Presentation by several speakers at the Hydrology Suite User Days (Day 2) - wflow and HydroMT, during the Delft Software Days - Edition 2023 (DSD-INT 2023). Wednesday, 29 November 2023, Delft.
6. Day 2 – MORNING PROGRAMME
Arrival (8:30-9:00)
wflow/HydroMT
User
Day
9:00 – 9:10 Opening and Welcome Hessel Winsemius, Hydrologist at Deltares
9:10 – 9:35
Introduction to wflow and HydroMT-wflow: Overview and New
Developments
Ali Meshgi, Product Owner at Deltares
9:35 – 10:00 Demo HydroMT-wflow, building a model from scratch Joost Buitink, Hydrologist at Deltares
10:00 – 10:20
Global Water Watch, a platform supporting the democratization of
information about global water resources
Björn Backeberg, Senior researcher at
Deltares
10:20 – 10:40 Coffee break
10:40 – 11:00
Multi-threading in wflow_sbm: rapid derivation and model runtimes for
Europe on high-performance clusters
Ruben Imhoff, hydrometeorologist at
Deltares
11:00 – 11:20
Application of wflow for water balance modelling of surface reservoirs
in Burkina Faso
Daan te Witt, Water management
consultant at Witteveen & Bos
11:20 – 11:40
wflow_sbm application for one-way model coupling with groundwater
flow model in Bandung, Indonesia.
Steven Rusli, PhD Student at Wageningen
University
11:40 – 12:00
Flood hazard modelling with coupled wflow and SFINCS models,
supported by HydroMT
Dirk Eilander, Hydrologist at
Deltares/assistant professor at VU
Amsterdam
12:00 – 12:20
Machine learning for predicting spatially variable lateral hydraulic
conductivity: a step towards efficient hydrological model calibration and
global applicability
Awad Mohammed Ali - PhD student at
Wageningen University
12:20 – 12:30 Closing remarks Hessel Winsemius, Hydrologist at Deltares
Group photo and Lunch break (12:30 – 13:30)
7
7. Ali Meshgi
Deltares
Joost Buitink
Deltares
Björn Backeberg
Deltares
Ruben Imhoff
Deltares
Daan te Witt
Witteveen+Bos
Steven Rusli
Wageningen University
Dirk Eilander
Deltares/VU
Awad M. Ali
Wageningen University
Speakers of Hydrology Suite User Days (Day 2)
8. Ali Meshgi
Ali Meshgi is a senior data scientist and urban groundwater
hydrologist at Deltares, Department of Catchment and
Urban Hydrology since 2018. He holds a Ph.D. in Civil and
Environmental Engineering from the National University of
Singapore and has 15 years of experience in both the
private and government sectors, as well as in universities.
His expertise spans various areas, including groundwater
and hydrological studies, computer modelling and
programming, data analytics, machine learning, project
management/lead, and product owner. Currently, he holds
the role of product owner for HydroMT and wflow.
Introduction to wflow and HydroMT-wflow: Overview and New Developments
Product Owner at Deltares
10. Why do we need models?
11
See level rise
Subsidence
Population growth
Urbanization
Climate change
This requires all the smart minds and all the smart technology
11. 12
Why do we need models?
We (approximately) know what happened…
• In the past
• At a certain location
12. 13
?
➢ What will happen in the future?
➢ And what happens at other
locations?
➢ What are the causes of current
issues?
➢ Are the land-use management
plans and mitigation measures
efficient and sustainable?
Why do we need models?
Climate change Population growth
13. 14
Why do we need models?
Model can help with:
➢ Better understanding of a complex
system
➢ Assess effects of scenarios or
measures and strategies
➢ Predict into the future
➢ …
14. • Open Source & free
• For catchment-scale rainfall-runoff
modelling
• Fully distributed (grid-based)
• Online documentation and
examples
• Different hydrological modelling
concepts
Hydrological modeling with wflow: General info
15
15. HBV
SBM
16
Hydrological modeling with wflow: Concept
Currently, most developments are in SBM concept
• Conceptual
• Based on HBV’96 concept
• Implemented in the wflow
framework by Deltares
HBV concept
• Flexible process-based model
FLEX-Topo concept
• Physically based
• Based on Topog-SBM concept
• Derived from the CQflow
• Developed and maintained by
Deltares
SBM concept
17. Hydrological modeling with wflow: Use case examples
18
Sediment inflow to reservoirs/lakes
Groundwater recharge calculation
Pollutant fate and transport input
Reservoir inflow calculation
Data-scarce region use
Large-scale application
Inputs for flood modelling
18. Hydrological modeling with wflow
19
o Improve computational efficiency
o Add Floodplain routing
o Improve clearness of
documentations
o Enhance software stability
Recent Developments
Moving towards v1.0 by:
o Providing comprehensive
documentation
o Ensuring the software is stable
and free of critical bugs
o Defining a support and
maintenance plan
Upcoming features and planning
Other plannings:
o Including water demand analysis
o Better quantifications of NBS
o Online coupling with MODFLOW
& RIBASIM
19. Key Characteristics
& Vision
Integration Capability
• Integrates with other
models
• Open source & free
• Easy to set up using HydroMT-wflow
• Clear documentation and support
• Windows and Linux compatibility
Usability
• Simulates a wide range of
hydrological processes and
phenomena
Versatility
• Provides reliable results
under a variety of conditions
Robustness
• Ongoing research on creating
an emulator for wflow using a
graph neural network
Computational Efficiency
• Ongoing pilot study on
cloud storage
Data Storage Solution
• Works at different
spatial and temporal
scales
Scalability
• Accurately predicts
hydrological variables
Accuracy
Hydrological modeling with wflow: Key Characteristics & Vision
20. 21
Hydrological modeling with wflow: Modelling Process
• Physical catchment
characteristics
• Hydro-
meteorological data
• Satellite data
Data
collection
• Setting up the model
from data and the
understanding of the
relevant hydrological
processes
Model set-up
• Getting values for
the parameters
Calibration
• Confronting the
model output with
observations
Evaluation
21. 22
Models need data…
• Base model data:
➢ E.g.: elevation, land-use, river
cross-sections…
Construction data
• Data to execute the model
➢ E.g.: meteorological data
(precipitation, wind…),
Execution data
• Data to estimate the efficiency of
the model
➢ E.g.: discharge, water level,
flood maps…
Validation data
22. 23
Time
1) Month 1-3
2) Month 3-6
3) Month 6
4) Month 7
5) Month 8-12
6) Month 12
The classical approach for Water Resources
Management projects:
1) Collect local data
2) Collect more local data
3) Collect even more local data
4) Analyse data
5) Build a model
6) Throw away 80% of the collected data
Models need data…
23. 24
GIS
processing
HydroMT was born!
