Ph.D. dissertation · Jackson State University · defense February 10, 2027
Accessing, Refining, and Combining Open Earth Data for a Modular, Parsimonious, Physics-Based Hydrological Model
Physics-based hydrological models can, in principle, predict floods anywhere, because conservation of mass and momentum holds in every basin. In practice they are used mostly where floods are already observed: their parameters are tuned to past hydrographs, their inputs must be gathered by the user, and their physics is fixed when they are built. This dissertation asks whether open Earth data, made accessible, placed at the local scale, and read for interacting hazards, can feed an open, modular, parsimonious model that runs from an outlet alone and still predicts local floods.
Advisor
Prof. Rocky Talchabhadel (Chair)
Committee
Dr. Nawa Raj Pradhan (ERDC), Dr. Sadik Khan, Dr. Kejun Wen, Dr. Robert W. Whalin
Study regions
Mississippi River Basin and Gulf Coast; central Nepal Himalaya
Status
Proposal defended October 2026; final defense February 10, 2027
Figure 1Graphical abstract of the dissertation.
The problem and the idea
Data that speak, physics that translates.
Physics-based models and open Earth data fail in mirror image. Models carry mechanism but lack coverage; open data carry coverage but lack mechanism. Each is the other's missing half, yet they are rarely joined, and the reason is easy to miss: calibration. A model tuned to observed discharge absorbs the errors of its inputs in its parameters, so it cannot tell good data from bad, and it needs a discharge record that the places of greatest flood exposure lack.
So give the model only an outlet. Its inputs come from open Earth data, made accessible, placed at the local scale, and read for interacting hazards. It is open and modular, so its users can run the physics provided or test their own. It is parsimonious: few parameters, each starting from open data or a published equation, with calibration kept for the few that need it. With little to tune, errors in the data surface in the hydrograph instead of hiding in the parameters, and the model becomes an instrument for judging the data that drive it.
In the Kathmandu Valley the same untuned model simulates eight floods with a median Nash–Sutcliffe efficiency of 0.54 when driven by raw satellite rainfall and 0.81 when driven by ground stations. The data speak, and the physics shows which to believe.
Not every analysis feeds the model. Each result below carries a tag:
Model inputShapes the modelIndependent finding
Two regions in every content chapter
The Mississippi River Basin and Gulf Coast are data-rich: dense USGS streamflow and well networks, radar rainfall, airborne soil-moisture campaigns, and tide gauges let every refined product and every model run be checked against independent observations. The Nepal Himalaya is data-sparse and steep: the Bagmati River above Khokana drains the Kathmandu Valley (606 km²) and is monitored by 15 long-record daily stations of Nepal's Department of Hydrology and Meteorology (Figure 2).
Figure 2Study areas: (a) Mississippi River Basin and Gulf Coast; (b) central Nepal Himalaya.
Chapter 1 · Introduction
Access
Understanding the water balance through open Earth data
What does each dataset say on its own, and who can hear it? Two new indices of rainfall timing, a 76-year heatwave record, and a platform that opens the data to people who do not program.
1.6.1When the wettest days arrive
Published · ESS 2026
Independent findingShapes the model: antecedent wetness
Standard indices measure how much rain falls and how intense it is, but not when the wettest days arrive, and clustered wet days fall on soil the previous storm left wet. I defined two families of indices for a share of the year's rain: the fewest days that hold it (intensity, WD) and the shortest run of consecutive days that holds it (timing, WDR), computed them from ERA5 worldwide for 1980–2024, and tested them against gauges (Figure 3).
Timing changes independently of intensity: the Amazon reaches its annual total in 15–18% fewer days, while eastern North America needs 12–25% more. Monsoon regions cluster their wettest days; mid-latitudes scatter them.
Gauges confirm the timing: half of the annual rain falls within 54 consecutive days at 15 stations in the Kathmandu Valley and within 114 days at 16 NOAA stations in Mississippi.
Gauges do not confirm the amounts: ERA5 gets the windows right but gives 64% too much rain in the Kathmandu Valley, the first sign of the bias corrected in Chapter 2.
Figure 3Chronological clustering indices WD25R, WD50R, and WD75R from ERA5: 1980–2024; 1980–1999; 2000–2024.
Heat drives the demand side of the water balance, but heatwave records are often short, regional, or defined inconsistently. A heatwave day here is one whose daily maximum temperature exceeds its day-of-year 90th percentile, computed from 76 years of ERA5 (1950–2025) and summarized for the 44 IPCC AR6 land regions (Figure 4; Animation 1).
