areas (Mason et al., 2014) SAR data are also frequently used in combination with other remotely sensed ob-servations to enhance flood inundation predictions Topographic data derived from digital elevation models (DEMs) are routinely used to constrain SAR-based flood estimates to areas likely to experience flooding (Brakenridge et al., 1994, 1998;Huang et al., 2017) A number

of studies have also demonstrated the potential of combining SAR data with optical data or other a priori surface water datasets to allow for multi-scale flood assessment (Martinis et al., 2013), automated training data selection (Huang et al., 2018;Westerhoff et al., 2013) or an in-crease in observation density during flooding events (Chaouch et al.,

2012)

Despite the range of methods developed for monitoring floods with SAR data, few studies have employed dense time series stacks of SAR or optical imagery towards this goal, due to challenges in accessing and processing such large volumes of data Open-access data policies like those of the NASA-USGS and Copernicus programs, and cloud-com-puting platforms like the GEE are now poised to facilitate the ex-ploration of methods that exploit entire archives of multiple satellite missions, with the aim of improving observation-based flood mon-itoring systems In this paper, we describe a new method for monmon-itoring floods using dense time series of Sentinel-1 SAR and Landsat data on the

GEE Our algorithm uses temporal SAR Z-scores, computed against a

baseline during which no flooding is assumed to have occurred In describing our algorithm, we pose the following questions:

1 What influence does flooding have on Z-scores derived from vertical

transmit vertical receive (VV) and vertical transmit horizontal re-ceive (VH) SAR backscatter time series?

2 How accurate are flood predictions using Sentinel-1 derived

Z-scores, and how well do they compare with existing operational flood maps?

We defined floodwater in this study as new water appearing above the ground surface or vegetation canopy (if present) in comparison to a historical reference period The algorithm described in the following section was accordingly designed to detect these instances of “un-expected” floods

2 Site and event descriptions

In this study, we demonstrated our algorithm over a range of flood events across the globe in the past two years and focused our evaluation

of the method over three of these events: (1) Hurricane Harvey in August 2017 over Houston, Texas; (2) Floods in March 2018 in Thessaly, Central Greece; and (3) Cyclone Ava in January 2018 over the east coast of Madagascar We selected these sites to demonstrate the utility of Sentinel-1 time series data over a range of acquisition stra-tegies and consequent data densities (Fig 1) Operational 12-day revisit imagery were acquired over Eastern Madagascar, while the 6-day re-visit imagery programmed for Europe were available over Thessaly, Greece Houston, Texas represents a special case where imagery ac-quired under an additional observation mode (described inSection 3.1) were available Each of these sites and events are briefly described below and shown inFig 1

2.1 Houston, Texas: Hurricane Harvey

The 2017 hurricane season along the southeastern coast of the United States, including Puerto Rico, is being recognized as the most costly in the country's history, with thousands of deaths during the

Halverson, 2018) Hurricane Harvey, which made landfall on the coast

of Texas on 2017-08-23, brought with it unprecedented levels of

flooding to the city of Houston and surrounding areas and resulted in an

estimated 70 deaths (Jonkman et al., 2018) This paper focusses on the

area to the west of Houston, which is dominated by cropland

inter-spersed with forests, woody wetlands and coastal emergent wetlands

2.2 Thessaly, Greece: 2018 spring floods

The Thessaly region in Central Greece is traversed by the Pinios

River, which is the third longest river in Greece and drains an area of

10,700 km2northwards into the Aegean Sea (Migiros et al., 2011)

Between 1880 and 2010, 16 to 25 major flood events were recorded in

the Thessaly region, one of which in 1907 was the most severe flood

recorded in Greece during this period (Diakakis et al., 2012) Intense

rainfall between the 21st and 26th of February, 2018, caused

cata-strophic flooding in the region, particularly throughout the agricultural

plains between the cities of Larissa and Trikala, with mostly cropland

and several villages in the region being heavily impacted (Davies,

2018)