Models need data…
➢ Operate on these datasets.
➢ They are easy to use.
➢ Support the modelers by handling the "hard" and
"boring" work.
➢ Support the development of enhancements for these
datasets.
➢ Support the reuse of methods and workflows for
different models.
➢ Good data becomes more widely available.
➢ Data becomes available at higher temporal and spatial
resolutions.
➢ As the data improves, we aim to enhance our models.
Need for tools
24. 25
✓ Reproducible
✓ Automated
✓ Fast
✓ Flexible
✓ Scalable
Open-source Python package that facilitates
automated and reproducible pre- and
postprocessing of spatial geoscientific models
Build Update
Plot Stat
HydroMT, Model Building and Analysis: General Info
27. 28
• Open Source
• Available on Github
• MIT license
• Releases on pypi and conda for
easier installation
• Online documentation and
interactive examples with Binder
• Works via command line
HydroMT, Model Building and Analysis: General Info
29. 30
HydroMT, Model Building and Analysis:
Use case example
❑ Rapid model building /updating globally
❑ Small to large applications
❑ Operate on global and local datasets (data
harmonization)
❑ Support developments in enhancing local
datasets
30. 31
HydroMT, Model Building and Analysis:
Use case example
❑ Rapid model building /updating globally
❑ Small to large applications
❑ Operate on global and local datasets (data
harmonization)
❑ Support developments in enhancing local
datasets
31. 32
Mines locations
Regional statistics
HydroMT, Model Building and Analysis:
Use case example
❑ Rapid model building /updating globally
❑ Small to large applications
❑ Operate on global and local datasets (data
harmonization)
❑ Support developments in enhancing local
datasets
Global population grid
32. 33
HydroMT, Model Building and Analysis:
Use case example
❑ Rapid model building /updating globally
❑ Small to large applications
❑ Operate on global and local datasets (data
harmonization)
❑ Support developments in enhancing local
datasets
33. 34
HydroMT, Model Building and Analysis
1. Improve the versatility and
reproducibility of our data
catalogue
2. Improve usability by introducing
model checking
3. Improve clearness of
documentations
4. Improve developer velocity by
focusing on CI automation
Recent Developments
1. Moving towards v1.0 by:
o Stabilising our external API
o Defining a support and maintenance
plan
o Improving the robustness of the
code
2. Enabling Cloud support by:
o Making data better available for
collaboration
o Optimising the data for cloud use
o Optimising the code to take
advantage of cloud architecture
Upcoming features and planning
34. wflow & HydroMT Development Team
Joost Buitink
Developer
Willem van Verseveld
Developer
Laurene Bouaziz
Developer
Brendan Dalmijn
Developer
Ali Meshgi
Product Owner
Dirk Eilander
Developer
Hélène Boisgontier
Developer
Sam Vente
Developer
Tjalling de Jong
Developer
Jaap Langemeijer
Developer
Anaïs Couasnon
Developer
35. How do we work?
36
Start
Market &
research
Agile product
development
Design
Capacity
planning
Development
Test
Deploy
Deliver
36. General links for wflow and HydroMT
37
Documentation for HydroMT
How to access and install HydroMT
Feature application example for HydroMT
hydromt@deltares.nl
Documentation for wflow
How to access and install wflow
Feature application example for wflow
wflow@deltares.nl
HydroMT webpage
wflow webpage
40. Joost Buitink
Demo HydroMT-wflow, building a model from scratch
Hydrologist at Deltares
Joost Buitink works as a hydrologist at Deltares. He applies
wflow in different catchments around the world, with a focus
on understanding the effects of climate change on the
hydrological response. Besides the work in projects, he is
also one of the developers of wflow and hydromt_wflow
41. Translate data into a wflow model in only
a couple of steps!