Heatwave months increased in all 44 of 44 land regions, by 0.26 to 1.20 months per decade. This is heat alone; Chapter 2 asks what happens when rainfall fails at the same time.
Figure 4Trend in heatwave months, 1950–2025 (ERA5).Animation 1Heatwave months per year in 10-year windows, 1950–2025, with the land-mean series (ERA5).
1.7Data for everyone: the GeoClimate Intelligence Platform
Published · EMS 2026 · ASCE 2026
Independent findingModel input: same data catalog
Cloud platforms removed the cost of storage and computation, but not the need to program. The GeoClimate Intelligence Platform puts Google Earth Engine behind a no-code web interface. The published papers describe its first five modules; it has since grown to nine tools, including snow depth and soil moisture (Figure 5).
1,042 datasets and 23 climate indices without code.
Gulf Coast (DAYMET, 1980–2023): the wettest day intensifies by 0.44 mm per year and the longest dry spell shortens by 0.39 days per year; design values come directly, 107 mm for the 10-year and 154 mm for the 100-year one-day rainfall.
Access is not trust: maximum temperature transfers from grid to station (R² 0.96 to 0.99 at ten airports), daily rainfall does not (R² 0.41 to 0.68), and DAYMET underestimates the days above 100 mm.
Sixteen undergraduates used the platform in Advanced Water Resources Engineering (Fall 2025).
Figure 5Climate Intelligence Hub: index time series with trend tests; index maps.
1.8Unlocking published methods: SERVES and GLOBAL-SAM
In progress
Model input: SERVES soil deficit
SERVES root-zone soil moisture (Pradhan, 2019; Pradhan et al., 2020) and GLOBAL-SAM snow depth are algorithms from the U.S. Army Engineer Research and Development Center. I implemented both as Google Earth Engine applications inside the platform and evaluated SERVES against SMAP Level 4 and 2,643 in situ stations (ISMN, SCAN, SNOTEL, USCRN, and NEON), asking where it works (Figure 6).
Useful where vegetation controls soil water: tropical climates (median r 0.64), polar sites (0.55), and the growing season (JJA 0.33).
Not in cold (about −0.2) and temperate (0.03) climates outside the growing season, where SMAP L4 remains skillful.
Next: GLOBAL-SAM snow depth against SNOTEL.
Figure 6Median correlation with in situ root-zone soil moisture, SMAP L4 and SERVES: by climate zone; by season.
Chapter 2 · Data science
Refine
Making coarse data useful: mass-conserving downscaling, bias correction, and compound hazards
Coarse data become local in three ways: downscale under an exact conservation constraint, correct with ground observations for a stated use, and combine hazards into compound and cascading records. Every method is judged against data it never saw.
2.2.1Groundwater storage from GRACE, 55 km to 5 km
Manuscript in preparation
Independent finding: groundwater for aquifer managementModel input: baseflow check
GRACE is the only observation of groundwater from space, but at about 55 km, and downscaled fields often create or lose water. An explainable ensemble of gradient-boosted trees (XGBoost, LightGBM) and a neural network learns how terrestrial water storage relates to its drivers; applied at 5 km, 121 cells per GRACE cell, a residual correction makes the fine field average exactly to GRACE (Figure 7; Animations 2 and 3).
Held-out test against GRACE: R² = 0.95.
Against data GRACE never saw: High Plains aquifer storage change matches USGS estimates (r = 0.84), and the drought drawdowns of 2005–2006 and 2012 in the Mississippi Embayment are visible. Validated against 825 groundwater wells.
Where it stops working. Only about one third of the long-term High Plains decline is captured, because pumping is not a predictor, and inside a single GRACE cell the detail adds little skill against individual wells (r ≈ 0.02). The product is right at the scale where aquifers are managed, and its limits are known.
Figure 7Monthly mean groundwater storage anomaly, Mississippi River Basin, downscaled to 5 km.
Animation 2Monthly GRACE storage anomaly over the Mississippi River Basin at its native 55 km grid, with the basin mean; the 2012 drought stands out.Animation 3GRACE at 55 km beside the downscaled 5 km field, month by month; averaged back to 55 km, the fine field is exactly GRACE.
2.2.2Soil moisture from SMAP, 9 km to 1 km, and what it adds inside the pixel
Analysis complete
Shapes the model: storms need physicsModel input: soil-moisture state
Daily 1 km soil moisture that preserves SMAP exactly: each 9 km pixel holds 81 cells of 1 km, and the model predicts each cell's departure from its parent, not the parent itself. The test that matters is whether the detail inside the pixel is real (Figure 8; Animation 4).