2.3 Eastern Madagascar: Cyclone Ava

The southwest Indian Ocean is a hotspot for tropical cyclones,

which are the most significant natural hazard for Madagascar

(Ganzhorn, 1995) Historical records suggest that the number of

tro-pical cyclones making landfall in Madagascar has increased from the

19th century to more recent years (Nash et al., 2015), with some

ana-lyses suggesting a link to increases in sea surface temperature (Mavume

et al., 2009) Cyclone Ava made landfall on the northeast coast of

Madagascar on the 3rd of January, 2018 and affected an estimated

123,000 people over the following five days as it moved southwards

along the eastern coast (ReliefWeb, 2018) The affected regions

de-scribed in this study are dominated by an agricultural mosaic land

cover, comprised of a patchwork of forests, woodlands and croplands

3 Data and methods

3.1 Sentinel-1 data

Sentinel-1A, launched in April 2014, and Sentinel-1B, launched in

April 2016, are the first among a series of Earth imaging satellite

constellations operated under the ESA Copernicus program Data from the Sentinel-1 satellites are operationally acquired in four imaging modes: interferometric swath (IW), strip map (SM) extra wide-swath (EW) and wave (WV) mode, each with different acquisition configurations An overview of these imaging modes and their asso-ciated specifications is given inTorres et al (2012), and a brief sum-mary of the two modes used in this study (IW and SM) is given here Since all SAR sensors are side-viewing instruments, the imaged land surface is illuminated over a range of view and incidence angles Even though methods exist to correct for the effects of incidence angle and terrain (Small, 2011), radiometric variations due to local incidence angle may still persist after pre-processing IW imagery are acquired at incidence angles between 31° and 46° and SM imagery may be acquired

at incidence angles ranging from 20° to 47° (Torres et al., 2012) While

IW images are consistently acquired across swaths spanning these an-gles, SM imagery span narrower swaths, and the precise range varies from site to site Each of the two Sentinel-1 satellites orbits the globe once every 12 days, allowing for a joint potential 6-day repeat fre-quency over the equator and a 3-day revisit frefre-quency when both as-cending and desas-cending orbits are considered Operational IW imagery are acquired at nominal 6-day intervals over Europe and nominal 12-day intervals over the rest of the Earth's surface, with higher repeat frequencies at higher latitudes and areas where targeted acquisitions are planned

The GEE hosts Sentinel-1 Ground Range Detected (GRD) data ac-quired in EW, IW and SM modes, pre-processed using tools available through the ESA Sentinels Application Platform (SNAP) software package First, restituted orbit files were applied to the GRD images, resulting in geometric accuracies within 10 cm (Prats-Iraola et al.,

2015) GRD and thermal noise removal, which masks artificially low backscatter pixels found at the edge of the image swath, was then carried out for all imagery acquired after 2018-01-12 (Ali et al., 2018)

We applied additional extreme value thresholds to VV and VH back-scatter (-35 dB and -40 dB, respectively) for images included in the historical baseline period (described in the following section), to further reduce the impact of these border regions Radiometric calibration was

then carried out to produce the unitless backscatter intensity (σ int) (Sabel et al., 2012) The images were then terrain geo-coded using a digital elevation model (DEM) from the Shuttle Radar Topography Mission (SRTM) or the ASTER Global DEM for latitudes above 60° North

or South Finally, backscatter intensity was converted to backscatter

coefficient (σ0) measured in decibels (dB) according to Eq.(1):

Fig 1 Overview of validation sites described in this study, with colours representing the number of Sentinel-1 observations available on the GEE for all of 2017 and

validation regions shown as hatched polygons Left: Houston, Texas Centre: Thessaly, Greece Right: Eastern Madagascar Note that different value ranges are used for each site

= 10log int

All imagery were projected to WGS84 latitude/longitude projection

(EPSG:4326) during terrain correction and export of image products

3.2 Temporal SAR statistics

Our approach to monitoring floods using temporal anomaly

mea-sures consisted of several stages, demonstrated inFig 2 andFig 3

First, a temporal subset of all SAR backscatter data was selected for

each pixel based on a historical baseline for each flood event This

baseline was first defined using a priori knowledge of surface water dynamics at each respective site For example, inspection of several SAR pixel time series suggested that no floods had occurred between 2017-06-01 and 2017-08-15 in Houston, so these dates were chosen as the bounds of the historical baseline for that site All Sentinel-1 images acquired between those dates were then used to compute the temporal mean backscatter coefficient (o) (Fig 3B) and standard deviation

backscatter coefficient (std(σ o)) (Fig 3C) These statistics were com-puted separately for polarization modes (VV and VH), as well as for subsets of the baseline image stack based on orbital direction (as-cending and des(as-cending) and acquisition mode (IW and SM) The