29-11-2023
Demo HydroMT-wflow
42. Building a model
Steps to build a model
1. Select the data you want to use by
preparing a data catalog or using a pre-
defined one
2. Select your region of interest
3. Select the model components you want to
build and the options (data source,
resampling method, etc.) by preparing a
model configuration file or using a pre-
defined one
4. Build your model using the Command
Line Interface (CLI) or Python
> hydromt build model “my_model” “{‘basin’: [x,y]}” -i model_build.ini -d data_catalog.yml -vv
43
43. 1. Select the data you want to use
• Select the data you want to use and have it ready in one of the supported raster (tif, ascii, netcdf,
zarr…), vector (shp, geojson, gpkg…) or geospatial time-series formats (netcdf, csv…)
• If needed, create your own data catalog with a reference to your selected data, including:
− Path to the data location
− Data type and hydromt driver
− Rename and unit conversion
− Others…
• Or use a pre-defined data catalog:
− deltares_data → Deltares P drive
− artifact_data
− demo_data/data_catalog.yml
HydroMT documentation:
https://deltares.github.io/hydromt/latest/user_guide/data_prepare_cat.html
44
44. 2. Select your region of interest
• Select the region of interest for which you
would like to build a wflow model
• The type of region:
− basin: full basin until sea outlet (eg. Rhine)
− subbasin: sub-basin until a specific point
outlet (e.g Moselle, tributary of the Rhine)
• Defining the location of your region: options:
− Point location [x, y]
− Bounding box [xmin, ymin, xmax, ymax]
− Geometry file ‘my_region.shp’
− Basin ID [ID1]
HydroMT documentation:
https://deltares.github.io/hydromt/latest/user_guide/model_region.html
45
45. 3. Select the model components to build
• Start from a template configuration file:
https://deltares.github.io/hydromt_wflow/latest/user_guide/wflow_build.html#configuration-file
• Edit / add / remove sections to select which components
to prepare and change the building options such as data
source, resampling method, default values, etc.
HydroMT documentation:
https://deltares.github.io/hydromt/latest/user_guide/model_config.html
46
46. 4. Run HydroMT
• Build your model using the HydroMT CLI (see example below)
− Call hydromt
− Select build to prepare a model from scratch to update an existing model
− Select which type or model, here wflow
− Path where the generated model will be stored: e.g. folder called “my_model”
− Select the region of interest, e.g. for a basin containing point x,y: “{‘basin’: [x,y]}”
− Path to the model configuration file listing the components to build with the –i option: e.g. -i wflow_build.ini
− Path to the data catalog with the input data sources with the –d option (several catalogs possible),
e.g. -d data_catalog.yml
− Change verbosity with -v option, e.g. -vv for extra verbosity
> hydromt build wflow “my_model” “{‘basin’: [x,y]}” -i wflow_build.ini -d data_catalog.yml -vv
HydroMT documentation:
https://deltares.github.io/hydromt_wflow/latest/user_guide/wflow_build.html
47
48. Björn Backeberg
Global Water Watch, a platform supporting the democratization of information
about global water resources
Project Leader at Deltares
Björn Backeberg is an experienced ocean expert with over
15 years of experience in academia, applied sciences, and
industry. He specializes in understanding how the ocean
affects our climate, using computer models, data analysis,
and satellite information. He obtained his PhD at the
University of Cape Town and played a key role in creating a
system to predict ocean conditions in South Africa.
Throughout his career, he has held various roles, including
being a lead researcher and working on ocean and coastal
data management. Currently, he works as a Senior
Researcher and Project Leader at Deltares, where he's
involved in European Commission projects and manages
the BlueEarthData Platform. He's also a member of the
Deltares Young Science Council.
49. A platform supporting the
democratisation of information on water
resources
29 November 2023
Global Water Watch
Björn Backeberg
Marine and Coastal Systems
Environmental Hydrodynamics and Forecasting
50. Table of contents
51
• Sponsor and partners
• The Team
• The Challenge
• Our Solution
• Demo
• Roadmap
• Conclusion
53. The challenge
54
HIGH-RESOLUTION DATA ON WATER RESOURCES ARE FREQUENTLY NOT AVAILABLE
Example: Reservoirs / Dams
Critical
infastructure
Anticipate water
resource trends
Data-driven
management
Peace and secutiry
(deliver facts)
54. Our solution
55
Open Source Water Datasets Water Analytics Water Stories
Support the democratisation of information
on water resources
• Near real-time methods to estimate
reservoir area and volume dynamics from
satellite data
• Access to global datasets through a hosted
web application and API.
55. Global Water Watch platform demo
Follow along at
https://www.globalwaterwatch.earth/map/
56. The Global Water Watch Platform
57
Our current reservoirs time series database
(1984-2023) includes 71 208 reservoirs
91. Roadmap
92
Platform enhancement
and community building
• Improve and extend
reservoir database
• Add AI-based dam
detection and mapping
• Add uncertainty
estimates
• Monitor very large
reservoirs
• Outreach and uptake
• Impact stories
2024
Extend platform scope
• Add surface water
changes
• Citizen science
capabilities
2025
Extend platform scope
• Add data-driven and
process modelling
based forecasting
capabilities
• Extend to monitoring
and forecasting
rivers and lakes
2026
92. 93
EARTH OBSERVATIONS
MACHINE LEARNING
WATER EXPERTISE
CLOUD COMPUTING
FARMERS
POLICY MAKERS
GOVERNMENTS
INSURANCES
HYDROPOWER
DRINKING WATER
CONSERVATION
Summary
• Example of co-creation and research to
operations
• Platform monitors water dynamics on a
planetary scale
• Intuitive web app with data analytics
supporting decision-making and
storytelling
• API for downstream application
development and integration
93. Thank you on behalf of the team
94
Join the community:
https://www.globalwaterwatch.earth/
Help us improve the platform!
