image
image
image
image
1
Rainfall estimates from opportunistic sensors
in Germany across spatio-temporal scales
Maximilian Graf
1,3
, Abbas El Hachem
2
, Micha Eisele
2
, Jochen Seidel
2
, Christian Chwala
1,3
,
Harald Kunstmann
1,3
, and András Bárdossy
2
1 KIT, IMK-IFU, Garmisch- Partenkirchen, Germany,
maximilian.graf@kit.edu
2 IWS, University of Stuttgart, Germany
3 IGUA, University of Augsburg, Germany

image
image
image
image
image
image
2
Opportunistic sensors (OS) can be used for
rainfall monitoring
Commercial Microwave Links (CMLs)
§
~ 4000 CMLs
§
fixed set of CMLs with custom real time
application
1
together with Ericsson
§
10 to 40 GHz with 0.3 to 30 km length
Personal Weather Stations (PWSs)
• up to 20,000 PWSs from netatmo
• number of PWSs is increasing
Other examples from a growing number of opportunistic sensor for environmental monitoring
Smart phones
à
temperature, pressure, light
Windshield wipers
à
rainfall binary info from windshield wipers
Satellite TV link path
à
rainfall
Surveillance cameras
à
rainfall
1
Chwala et al. 2016, AMT

image
image
image
image
image
image
3
Commercial Microwave Links
à
Relation between attenuation and rain rate is defined as
Standard
Marshall-
Palmer DSD
! = #$
%
[dB/km]
[mm/h]
A-R power law:
Chwala and Kunstmann, 2019 (Wires
)
Polz et al., 2020 (AMT)
Graf et al., 2020 (HESS)
more information
on CML processing

image
image
image
image
4
Personal weather station (PWS)
indoor/outdoor units to measure:
• temperature
• humidity
• pressure
• CO2
• wind
• rainfall
wireless weather station for the ”smart home” here from Netatmo
http://www.aragonvalley.com/
manufacturer's specifications
• range of 0.2–150 mm/h
• precision of 1 mm/h
• 13 cm diameter

image
image
image
image
image
image
image
image
5
Personal weather station (PWS)
https://weathermap.netatmo.com/
Owners of a netatmo PWS can
access data from all other publicly
shared PWS via an API

image
image
image
image
image
6
Evaluating rainfall estimates through scales
What are the challenges?
an adequate quality control routine has to be used for opportunistic sensors
à
remove only as much data as necessary to profit from high number of sensors
interpolate sensors individually and in combination
find suitable reference data sets to evaluate rainfall estimates from OS

image
image
7
Processing and Interpolation
raw CML data
(attenuation data)
raw PWS data
(rainfall data)
rainfall estimation
with pycomlink
indicator correlation filter
bias correction
event based filter
DWD
auto
rain
gauge data
block kriging interpolation
rank information
hourly rainfall maps
seven combinations of
CML, PWS and DWD
auto
distribution functions
estimated variogram
Assumption:
even if the exact OS
values are wrong,
their rank within one
sensor should be fine

image
image
image
8
Data availability after filtering
~ 92% of the data are assumed to be ok and used for seven interpolated products

image
image
image
image
image
9
Evaluating rainfall estimates through scales
Concept of evaluation
seven interpolated products with hourly resolution which
consist of PWS, CML and DWD (hourly, automatic stations)
and their combinations
evaluation of seven interpolated products for three scales
1
DWD
daily
≠ DWD
hourly
, these are two different
gauge dataset with different locations
scale
region
temporal
n stations
data provider
country
Germany
daily
1062
DWD (manual gauges)
regional
Rhinland-
Palatinate
hourly
169
Agrometeorological Agency of
Rhinland-Palatinate
local
Reutlingen
hourly
12 (10)
Municipality of Reutlingen

image
image
image
10
Country-wide, daily scale: Germany
performance of interpolated products for 1062 manual,
daily rain gauges from DWD (DWD
man
)

image
image
image
image
11
Country-wide, daily scale: Germany
Pearson's correlation coefficient (-)
bias (%)
Kling-Gupta efficiency (-)
false positive rate (-)
§
OS products correlate similar or better to the reference than one of DWD rain gauges