Fig 2 Time series VV backscatter (σ0VV ), VV backscatter anomaly (∆σ0VV ) and VV Z-score (Z VV) for a single pixel located southwest of Houston, Texas Data markers

correspond to the acquisition modes (IW or SM) and orbital directions (ascending: ASC, or descending: DSC) available for this pixel Zero anomaly and Z-scores are

shown as a dotted grey line in the middle and bottom panels and an example flooding Z-score threshold of −2.5 is shown as a broken green line in the bottom panel (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig 3 Demonstration of the summary statistics computed for a flood event in Kale, Myanmar on 2017-09-19 A: VV backscatter coefficient image from 2017-09-19.

B: Mean VV backscatter coefficient for the baseline period between 2017-04-01 and 2017-09-01 C: Standard deviation VV backscatter coefficient for the same

baseline period D: VV Z-score corresponding to the image in (A).

backscatter anomaly (∆σ0) and Z-score (Z) were then computed for each

observation acquired at time t under polarization mode p, sensor

ac-quisition mode m, and orbital direction d according to Eqs.(2) and (3),

respectively:

=

=

std

p m d

3.3 Landsat data and historical inundation

The temporal SAR statistics described above provide an objective

baseline from which to measure backscatter anomalies that can be

re-lated to “new” floods within a given time period However, it is also

important to determine whether these floods may occur as part of a

regular intra-annual surface water regime (e.g., seasonal wetlands and

other water bodies) or are unexpected and potentially catastrophic To

aid in contextualizing the SAR anomaly measures, we used historical

Landsat data available on the GEE to map previously permanently or

seasonally inundated pixels using two different data products or

algo-rithms We relied on Landsat due to the fact that no SAR data acquired

before 2016 are available on the GEE First, we accessed the

Landsat-based Monthly Water History dataset produced by the Joint Research

Centre (JRC) of the European Commission (Pekel et al., 2016) and

computed water occurrence probability from all available images on

the GEE Second, we applied the Dynamic Surface Water Extent (DSWE)

algorithm described inJones (2015)to all available Landsat-5 TM and

Landsat-7 ETM+ images acquired between 2000 and 2015 on the GEE

to expand the range of water body types included in our algorithm

DSWE employs a series of thresholds and decision rules to identify dry

land, open water and partial water (surface water mixed with

vegeta-tion and/or soil targets), from each individual cloud-masked Landsat

image Landsat-8 OLI data were excluded from this step of the analysis

due to spectral and radiometric adjustments on the OLI sensor that

could affect the performance of the original DSWE thresholds (Vermote

et al., 2016) We computed the historical probability of open water from both the JRC and DSWE datasets as the proportion of all un-masked pixels classified as open water We additionally computed the probability of partial water from the DSWE-classified images These class probabilities were used to classify water pixels based on their historical occurrence, as described in the following section

3.4 Flood classification

Our classification scheme is shown inFig 4 First, we identified pixels with permanent open water (POW), defined as having a prob-ability of open water > 95% in either the JRC or DSWE products We

then applied Z-score thresholds on both the VV and VH Z-score images.

If both the VV and VH Z-scores were lower than these thresholds for a given pixel, that pixel was assigned a high-confidence flood label If

only one of the two polarizations resulted in a Z-score lower than the

thresholds, the pixel was assigned a moderate-confidence flood label Although VV backscatter is generally used alone to detect floodwater,

we included both polarizations to provide an additional line of evidence for flooding For example, Bragg scattering of VV energy over rough water surfaces may mask the presence of water, in which case a ne-gative VH Z-score would still provide evidence for flooding We tested a combination of Z-scores for each polarization over each study site to assess their effects on classification accuracy Finally, all pixels not classified as POW but with a historical inundation probability > 25%, where inundation was defined as having either an open water or partial water DSWE class label, were assigned a class label modifier describing the likelihood of prior inundation In summary, the class labels em-ployed in this study can be represented as a 2-D matrix (Fig 4), in which SAR-based flood confidence forms one axis and Landsat-based historical inundation information forms the other