96. Ruben Imhoff
Multi-threading in wflow_sbm: rapid derivation and model runtimes for Europe
on high-performance clusters
Hydrometeorologist at Deltares
Ruben Imhoff is a hydrometeorologist working in the
Operational Water Management and Early Warning
Department at the research institute Deltares. He primarily
focuses on hydrological modeling and short-term
hydrometeorological forecasting systems for rapidly
responsive catchments and urban areas, often utilizing
radar-based methods like nowcasting. This aligns with
Ruben's academic background, as he recently obtained his
Ph.D. in rainfall nowcasting for flood early warning from
Wageningen University in the Netherlands.
97. Ruben Imhoff
Willem van Verseveld
Joost Buitink
Albrecht Weerts
Rapid deployment and model runtimes
for Europe on high-performance clusters
29 November 2023
Multi-threading in
wflow_sbm
98. Table of contents
99
• The ACROSS project: low-latency weather forecasts, climate projections and subsequent
hydrological model simulations
• Speeding up the wflow_sbm model using Julia’s multi-threading
− From Python to Julia
− Multi-threading improvements
− Tests on Karolina
• Rapid deployment of high-resolution hydrological model whenever new forcing is available
− Direct interaction with new meteorological forecasts and climate model simulations
− Model setup for Europe
− Integration of all steps using Streamflow workflow manager
− Model validation
99. The ACROSS project: low-latency hydrological model
simulations with wflow_sbm
100. The ACROSS project: low-latency hydrological model
simulations with wflow_sbm
101
Forecasts Climate projections
Storage Storage
Obtain, regrid and pre-process
forcing data for wflow_sbm
models
Data / observations
Download data through API
Hydrological simulation
Post processing
101. The ACROSS project: low-latency hydrological model
simulations with wflow_sbm
102
System by ECMWF and MPI: “BORGES” (Better
ORGanized Earth System model data) data system
Needs:
• Direct interaction and pre-processing of
forcing
• Fast model runtimes
102. Speeding up the wflow_sbm model using Julia’s multi-
threading
104. Testing Julia’s multi-threading with wflow (for the Rhine)
105
Kinematic wave land and river
Kinematic wave land – local inertial river
1D Floodplain routing
105. Multi-threading improvements
106
Tested solutions to improve multi-threading:
• Minimize overhead by using a minimum ‘basesize’, which sets the number of elements of the iteration space
processed by each thread.
• Replacing Julia’s @threads macro with the @spawn macro in combination with a minimum ‘basesize’.
• Using low overhead threading of Polyester.jl, as a replacement of the native Julia threading functionality.
• Setting Julia environment variable JULIA_EXCLUSIVE=1 to pin threads to CPU-cores.
For the local-inertial routing:
• This routing scheme does not rely on a specific execution order → suited for loop vectorization
• Further speedup of computations by using SIMD (Single Instruction, Multiple Data) vectorization and loop
reordering, provided by the Julia package LoopVectorization.jl.
107. Rapid deployment of high-resolution hydrological
model whenever new forcing is available
108. Direct interaction with new meteorological forecasts
and climate model simulations
109
Forecasts Climate projections
On-the-fly regridding and
downscaling of forcing
variables for wflow
FDB
BORGES
FDB request with pyfdb
and pre-processing
Karolina HPC
wflow_sbm
109. Model setup for Europe
• High-resolution (1 km2) basin delineation based on MERIT DEM (90x90 m)
− Automatic basin delineation using HydroMT (Eilander et al., 2023, Journal of Open Source Software)
− Automatic parameterization of land, soil and river parameters (Imhoff et al., 2020, Water Resources Research; Eilander et al.,
2021, Hydrology and Earth System Sciences)
• Merging of small catchments with larger ones → reduction to 108 basins that can be run in parallel
110
European catchments at 1-
km resolution
Merged catchments after
12,000 km2 catchment area
constraint
110. Integration of all steps using StreamFlow workflow manager
111
wflow_sbm models
for 108 catchments
Run in parallel with Julia
multi-threading on Karolina
Post processing
Forecasts Climate projections
FDB
BORGES
On-the-fly regridding and downscaling
of forcing variables for wflow
FDB request with pyfdb and pre-
processing
111. Results
112
• ECMWF forecast speed up in absolute amounts
not enormous for Rhine and Meuse, but:
− Based on 1 member: with 100 ensemble members,
this becomes 100x the shown speed up
− Forecast up to 240h in advance, true usage up to 15d
and 46d in advance by Rijkswaterstaat
− True power in running this for all of Europe at once.