By the usual test the product works: r ≈ 0.8 against Sentinel-1-based labels.
Against radar-free independent data, r ≤ 0.35: agreement is not accuracy.
The best lasting within-pixel pattern among the products tested, but a storm response of only 3 to 10%, where a rain-forced model reaches 40 to 50%. Learned maps carry patterns; storm dynamics need physics.
Figure 8The daily 1 km soil moisture product: one morning; SMAP at 9 km; 1 km median; anomaly by land use; irrigated fraction; interval width.Animation 4Daily 1 km soil moisture (latest pass within 3 days), with a zoom on central Iowa where every 1 km cell is visible.
2.3Satellite rainfall that lost its gauges
Manuscript in preparation
Model input: rainfall for every Kathmandu run
IMERG V07 over the Kathmandu Valley became too wet after 2016. Three records are combined to find out why and to correct it: IMERG (30 min, 1998–2025), 15 DHM stations with reliable daily totals, and nine S4W automatic weather stations (10 to 15 min, 2018–2025) (Figure 9).
The DHM stations see 78% of IMERG's rainfall before the break and 55% after; three homogeneity tests agree.
The GPCC gauges behind IMERG's adjustment here fell from about 50 to none; on the same days the S4W stations give a ratio of 0.60 and the DHM stations 0.57. The satellite did not change; its gauges disappeared.
Volume correction keeps bias within ±4% in every period; peak correction lowers the 30-minute heavy-rain error at independent S4W stations from 54% to 40%.
First flood test with the model unchanged, eight floods: median efficiency 0.54 with raw IMERG, 0.62 with IMERG scaled to DHM totals, 0.81 with the S4W stations.
Figure 9IMERG V07 and DHM stations, Kathmandu Valley: seasonal rainfall; step after 2016; double-mass curve.
2.4Compound heatwave–drought
Manuscript in preparation
Independent findingModel input: antecedent states (proposed)
A global record of months that are both heatwave months and drought months (SPEI), 1950–2025, and how water storage in the Mississippi River Basin responds to them (Figure 10; Animation 5).
Compound months more than doubled, from 0.80 per year (1950–1979) to 1.70 (1996–2025), with a significant rise over 65% of land and the fastest rise in the driest regions.
The shifts come in order: heatwaves in 1986, compound months in 1996, drought in 2000. Heat increasingly comes first and dries the land before rain fails.
Mississippi River Basin (in progress): a compound month is followed by a storage loss of 0.81 cm, a drought-only month by 0.41 cm, about twice as much with slower recovery; 74% of warm-season drought months there are also heatwave months.
Figure 10Compound heatwave–drought months: mean annual frequency, 1950–2025; change between the first and last 30 years.Animation 5Compound heatwave–drought months per year in 10-year windows, 1950–2025, with the land-mean series.
2.5Compound flooding from storm surge, rainfall, and rivers
Manuscript in preparation
Independent finding: attribution to sea levelShapes the model: sea-level boundary
At 79 NOAA tide gauges (tide removed), 1950–2025, joint exceedance of the 99th percentiles of surge and rainfall is compared with what independence would give, the trend is attributed, and what follows downstream is traced (Figure 11; Animation 6).
Joint extreme surge and rainfall occur 6.9 times more often than under independence, significant at 78 of 79 gauges; on the Gulf Coast more than half are near a tropical cyclone.
Compound flood days rose from 1.03 to 1.79 per year; with sea level held at its earlier level there is no increase. Sea-level rise, not changing storm dependence, explains the rise.
The day after an extreme surge, river floods are twice as likely as chance for about a week: compound becomes cascade, with Hurricane Florence as the example.
Figure 11Compound flooding at U.S. tide gauges: joint extremes against chance; change in flood days, observed and at 1950–1987 sea level.Animation 6Hurricane Florence, September 2018: surge at the tide gauges, rain, and the Cape Fear River rising days later at Wilmington, North Carolina.
Chapter 3 · Core
Translate
MRRpy, an open, modular, parsimonious hydrological model: governing equations, numerical solution, and verification
An outlet in, a flood out. MRRpy, developed in this work, combines Green–Ampt infiltration with recovery, variable source area, and impervious runoff without double counting, and routes water with schemes derived from the Saint-Venant equations, the implicit one solved exactly in linear time on the drainage tree. Every parameter starts from open data or published physics (Figure 12).
Figure 12MRRpy concept: Earth observation to model grid; runoff generation in every cell; routing on the D8 tree; outlet hydrograph.