3.5 Validation

We validated our flood predictions over the study sites described in Section 2using two different datasets First, we performed an accuracy assessment on randomly drawn samples of flooded and non-flooded

Fig 4 Classification tree and colour key indicating floodwater classes included in this study The open water probability, P(ow), and total inundation probability, P

(in), were derived from historical Landsat data Permanent open water (POW) was assigned to all pixels with a P(ow) > 95% Other flood classes were derived from

VV and VH Z-scores (ZVVand ZVH, respectively) with an applied threshold (thdVVand thdVH, respectively)

pixels over the Houston site, using 3-m PlanetScope imagery (Planet

Team, 2017) acquired on 2017-08-29 as reference data We stratified

flood maps produced on the GEE from Sentinel-1 data acquired on the

same date by SAR flood classes Since our objective with the accuracy

assessment was to validate the SAR-based flood labels and not the

previously published Landsat-based algorithms, we excluded all pixels

with historical inundation class labels We rejected any samples that

included mixed water/non-water pixels when overlaid onto Planet data,

ensuring that only pure water or land pixels were included in the

va-lidation After removing seven samples where visual interpretation was

not possible or extreme Z-score values were found, a total of 493

samples taken from predicted 2017-08-30 flood and non-flood classes

over Houston, Texas remained We labeled each remaining sampled

pixel as “flooded” or “not flooded” based on visual interpretation of

true colour RGB composites of PlanetScope imagery at the same

loca-tion From these interpreted samples, we computed the overall

accu-racy, user's accuracy and producer's accuracy using both a count-based

confusion matrix as well as an area-normalized proportion matrix,

which adjusts confusion matrix counts based on the relative area

re-presented by each stratum (Stehman et al., 2003) Using the

Planet-Scope samples, we also computed accuracies for flood maps derived

using a series of VV backscatter thresholds ranging from −24 to

−15 dB for the purpose of comparison with the accuracies of the

Z-score derived maps

Second, we assessed the similarity between our predicted flood

maps and operational flood monitoring data, using vector data from the

Copernicus Emergency Mapping Service (EMS) produced in response to

the flood events in Greece and Madagascar We first rasterized and

resampled the EMS vector data using nearest neighbor resampling to

match our 10-m resolution flood maps We then overlaid the rasterized

EMS vector data onto our predicted flood rasters and counted all

in-stances of agreement and disagreements among the predicted flood and

non-flood classes As with the validation of the Houston flood

predic-tions, we excluded all pixels where prior inundation had been detected

with Landsat data, limiting the comparison to the SAR-based

predic-tions

3.6 Application programming interfaces and software packages

We developed and executed our flood mapping algorithm using the

GEE javascript and python application programming interfaces (API)

The javascript API, also known as the GEE “playground”, features

several frames, including a coding, documentation and object query

frame, as well as a mapping/visualization frame (Gorelick et al., 2017)

In addition to mapping and querying pixels on the fly, the javascript API

also allows for the export of results to cloud storage for further offline

analysis The GEE addresses of all datasets used in this study are listed

inTable 1

All validation work was done using the R programming language (R

Development Core Team, 2008) and the ‘raster’ and ‘rgdal’ packages

(Bivand et al., 2017;Hijmans, 2016) The EMS comparison work was

carried out in python using the ‘osgeo/gdal,’ ‘numpy’ and ‘rasterio’

packages (GDAL/OGR contributors, 2018;Gillies et al., 2013;Oliphant,

2006)

4 Results

4.1 Houston, Texas Z-score distributions among the PlanetScope samples taken from the

Houston site for both VV and VH polarizations were approximately normal (Fig 5A), although Z-score distributions from samples in the

flood class (n = 152) were slightly skewed towards low values VV and

VH Z-scores from non-flood samples (n = 341) had means of

−0.76 ± 1.6 (s.d.) and − 0.047 ± 1.5 (s.d.) respectively, and

medians of −0.69 and 0.15, respectively VV and VH Z-scores from

flood samples had means of −3.2 ± 1.8 (s.d.) and − 3.5 ± 2.5 (s.d.) respectively, and medians of −3.0 and − 3.2, respectively VV and VH Z-scores across all samples were moderately correlated with each other (R2= 0.47, p < 0.001) (Fig 5B)