With multi-threading and parallel distribution using
StreamFlow, this can be done in approx. the same
time
Meuse Rhine
1 thread 32 threads Speed up 1 thread 32 threads Speed up
ERA5 daily run from 1988
– 2022
8.6 h 4.6 h 1.9x 33.2 h 9.3 h 3.6x
ECMWF forecast, hourly
up to 240 hours
178 s 118 s 1.5x 321 s 156 s 2.1x
112. Summary
113
ACROSS project: low-latency weather forecasts, climate projections and subsequent hydrological
model simulations
− Collaboration with ECMWF and MPI
Sped up wflow_sbm by optimally using Julia’s multi-threading
− Reduction in overhead with multi-threading (run times do not go up anymore with many threads)
− 2 – 11 times faster on top of speed up due to transition to Julia
Rapid deployment of high-resolution hydrological model whenever new forcing is available
− Direct interaction with new meteorological forecasts and climate model simulations
− Model setup for Europe
− Integration of all steps using StreamFlow workflow manager
115. Daan te Witt
Application of wflow for water balance modelling of surface reservoirs in
Burkina Faso
Hydrologist at Witteveen+Bos
Daan te Witt has over 3 years of experience using wflow.
Daan graduated in 2021 from Delft University of Technology
in Water Management. His master thesis was about the
calibration of a wflow model using the spatio-temporal
patterns of satellite datasets to simulate the discharge in the
Volta river basin in West-Africa. Daan works as a
hydrologist at the Water Management department of
Witteveen+Bos for 2 years. He set up and applied wflow
models for different water resources projects all over the
world (Indonesia, Brazil, Guatemala, Kazakhstan, Burkina
Faso)
116. Rehabilitation of Large Water
Reservoirs in Burkina Faso
Phase 1: Preliminary design for reservoirs Boulsa, Boussouma, Loumbilla and Toécé
wflow user day, 29 Nov 2023
117. Goal of the project
118
˗ Problems in the reservoirs:
· Siltation
· Large (evaporation) losses
· Water quality
˗ Goal: To develop an integrated set of measures
to improve:
· reservoir storage capacity
· limit siltation
119. Project area
120
˗ Burkina Faso
˗ 4 Reservoirs
˗ Volta & Niger catchments
˗ Toécé, Loumbila, Boussouma
and Boulsa
120. Project area
121
˗ Burkina Faso
˗ 4 Reservoirs
˗ Volta & Niger catchments
˗ Toécé, Loumbila, Boussouma
and Boulsa
121. Application of wflow
122
˗ Simulation of discharge dynamics
· Water balance
· Extreme value analysis
· Effects of climate change
˗ Simulation of erosion and sedimentation processes
122. Application of wflow
123
˗ Simulation of discharge dynamics
· Water balance
· Extreme value analysis
· Effects of climate change
˗ Simulation of erosion and sedimentation processes
123. Model setup
124
˗ Modules
· wflow_sbm
· Reservoir module
· Sediment module
˗ Data
· Model basis was set up using HydroMT
· We added precipitation data (ERA5, CHIPRS, local data)
· We added evaporation data (ERA5, GLEAM, local data, and more)
· We added reservoir data (GRanD/HydroLakes)
126. Evaporation data
127
˗ Important factor in this region
˗ Huge differences between
available datasets
˗ Little local data available
˗ Ended up using the local data
but corrected by ERA5
127. Calibration of the hydrological model
128
˗ Little (reliable) discharge data was
available
˗ Why are we not able to simulate
the first discharge peak of the
year?
· Crust formation?
· Low infiltration capacity?
˗ Peak flows are often overestimated
128. Calibration of the reservoir module
130
˗ Calibration of reservoirs using
reservoir parameters:
· max volume
· target fraction full
· target minimum full fraction
· demand
· max release
· area
129. Calibration of the reservoir module
131
˗ Only WL-timeseries for 2 reservoirs
˗ Used satellite data to detect the
surface area of other 2 reservoirs
130. Calibration of the reservoir module
132
˗ Would be nice to add:
· Infiltration
· Abstractions
· HAV-relations
132. What we also did
135
˗ Set up an erosion/sediment model
· Identify erosion hotspots
· Estimate the sedimentation
˗ Estimated effects of climate change
˗ Calculated the effect of different measures
on the water balance
· Solar panels, dredging, catchment
reforestation, raising the dam lvl., etc.
133. Conclusions & Recommendations
136
˗ Easy to set up a hydrological model in a data scarce regions using wflow
˗ However, obtaining good forcing data remains an issue
˗ And model calibration with limited calibration data is a challenge
˗ Nevertheless, we were able to set up a good reservoir water balance and sediment
model, which formed the basis for developing measures for the reservoirs
˗ It would be nice to add some functionalities to the reservoir module
134. Other project in which we applied
wflow
Country
Vrbas river Croatia
Masterplan Welang river Indonesia
Rehabilitation of 4 reservoirs in Burkina Faso Burkina Faso
Backwater curve impact fence Rio Las Vacas Guatemala
Masterplan remining Gelado dam Brasil
Water balance of a Lagoon Ivory Coast
Disaster Risk Reduction Kazachstan Kazachstan
Water as leverage Cartagena Colombia
Questions?
137
136. Steven Rusli
wflow_sbm application for one-way model coupling with groundwater flow
model in Bandung, Indonesia.
PhD Student at Wageningen University
Steven Reinaldo Rusli obtained his bachelor's degree in the
Civil Engineering Department at Universitas Katolik
Parahyangan, Indonesia. His undergraduate thesis delved
into flood control in a tidal-influenced river on Borneo Island.
Following stints in engineering consultation and part-time
teaching, he pursued a double degree in Water Resources
Engineering from Universitas Katolik Parahyangan,
Indonesia, and Hohai University, China, graduating in 2015.