image
image
image
12
Country-wide, daily scale: Germany
Pearson's correlation coefficient (-)
bias (%)
Kling-Gupta efficiency (-)
false positive rate (-)
§
OS products correlate similar or better to the reference than one of DWD rain gauges
§
interpolated CMLs show a negative bias and high false positive rate mainly due to their
uneven spatial distribution in relation to the DWD
man
gauges

image
image
image
13
Regional, hourly scale: Rhineland-Palatinate
Pearson's correlation coefficient (-)
bias (%)
Kling-Gupta efficiency (-)
false positive rate (-)
performance of interpolated products compared to 169
hourly rain gauges operated by the
Agrometeorological Agency of Rhineland-Palatinate

image
image
image
14
Regional, hourly scale: Rhineland-Palatinate
Pearson's correlation coefficient (-)
bias (%)
Kling-Gupta efficiency (-)
false negative rate (-)
§
combination of OS performs better than combination of OS with DWD
§
False negative rate of OS and combinations is lower than DWD or radar
à
Even though OS do not measure at the validation stations (RLP) they perform reasonable in comparison
to radar measurements at such locations

image
image
image
15
Local, hourly scale: Reutlingen
performance of interpolated products compared to
10 hourly rain gauges operated by the Municipality
Reutlingen (RT)

image
image
image
16
Local, hourly scale: Reutlingen
§
with sparse spatial coverage (no gauge in the Figure), interpolated DWD gauges perform worse
than OS for this local example
§
OS and combinations perform similar good as radar products
à
while PWS have better correlation, CML improve the bias

image
image
image
17
Map example country-wide

image
image
image
18
Example Reutlingen
Pluvio
CML
PWS

image
image
image
19
Example Reutlingen

image
image
image
20
Example Reutlingen

image
image
image
image
21
Example Reutlingen

image
image
image
22
Example Reutlingen

image
image
image
23
Example Reutlingen

image
image
image
24
Example Reutlingen

image
image
image
image
image
25
Conclusion
Goal
• Estimation of rainfall in Germany with opportunistic sensors
Challenge
• OS need extra care during quality control and processing
Results
• OS can yield rainfall estimates of reasonable high quality
Graf, M., El Hachem, A., Eisele, M., Seidel, J.,
Chwala, C., Kunstmann, H., & Bárdossy, A. (2021).
Rainfall estimates from opportunistic sensors in
Germany across spatio-temporal scales.
Journal of Hydrology: Regional Studies
,
37
, 100883.

image
image
26
Outlook
DFG proposal HiPOSY
High-resolution Precipitation Products from Opportunistic Sensors and their
Impact on Hydrological Modelling
1. Improve OS sensor data quality control
2. Improve rainfall fields using OS data
3. Evaluate new rainfall products using hydrological modelling
4. Assess uncertainties using inverse hydrological modelling
5. Evaluate OS rainfall products with respect to rainfall statistics

image
image
image
image
image
image
27
Acknowledgments
We want to thank Ericsson Germany, in particular the IT team, for
their support with the CML data acquisition
and HGF, DFG and BMBF for funding and supporting our research.

image
image
image
28
Indicator correlation filter
indicator correlation (IC):
rank correlation of individual PWS, CMLs or DWD
auto
99% quantile to their next neighbors
PWS and CML are removed when their IC is lower then the IC with the next DWD
auto
station
Assumption: even if the exact values are wrong, their rank within one sensor should be fine

image
image
image
29
Bias correction and event based filter
bias correction
: precipitation distribution function of DWDauto are used to adjust
OS values
event based filter:
square root of each OS precip value is compare to the
estimated variogram value of the next 30 DWD
auto
gaugea in order to remove
(mostly) faulty zeros

image
image
image
image
image
image
image
30
Interpolation Framework: (Block-) Kriging
Include uncertainty of opportunistic sensors
Account for line characteristic of CMLs
DWD
auto
PWS
CML

image
image
image
image
image
image
31
31
CML processing
5
4
3
2
1
Remove erratic data
Rain event detection
Attenuation from baseline level
Compensate wet antenna attenuation
Derive rain rate
! = #$
%
CML derived rain rate
(hourly resampled)
reference rain rate