We estimated count-based and area-normalized overall accuracies for various combinations of VV and VH Z-score thresholds (Fig 6) Maximum count-based overall accuracies were achieved when the VV

Z-score threshold was set to −3.0 and the VH Z-score threshold ranged

from −1.5 to −3.0 Using Z-score thresholds of −3.0 for both polar-izations (Table 2), the overall count-based accuracy was estimated at 0.854 and the overall area-normalized accuracy at 0.868 ± 0.0309 (95% confidence interval) Count-based user's and producer's ac-curacies were estimated at 0.803 and 0.697, respectively Area-nor-malized user's and producer's accuracies were estimated at 0.803 ± 0.0719 (95% c.i.) and 0.330 ± 0.0783 (95% c.i.), respec-tively

Maximum area-normalized overall accuracy was achieved when the

VV and VH Z-score thresholds were set to −2.0 and − 2.5, respectively

In this case, the overall count-based accuracy was estimated at 0.805 and the overall area-normalized accuracy at 0.898 ± 0.0278 (95% confidence interval) (Table 3) Count-based user's and producer's ac-curacies were estimated at 0.635 and 0.868, respectively Area-nor-malized user's and producer's accuracies were estimated at 0.635 ± 0.0680 (95% c.i.) and 0.525 ± 0.0829 (95% c.i.), respec-tively

We conducted a similar analysis using flood maps derived by ap-plying VV-backscatter thresholds We achieved a maximum overall accuracy of 0.849 ± 0.016 (95% c.i.), with a user's accuracy of 0.787 ± 0.075 (95% c.i.) and a producer's accuracy of 0.399 ± 0.081 (95% c.i.) using a VV threshold of −16 dB

4.2 Thessaly, Greece

Comparison statistics of our predicted flood maps and those of the Copernicus EMS data for Thessaly, Greece are shown inTable 4 Overall

agreement rates of 0.966, 0.981 and 0.985 were estimated using Z-score

thresholds of −2, −2.5 and − 3, respectively Agreement within the predicted flood class (analogous to user's accuracy) ranged from 0.432 for a Z-score threshold of −2 to 0.793 for a Z-score threshold of −3 Agreement within the predicted non-flood class (analogous to produ-cer's accuracy) ranged from 0.629 for a Z-score threshold of −2 to 0.776 for a Z-score threshold of −3 Both PlanetScope (Fig 7, top) and

Sentinel-1 backscatter and Z-score time series (Fig 7, bottom) over a large polygon omitted by our algorithm failed to confirm that they were

Table 1

Datasets and corresponding GEE asset addresses used in this study

in fact flooded, suggesting that this disagreement is not due to any error

on the part of our algorithm Given the consistently low VV and VH

backscatter through time for a pixel chosen within this polygon (Fig 7,

bottom), it is likely that this is a smooth, non-inundated land surface

feature that was erroneously mapped as water in the EMS vector dataset

and therefore not a true omission error

4.3 Eastern Madagascar

Comparison statistics of our predicted flood maps and those of the

Copernicus EMS data for Eastern Madagascar are shown inTable 5

Overall agreement rates of 0.979, 0.984 and 0.985 were estimated

using Z-score thresholds of −2, −2.5 and − 3, respectively Agreement

within the predicted flood class ranged from 0.533 for a Z-score

threshold of −2 to 0.707 for a Z-score threshold of −3 Agreement

within the predicted non-flood class ranged from 0.840 for a Z-score

threshold of −2 to 0.905 for a Z-score threshold of −3 Although

overall agreement rates were very close to that of Thessaly, Greece,

agreement within the predicted flood class was significantly lower,

especially among the northernmost sites (Error! Reference source not found.) Visual inspection of our flood maps over Eastern Madagascar

revealed that these “extraneous” flood pixels were usually located ad-jacent to EMS-mapped polygons An example near the coastal town of

Fig 5 A: Violin plot of VV and VH score distributions for reference samples taken over the Houston, Texas region on 2017-08-30 B: Scatterplot of VV and VH