His master's thesis focused on assessing parameter
uncertainties in hydrological models across different
temporal variabilities. In 2019, he embarked on his PhD at
Wageningen University and Research, researching
groundwater flow modeling in the data-scarce Bandung
groundwater basin, a region heavily reliant on groundwater.
137. Hydrology Suite User Days – wflow and HydroMT
wflow_sbm application for one-way model coupling with groundwater
flow model in Bandung, Indonesia
140
138. • To understand the current situation in the designated study area
Motivation
Rusli, et al. (2021)
141
139. Motivation
• Bandung groundwater basin, Indonesia
▪ ±1700 km2 (two cities and three regencies)
▪ Steep slope (±2500 m ASL to ±650 m ASL)
▪ ±2000 mm of annual rainfall
▪ Discharge measurement at Nanjung ±75 m3/s
▪ Volcanic deposits (breccia, tuffs, and clays)
▪ Multi-layer aquifers (lake Bandung)
▪ ±11 million population by the end of 2023
▪ ±65% of domestic water supply from groundwater
▪ Rapid industrial development
▪ Estimated groundwater abstraction of 400Mm3/year
142
140. Motivation
• Estimating water balance components to determine water storage status
Data
Average value
(mm/day)
Precipitation
Rainfall station 5.40
SACA&D 6.47
TRMM 7.50
CHIRPS 7.80
Actual evaporation
GLEAM 3.13
ERA5 2.82
Discharge
Observation 3.65
GloFAS-ERA5 6.12
Groundwater abstraction estimates 0.57
Basin storage changes
‘lowest’ combination -4.42
‘highest’ combination +0.76
𝐼 − 𝑂 = ∆𝑆
Precipitation
Evaporation, river discharge,
groundwater abstraction
Storage (groundwater, surface
water, soil moisture)
143
143. One-way coupling of wflow_sbm and MODFLOW
Hydrological
simulation
using wflow_sbm
Groundwater flow
simulation
using MODFLOW
Model
parameterization
Model forcing Model results
CHIRPS rainfall
ERA5 PET
Groundwater
recharge
Groundwater
recharge
(Pedo)transfer
functions
with global dataset
Field-based
parameter estimates
Model transient
spin-up
Soil moisture
change
Basin-integrated
groundwater
budget
Groundwater
storage change
Water storage
change
(simulation)
GRACE dataset on
water storage
change
Basin-scale
groundwater
status assessment
Evaluation
comparison
Calibration data
River discharge
Groundwater
table elevation
data
One-way model coupling
Rusli, et al. (2023a)
146
146. Results: total water storage changes
Result: comparable response and behavior of water storage changes
Limitation: spatial scale & domain differences (±1700 km2 vs ±12000 km2)
Rusli, et al. (2023a)
149
147. On-going research: EWT data involvement
Rusli, et al. (2023b)
Rusli, et al. (2023b)
Rusli, et al. (2023b)
150
150. Take-home messages
▪ wflow_sbm flexibility to be applied in many areas
▪ The possibility of one-way coupling the wflow_sbm model with groundwater flow
model (MODFLOW)
▪ Opportunity to calibrate the one-way coupled model with satellite data (GRACE)
▪ Even further usage is feasible (aquifer interaction, future projection, etc)
153
151. References
▪ Rusli, S., A. Weerts, A. Taufiq, and V. Bense (2021). “Estimating water balance components and their uncertainty
bounds in highly groundwater-dependent and data-scarce area: An example for the Upper Citarum basin”. Journal
of Hydrology: Regional Studies 37, 100911. DOI: https://doi.org/10.1016/j.ejrh.2021.100911
▪ Rusli, S., V. Bense, A. Taufiq, and A. Weerts (2023a). “Quantifying basin-scale changes in groundwater storage
using GRACE and one-way coupled hydrological and groundwater flow model in the data-scarce Bandung
groundwater basin, Indonesia”. Groundwater for Sustainable Development 22, 100953. DOI:
https://doi.org/10.1016/j.gsd.2023.100953
▪ Rusli, S., A. Weerts, S. Mustafa, D. Irawan, A. Taufiq, and V. Bense (2023b). “Quantifying aquifer interaction using
numerical groundwater flow model evaluated by environmental water tracer data: Application to the data-scarce
area of Bandung groundwater basin, West Java, Indonesia”. Journal of Hydrology: Regional Studies. accepted
▪ Rusli, S., V. Bense, S. Mustafa, and A. Weerts (2024). “Assessing the significance of global climate model
projections and groundwater abstraction scenarios on future basin-scale groundwater availability”. Hydrology and
Earth System Sciences. In preparation.
154
152. Dirk Eilander
Flood hazard modelling with coupled wflow and SFINCS models, supported by HydroMT
Hydrologist at Deltares,
Assistant Professor at VU
Dirk Eilander is an expert in flood risk modeling. His work
focuses on improving flood risk management anywhere
globally in co-creation with stakeholders, through
development of models, methods, and datasets. He
obtained his PhD degree at the Vrije Universiteit on large-
scale compound flood risk modeling in coastal deltas.
Currently, he combines his work as hydrologist and lead-
developer of HydroMT at Deltares with an Assistant
professorship position at the Vrije Universiteit in
Amsterdam.