3.2–3.4Runoff physics and routing you can switch
Complete
One setting swaps the physics. Runoff comes from infiltration excess (Green–Ampt with soil recovery between storms), saturation excess (a variable source area that grows as the divide store fills), and impervious surfaces, alone or together. Routing uses a kinematic wave, Muskingum–Cunge, or a semi-implicit diffusive wave, and since version 0.3.0 a local-inertial scheme (Figure 13; Animation 7).
September 2024, Bagmati: infiltration excess made 48% of runoff, impervious surfaces 28%, saturation excess 23%; the saturated area grew from 0.6% to 36% of the basin, then contracted.
Blue River 2018, nothing tuned: saturation excess alone gives the best overall skill (KGE 0.68), all three mechanisms together the best peak (−29%). Different basins, different dominant physics.
The same runoff routed with three schemes gives peaks that differ by a factor of almost three.
Figure 13The same storm routed with three schemes (MRRpy documentation example).Animation 7Saturated area (blue) spreading across the Bagmati basin in the September 2024 storm, with rainfall and the saturated share of the basin.
3.5An exact linear-time tree solver
Complete · method contribution
On a D8 grid every cell drains to exactly one neighbor, so the implicit diffusive scheme's system of equations lives on a tree. Eliminating from the leaves to the outlet and substituting back solves it exactly, in time proportional to the number of cells, stable by construction and with water conserved (Figure 14).
Matches a dense matrix solve on 150 random trees to 4 × 10⁻¹⁶; the water ledger closes to a median 3 × 10⁻⁶ of input over 14 basin runs.
Five years at Khokana: 31.6 million kinematic steps (5 s) against 1.4 to 2.0 million implicit steps (mean 77 to 115 s); the GPU speeds the kinematic scheme 34 times at 1.38 million cells.
Figure 14The implicit solver: (a) elimination and substitution on the D8 tree; (b) run time against grid size.
3.6Verification against analytical solutions
Nearly complete
Test
Result
Tree solver against a dense matrix solve, 150 random trees
error 4 × 10⁻¹⁶
Mass balance on a valley DEM
error 4 × 10⁻¹⁶
Steady flow against Manning normal depth
0.0% difference
Linear diffusion wave (Hayami)
peak within 7%, timing exact
Kinematic wave on a plane; flood front in a dry channel
analytical hydrograph and arrival
Green–Ampt recovery against the EPA SWMM engine
matches
Automated test suite, Python 3.9 to 3.12
1,893 tests pass
Tilted V-catchment
proposed
3.7Observed floods without event-wise fitting
In progress
A parsimonious model fed by open data has to work in a steep, monsoon, partly urban basin with sparse data (the Bagmati at Khokana) and in a mild, rural basin with rich data (the Blue River in Oklahoma), before any tuning (Figures 15 and 16).
Bagmati: median Nash–Sutcliffe efficiency 0.81, all eight peaks within 30%, one to two minutes per six-day run on a CPU.
Blue River: KGE 0.68 untuned. On this mild slope routing matters as much as runoff (all mechanisms: NSE −1.17 kinematic, 0.43 implicit diffusive).
What it does not do yet. At the Blue River, peaks arrive 25 to 30 hours early in every configuration, and the observed volume includes baseflow the model does not have yet. A linear groundwater store fed by recharge from the saturated zone is being added.
Figure 15Eight largest floods with complete records, Bagmati River at Khokana: observed and MRRpy.Figure 16Blue River, 2018: observed and simulated discharge; skill of every mechanism combination; routing scheme; rainfall interpolation.
3.8An open companion to GSSHA
In progress
I built and tested a GSSHA model of the Bagmati above Khokana as an established physically based reference. Both models route the same floods with the same rainfall and terrain; hydrographs, peaks, timing, and volumes are being compared event by event. The aim is not to replace GSSHA but to offer an open, fast companion that agrees with it where both apply.
GSSHA
MRRpy
Source
free, closed
open, MIT license
Interface
WMS (paid)
Python, command line, notebook, QGIS
Setup
project built by hand
an outlet and one configuration file
Data
supplied by the user
Earth Engine when local data are missing
Physics
comprehensive (groundwater, 2-D overland flow)
narrower, switchable by one setting
Solvers
established schemes
GPU explicit or exact implicit tree solver
3.11A model anyone can start
Complete
A model nobody can run changes nothing. pip install MRRpy, then a complete run is three lines of Python given an outlet. The documentation has a user manual and 11 worked examples, each shown four ways (notebook, command line, Python, QGIS), and a free six-module course follows the model's pipeline in the browser. A software paper is drafted for the Journal of Open Source Software (Figure 17; Animation 8).