Z-scores for the same reference samples (F = “flooded”; NF = “not flooded”)

Fig 6 Count-based (A) and area-normalized (B) overall accuracies estimated using different combinations of VV and VH Z-score thresholds Red shades indicate

higher accuracies and blue shades indicate lower accuracies (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 2

Confusion matrix from comparison between GEE-predicted (using Z-score

thresholds of −3.0 for both polarizations) and interpreted samples from PlanetScope imagery over Houston, Texas following Hurricane Harvey Actual number of samples and respective normalized area fractions per category are shown (NF = “not flooded; F = “flooded”)

Reference

Nosy Varika shown in Fig 8reveals that these pixels follow

valley-bottom tributaries leading into EMS-mapped polygons, suggesting that

these are in fact valid flooded pixels not mapped in the EMS dataset

5 Discussion

5.1 SAR backscatter anomaly trends

The first objective of this study was to better understand how SAR

backscatter Z-scores are affected by floods Assuming that per-pixel SAR

backscatter values are normally distributed over time during the

baseline period, the mean VV and VH Z-score values of −3.2

and − 3.5, respectively, for all Houston flood samples correspond to standard normal probabilities < 0.001, showing that floods elicit strong deviations from expected baseline backscatter values These deviations, demonstrated inFig 2andFig 5, are due to transitions from volu-metric or surface scattering mechanisms typical of rough land surfaces (e.g., rough soils or vegetation canopies) to specular reflection typical

of smooth open water surfaces Since water bodies present during the baseline period would already have low mean backscatter values, they

Table 3

Confusion matrix from comparison between GEE-predicted (using Z-score

thresholds of −2.0 and −2.5 for VV and VH polarizations, respectively) and

interpreted samples from PlanetScope imagery over Houston, Texas following

Hurricane Harvey Actual number of samples and respective normalized area

fractions per category are shown (NF = “not flooded; F = “flooded”)

Reference

Table 4

Agreement statistics between GEE flood prediction maps and Copernicus EMS

vector data layers for Thessaly, Greece using three Z-score thresholds applied to

both polarizations

Fig 7 Top: True colour RGB composite PlanetScope imagery over a portion of the Thessaly flooded site acquired on 2018-02-17 (left) and 2018-02-27 (centre and

right) Predicted flood classes (seeFig 4for colour table) and Copernicus EMS vector (hatched overlay) are shown on the top-right Bottom: VV and VH σ0backscatter

coefficient (top panel) and corresponding Z-scores (bottom panel) for the potential EMS commission error shown as a yellow ‘X’ in the top-right image The date for

the classified map in the top-right panel is indicated as a black arrow (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 5

Agreement statistics between GEE flood prediction maps and Copernicus EMS vector data layers for Eastern Madagascar using three Z-score thresholds ap-plied to both polarizations