153. Dirk Eilander
Email: dirk.eilander@deltares.nl
Flood hazard modelling with
coupled wflow and SFINCS
models, supported by HydroMT
SOURCE: Janicki, J. (2019) Flood extent delineation in Mozambique after Cyclone Idai using Sentinel-1 data.
154. Flood hazard modelling - intro
157
• Over the past 50 years (1970-2019) floods have
caused an average of 6,600 deaths and 22.32
billion USD in losses per year (WMO, 2021)
• Understanding the flood hazard is key to
prevention, mitigation, preparedness and response
(Sendai Framework)
155. SFINCS
158
• Fast hydrodynamic computations:
− Bathtub models too simple
− Advanced models like D-Flow FM
module of the Delft3D FM Suite:
high detail, but too slow for quick
scan analysis for 1000km coast
− Alternative: reduced-complexity
solvers
• Local Inertial equations based on Bates
et al. 2010
• Added processes for coastal compound
flooding (e.g., wind drag, waves)
• Subgrid schematisation
• Spatially varying infiltration & roughness
Leijnse et al. (2021) – Coast Eng
157. Case study 1: Sofala Mozambique
Reuters
Reuters
BBC
158. wflow – SFINCS model coupling with HydroMT
161
• Reproducible sensitivity experiment
with wflow-SFINCS model chain
− Rainfall data (ERA5, Chrips)
− Land cover data (Globcover, VITO)
• Workflow:
− Build SFINCS model (incl river inflow
points)
− Build wflow model based on the SFINCS
domain
159. Sensitivity analyses with HydroMT
162
• Reproducible sensitivity experiment
with wflow-SFINCS model chain
− Rainfall data (ERA5, Chrips)
− Land cover data (Globcover, VITO)
• Workflow:
− Build SFINCS model (incl river inflow
points)
− Build wflow model based on the SFINCS
domain
− 4x Update wflow and run simulation
− 4x Update SFINCS and run simulation
(2 shown)
• Results:
− (obvious) large sensitivity to rainfall
product
160. FAIR modelling with HydroMT
163
https://github.com/DirkEilander/hydromt-wflow-sfincs/
166. Flood adaptation – rehabilitation of drainage system
169
BASE Canal rehabilitation Flood reduction
167. Case study 3 - Operational forecasting in Australia
170
170
Leijnse et al. (2022) – ICCE22 conference
168. Take home message
171
• Coupled wflow + SFINCS models, supported by HydroMT, provide a strong toolset for
(compound) flood hazard modeling
• With HydroMT, the model setup up:
− is reproducible
− is fast
− can be updated to facilitate model experiments (e.g., different forcing, land use)
− can easily be improved with local data
− is scalable
Dirk Eilander
Email: dirk.eilander@deltares.nl
169. Awad Mohammed Ali
Machine learning for predicting spatially variable lateral hydraulic conductivity:
a step towards efficient hydrological model calibration and global applicability
PhD Student at Wageningen University
Awad M. Ali, a Sudanese Civil Engineering professional,
obtained his Bachelor of Science from the University of
Khartoum in 2018 and completed a Master's program in
Earth and Environment at Wageningen in August 2023. In
the three years between his undergraduate and
postgraduate studies, Awad worked at the Water Research
Centre in Sudan, contributing to research in hydrological
drought and dams’ management in the Nile basin. His
expertise includes remote sensing in ungauged basins,
hydrological modelling, and machine learning integration.
Currently, he's pursuing a Ph.D., focusing on developing a
global dataset for transparent water management in
transboundary basins, building on his four-month internship
at Deltares this year.
170. Awad M. Ali
Supervisors (Deltares): Ruben Imhoff and Albrecht H. Weerts
Supervisor (WUR): Lieke A. Melsen
29 November 2023
Machine learning for predicting spatially variable lateral
hydraulic conductivity: a step towards efficient
hydrological model calibration and global applicability
Hydrology Suite User Days
Day 2 - wflow and HydroMT
173
171. (Pedo)transfer Functions (PTFs)
Imhoff, R. O., van Verseveld, W. J., van Osnabrugge, B., & Weerts, A. H. (2020). Scaling point-scale (pedo) transfer functions to seamless large-
domain parameter estimates for high-resolution distributed hydrologic modeling: An example for the Rhine River. Water Resources Research,
56, e2019WR026807. https://doi.org/10. 1029/2019WR026807
• PTFs are predictive functions that
estimate soil hydraulic properties
from structural soil data
• To lower the number of calibrated
global parameters
• A sensitive parameter without a PTF
is KsatHorFrac (fKh0)
(Wannasin et al., 2021)
(Imhoff et al., 2020)
KsatVer
KsatHor
𝐾𝑠𝑎𝑡𝐻𝑜𝑟 = 𝑓𝐾ℎ0 × 𝐾𝑠𝑎𝑡𝑉𝑒𝑟
174
172. Problem statement (cont.)
Page 3
van Verseveld, W. J., Weerts, A. H., Visser, M., Buitink, J., Imhoff, R. O., Boisgontier, H., Bouaziz, L., Eilander, D., Hegnauer, M., ten Velden,
C., and Russell, B.: wflow_sbm v0.6.1, a spatially distributed hydrologic model: from global data to local applications, Geosci. Model Dev.
Discuss. [preprint], https://doi.org/10.5194/gmd-2022-182, in review, 2022.