Figure 17The MRRpy documentation at mrrpy.readthedocs.io.Animation 8The free interactive course: a D8 flow-direction explorer that runs in the browser.
Model applications with ground, satellite, and refined data in the Himalaya and the Mississippi River Basin
Where does open-data flood modeling with few parameters work, and where does it fail? Each use case is run with ground, raw Earth Engine, and corrected data, to measure how much each source changes the answer.
4.2An observed flood, at the outlet and along the way
In progress
For 30–31 July 2024, with S4W rainfall, MRRpy produces hydrographs at five DHM station locations as well as at the outlet. Tributaries peak first and the flood builds downstream (Figure 18; Animation 9).
At Khokana the Nash–Sutcliffe efficiency is 0.86, and the peak is 923 m³/s against 813 observed: 14% high and one hour late, with the right shape.
Figure 1830–31 July 2024: rainfall; hydrographs at five DHM station locations; observed at Khokana.Animation 9MRRpy routing the 31 July 2024 flood through every river cell to Khokana (log-scale flow), with observed and simulated discharge.
4.3Design floods from return-period rainfall
In progress
Generalized extreme value fits at the 15 DHM stations give 24-hour design rainfall of 78 to 194 mm for the 2- to 100-year storms; one run per return period gives design hydrographs at every station location (Figure 19).
Peaks at Khokana range from about 900 to 2,400 m³/s.
Corrected IMERG matches gauge design rainfall up to about the 10-year return period, and the September 2024 storm exceeded 11 of 14 station 100-year estimates.
Point rainfall is applied without areal reduction, so present peaks are upper estimates.
Figure 1924-hour design storms through MRRpy: hydrographs at Khokana; peaks at five locations.
4.4Five years, storms and dry spells
In progress
Because the soil dries and recovers between storms, a five-year run takes a few hours on a CPU with the implicit solver (Figure 20).
Driven by the S4W stations, it follows the floods and the seasonal cycle with a daily correlation of 0.88.
Between floods it drains to almost nothing: there is no baseflow yet, which is why the groundwater store comes next.
Figure 20Daily flow at Khokana, 2020–2024: observed and MRRpy; seasonal cycle; flow-duration curve.
4.5Floods without local rain: the 2026 outburst
Exploratory
For the glacier-triggered surge of 26 August 2026, the inflow was routed down the Bhote Koshi and Trishuli rivers as an upstream boundary (Figure 21; Animation 10).
The modeled front arrives at Syabrubesi, Betrawati, and Malekhu after 27, 52, and 175 minutes, against 23, 43, and 193 minutes reported.
This is exploratory: the released volume is assumed and roughness calibrated. Next is the 2020 Pearl River flood at Jackson, Mississippi, with the reservoir release as inflow.
Figure 21Surge arrival along the Trishuli River: reported and modeled.Animation 10The 26 August 2026 outburst surge entering as an upstream boundary and travelling down the Bhote Koshi and Trishuli, with modeled and reported arrival.
4.6–4.7Forecast rain, and the Mississippi River Basin
Proposed
Forecast rainfall through an interface, toward early warning: how early a warning is possible, and how wrong at each lead time. Weak skill of global forecasts at valley scale is a known risk and will be reported as a result.
Chapter 2 products as model inputs on gauged tributaries of the Mississippi River Basin: radar, gauge, and satellite rainfall; downscaled soil moisture and groundwater as initial states; floods after compound heatwave–drought months.
Chapter 5 · Conclusions and future work
Extend
Sediment, sea-level boundaries, and machine-learning surrogates
Prototypes within the dissertation; full development after it.
Sediment
Floods carry more than water
A terrain reconstruction released after the 2026 outburst, from stereo satellite imagery, shows at least 3.1 million m³ eroded and 0.8 million m³ deposited along the Bhote Koshi: a rare test for a sediment-continuity term in each routing cell.
Sea level
The coast pushes back
A tide-plus-surge water level at the outlet would let the model simulate what compound and cascading floods do to river levels. The implicit diffusive scheme already carries backwater; the new piece is the boundary itself, tested first on a Gulf Coast river.
Surrogates
Access at the start, access at the end
The design and long-term runs are training data. A surrogate that learns rainfall and soil state in and hydrograph out can be served in the GeoClimate platform for each study area, so a user gets a flood estimate in seconds without running a model.