would be expected to have Z-scores near zero, allowing for the

auto-matic discrimination between existing surface water bodies and new

inundation (i.e., flooding) based on these high-magnitude negative

Z-scores

The distribution of Z-scores shows a high degree of separability

between flood and non-flood samples taken over Houston during

Hurricane Harvey (Fig 5) However, we observed subtle differences in

the VV and VH score distributions among non-flood samples VH

Z-score values were normally distributed around a mean of zero, which is

not surprising, since no significant changes were expected in these

samples (Fig 5B) However, the mean Z-score for VV non-flood samples

was slightly below zero, and a possible second mode greater than zero

was observed, suggesting other land surface changes in response to the

hurricane, resulting in both negative and positive anomalies For

ex-ample, high winds brought on by the hurricane could damage or alter

vegetation canopies (e.g., orientation), resulting in decreased

back-scatter (Ramsey et al., 2015) Additionally, elevated soil and vegetation

moisture and enhanced double-bounce scattering during and following

the hurricane would give rise to an increase in enhanced dielectric

scattering off of these surfaces, resulting in increased backscatter

fol-lowing the storm (Lucas et al., 2010; Merot et al., 1994; Schmugge,

1980) In fact, the time series trajectory shown inFig 2shows a slight

increase in Z-scores following the hurricane, suggesting that the

ele-vated backscatter following the hurricane is persistent Finally,

wind-induced Bragg scattering over open surface water during hurricanes

would also give rise to positive Z-scores over open water bodies (Zhang

et al., 2014), but since our sampling strategy excluded open water

bodies (based on Landsat historical classifications), these anomalies are

not included in this analysis While the trends described here warrant

further investigation and could help to improve methods for

post-hur-ricane impact assessments, they had no impact on our classification

scheme, which targets high-magnitude negative Z-scores for flood

identification

5.2 Flood classification accuracy

Our second objective was to determine the accuracy of our classified

flood maps and their comparability with operational flood map

pro-ducts We estimated a maximum area-normalized accuracy of

89.8 ± 2.8% (95% c.i.) for the Houston case, and maximum

com-parability estimates of 98.5% for both Greece and Madagascar In all

cases, the choice of the threshold had notable impacts on the overall

accuracies, as well as the rates of commission or omission errors The count-based overall accuracies shown inFig 6A suggest an optimal VV

Z-score threshold of −3.0, with a broader range of optimal VH Z-score

thresholds This finding is consistent with the Z-score distributions shown inFig 5A, where VH Z-scores are distributed much more tightly

around a mean of zero, resulting in reduced confusion between flood and non-flood classes and greater flexibility with respect to threshold

selection Similarly, we achieved maximum overall agreement using

Z-score thresholds of −3 for the Greece and Madagascar sites Consistent with this observation, increasing the Z-score thresholds to −2 in the Greece and Madagascar cases reduced the agreement within the GEE-predicted flood class (analogous to user's accuracy) to a larger degree than the increase in agreement within the GEE-predicted non-flood class (analogous to producer's accuracy)

We found that the Z-score derived maps out-performed maps

de-rived from simple VV-backscatter thresholds, with estimated differ-ences in overall accuracies ranging between 1.6% to 9.8% While the user's accuracy of the VV-backscatter threshold maps was found to be marginally higher than that of the Z-score based maps, the large spread

of the confidence intervals indicates that this difference is highly un-certain On the other hand, the Z-score based maps had a significantly higher producer's accuracy than that of the VV-backscatter thresholded maps We therefore conclude that the Z-score method improves upon simple VV-backscatter thresholds by reducing the omission errors of the flood map, which is important for rapid-response products The im-provements that we observed in the Z-score derived maps are likely due

to the fact that SAR backscatter is impacted by variations in incidence angle across the sensor's field of view (Westerhoff et al., 2013), making selection of a static VV-backscatter threshold difficult In our algorithm,

we computed baseline statistics by grouping Sentinel-1 imagery by acquisition modes and orbital direction This method of computing baseline statistics ensures that Z-scores are influenced by historical and current conditions at the pixel level, and not by spatial variability caused by viewing geometry

Our validation has a number of limitations worth noting here First, area normalization, in which weights are assigned to samples based on their respective inclusion probabilities, is recommended for accuracy assessments and area estimations in land cover and land use change studies where validation samples have been drawn from stratified random samples (Olofsson et al., 2014;Stehman et al., 2003) However, the fact that the size of our “non-flood” stratum in our sampling design greatly exceeded that of the “flood” stratum resulted in a large disparity

Fig 8 Agreements and disagreements resulting from overlay of GEE-predicted floods and EMS vector data for the area surrounding Nosy Varika, Madagascar.