Challenges:
• Calibrate fKh0?
• Uniform fKh0?
• Large-scale model?
Available PTFs are still insufficient
(Araya et al., 2019)
fKh0
Soil
Properties PTF
wflow_sbm
(Deltares)
175
173. Objective & Research Questions (RQs)
Page 4
How accurately can ML algorithms predict fKh0 compared to a calibrated
benchmark?
RQ1:
Can ML-based fKh0 predictions improve wflow_sbm performance? How certain we
are?
RQ2:
How reliable and robust are ML-based fKh0 predictions if tested in a different
region?
RQ3:
To explore the possibilities of using Machine Learning (ML) algorithms in
estimating PTFs for fKh0 prediction
Main
Objective
176
174. Study area and data
Page 5
Data
Study Area
Great Britain
(1970 – 2015)
Loire Basin, France
(1978 – 2020)
Discharge CAMELS-GB Hydro platform
Precipitation CEH-GEAR E-OBS
Temperature &
Evapotranspiration
ERA5
Soil maps SoilGrids v1.0
• Great Britain: 551 subbasins
• Loire basin (117,000 km2): 559 subbasins
177
175. SoilGrids data in subbasin scale
Page 6
Step 1:
Rasterize wflow subbains
Step 3:
Average over layers
Trapezoidal rule
Organic carbon
content
(‰ or g Kg-1)
Sand (%) Clay (%) Silt (%)
Absolute depth
to bedrock (cm)
Bulk density
(kg/m3)
pH index
measured in water
solution (pH)
Organic carbon
content
(‰ or g Kg-1)
Sand (%) Clay (%) Silt (%)
Absolute depth
to bedrock (cm)
Bulk density
(kg/m3)
pH index
measured in water
solution (pH)
Step 2:
Average each layer over subbasins
SoilGrids 250m
Layer 1 Layer 1 Layer 1
Layer 1
Layer 1
Layer 1
178
176. Optimized 𝑓Kh0
Page 7
Duc, L. and Sawada, Y.: A signal-processing-based interpretation of the Nash–Sutcliffe efficiency, Hydrol. Earth
Syst. Sci., 27, 1827–1839, https://doi.org/10.5194/hess-27-1827-2023, 2023. 179
177. SoilGrids 250m
Soil attributes
Sand %
Silt %
Clay %
Org. carbon ‰
Bulk density
pH idx (water)
Bedrock depth
Optimized
(𝑓𝐾ℎ0)
ML-based PTFs
Training
(75%)
Test
(25%)
Random
Sampling
Standardization
Standardization
Mean
&
St.D.
Random Forest
(RF)
Boosted Regression
Trees (BRT)
Machine Learning
Algorithm
Parameter
Tuning
Prediction
Model Evaluation +
Feature Importance
5-Fold Cross
Validation
Page 8
Araya, S. N., & Ghezzehei, T. A. (2019). Using machine learning for prediction of saturated hydraulic
conductivity and its sensitivity to soil structural perturbations. Water Resources Research, 55(7), 5715-5737. 180
178. Tree-based models
Page 9
(a) Random Forest (RF) (b) Boosted Regression Trees (BRT)
Error
Iterations
+
+
+
RF: Breiman, L. (2001). Random forests. Machine learning, 45, 5-32.
BRT: Friedman, J. H. (2002). Stochastic gradient boosting. Computational statistics & data analysis, 38(4), 367-378. 181
180. Objective & Research Questions (RQs)
How accurately can ML algorithms predict fKh0 compared to a calibrated
benchmark?
RQ1:
Can ML-based fKh0 predictions improve wflow_sbm performance? How certain we
are?
RQ2:
How reliable and robust are ML-based fKh0 predictions if tested in a different
region?
RQ3:
To explore the possibilities of using Machine Learning (ML) algorithms in
estimating PTFs for fKh0 prediction
Main
Objective
183
183. Objective & Research Questions (RQs)
Page 14
How accurately can ML algorithms predict fKh0 compared to a calibrated
benchmark?
RQ1:
Can ML-based fKh0 predictions improve wflow_sbm performance? How certain we
are?
RQ2:
How reliable and robust are ML-based fKh0 predictions if tested in a different
region?
RQ3:
To explore the possibilities of using Machine Learning (ML) algorithms in
estimating PTFs for fKh0 prediction
Main
Objective
186
184. Study area and data
Data
Study Area
Great Britain
(1970 – 2015)
Loire Basin, France
(1978 – 2020)
Discharge CAMELS-GB Hydro platform
Precipitation CEH-GEAR E-OBS
Temperature &
Evapotranspiration
ERA5
Soil maps SoilGrids v1.0
• Great Britain: 551 subbasins
• Loire basin (117,000 km2): 559 subbasins
187
188. Conclusion
• ML-based PTFs obtained from RF and BRT effectively predicted fKh0 and RF slightly
outperformed BRT
• Two globally seamless fKh0 maps were generated and are readily available for use
• PTFs notably enhanced wflow_sbm performance compared to the uncalibrated scenario
(fKh0 = 100), increased the median KGE from 0.55 to 0.75
• The utilization of RF and BRT improved performance in around 75% of subbasins within
the Loire basin in France, resulting in a KGE that on average 0.06 higher
191
189. Hydrology Suite User Days
Day 2 - wflow and HydroMT
Thank you
for your attention