Original EMS polygons are shown as a hatched overlay

in weights assigned to samples drawn from the “non-flood” stratum As

a consequence, the area-normalized producer's accuracies were

typi-cally very low, reflecting the lack of representativeness of these

sam-ples For this reason, we report both count-based and area-normalized

accuracy statistics for the Houston, Texas validation

Second, the high degree of comparability between GEE and EMS

predictions is heavily influenced by disproportionately large areas of

agreed non-flooding Despite very high overall agreement rates, large

areas of disagreements reveal potential errors in the semi-automated

EMS workflow, including large commission errors in the Thessaly,

Greece case (Fig 7) In this example, low backscatter over a smooth

non-flooded surface was likely erroneously labeled as flooded in the

EMS dataset However, comparison of the backscatter and Z-score time

series with PlanetScope imagery around the flood date revealed that

this is a persistently non-flooded area While the purpose of this study

was not to evaluate existing EMS map products, these results highlight

the need for further improvement of algorithms and workflows used to

generate operational disaster monitoring products

5.3 Strengths of the method

The implementation of our algorithm on the GEE confers several

advantages over conventional flood mapping and monitoring methods

First, the temporal SAR backscatter Z-scores provide an objective

measure of change for individual image pixels Since Z-scores are

di-rectly related to standard normal probability distribution functions,

users can define thresholds based on a desired probability or confidence

level In contrast, the decision on where to set a backscatter threshold

can be complicated by variations in backscatter due to other factors like

view angle (Huang et al., 2018; Westerhoff et al., 2013) or Bragg

scattering (Zhang et al., 2014) It is important to note that the use of

Z-scores relies on the assumption that backscatter coefficient values are

normally distributed through the baseline period, which may not

ne-cessarily hold true when observations taken under various sensor

con-figurations are included in the baseline Specifically, the incidence

angle for each observation is known to have an effect on backscatter off

of inundated surfaces (Huang et al., 2018; Schlaffer et al., 2015;

Westerhoff et al., 2013), which is likely the cause of the systematic

difference between IW and SM backscatter values shown inFig 2(top

panel) We were able to remove most of this bias at the anomaly

computation stage by computing baseline statistics separately for

dif-ferent acquisition modes and orbital directions (Fig 2, middle panel)

When this bias is taken into consideration, the use of temporal Z-scores

allows for rapid monitoring of floods while simultaneously correcting

for systematic difference arising from varying sensor configurations

Second, the use of historical Landsat allowed for the

contextualization of Sentinel-1 Z-scores contemporary with the flood events, allowing for further discrimination between floods and his-torically permanent or seasonally inundated areas Since SAR Z-scores are based on statistics computed from observations within a limited baseline temporal window, historical Landsat-based inundation prob-abilities provide an important constraint for our classification algo-rithm For example, seasonally inundated wetlands like the coastal wetlands shown inFig 9may experience flooding above and beyond what has historically occurred Knowledge on where floods are occur-ring in previously non-flooded areas is an important component to flood monitoring systems, as these floods have the potential to be cata-strophic and likely require greater attention in disaster response and rehabilitation plans

Finally, our algorithm is based on pixel-based statistics derived from SAR and optical time series data, making it compatible with data cube computing architectures like the GEE (Lewis et al., 2017) Using GEE's web-based javascript application programming interface (API), we were able to generate time series flood maps for a number of recent flood events in several minutes (Fig 10), compared to a processing time of several days on a high performance computing cluster This perfor-mance improvement is not only due to the fact that all data are already downloaded and pre-processed on the GEE servers, but also because of the fact that pixel-based operations are easily parallelized, facilitating rapid data exploration and algorithm development (Gorelick et al.,

2017) Additionally, the rapid deployment of time series algorithms allows users to query and interpret single pixel time series on the fly, which can be incorporated into interactive web-based tools as demon-strated inFig 11

5.4 Limitations and potential improvements

Despite the strengths of our algorithm and the GEE platform in monitoring floods in near real-time, we encountered several limitations that warrant further study First, other land surface dynamics can give

rise to temporal Z-scores on the order of that observed in our flood

examples For example, the removal of forest or crop canopies is asso-ciated with a change in SAR backscatter signatures from those domi-nated by volumetric canopy scattering to those domidomi-nated by soil sur-face scattering (Shimada et al., 2014) While the latter backscatter signature is typically not identical to open surface water, such changes may be sufficiently large to generate large anomaly scores (Cian et al.,

2018) In some cases, these dynamics may be a result of regular in-undation patterns (e.g., flooded rice fields), which make selection of a non-flooded baseline with sufficient number of observations difficult In such cases, it will be possible to define baseline periods consisting of specific subsets of multiple years (e.g., only dry season imagery) as

Fig 9 Example of flooding along coastal wetlands south of Houston, TX on 2017-08-30 Predicted new (red) and recurring (purple) inundation (left panel) and

coincident PlanetScope imagery (right panel) are shown for comparison The full colour key used in this map is shown inFig 4 (For interpretation of the references

to colour in this figure legend, the reader is referred to the web version of this article.)