Work · Research · 14
Rainfall Estimation via Heterogeneous Data Fusion
Fusing ground sensors, radar and satellite imagery into one signal, and beating the industry baseline.
Private or proprietary. No public source link.
- Type
- Research
- When
- IRJMETS · Vol 7 Issue 3 · March 2025
- Role
- Co-author. Model design, fusion architecture and evaluation. Peer-reviewed and published.
- Scale
- Three heterogeneous sources fused · published IRJMETS Vol 7 Issue 3
In plain language
Measuring rain sounds simple and is not. Rain gauges are accurate but there are very few of them. Radar covers more ground but is biased. Satellites see everywhere and resolve almost nothing. Each one is wrong in a way the others are not, so this work combines all three into a single estimate, using a second model that learns which source to believe when they disagree. It beat the interpolation method the field treats as standard, and it matters most in exactly the places where flood warnings currently fail: the gaps between gauges.
The question
Three rainfall sources each fail differently and none is reliable alone. Can a fusion model outperform the standard interpolation baseline, particularly in the gaps between sensors?
Rain gauges are accurate and sparse. Radar is dense and biased. Satellite imagery covers everything and resolves nothing. Each one is wrong in a way the others are not.
The approach
A Hierarchical Probabilistic Ensemble Classifier fusing three sources (ground sensor networks, weather radar and Meteosat satellite imagery) into a single precipitation estimate, with a meta-learner stacking layer resolving disagreement between the base estimators.
The result
Probability of detection of 0.58, outperforming Kriging with External Drift, the industry-standard interpolation baseline, on the same evaluation set.
Why it matters beyond the number
Gauge networks create blind spots, and the blind spots are where flood early-warning fails. A fusion model that holds up between gauges is worth more operationally than a marginally better model that only works where a sensor already sits.
Status
Peer-reviewed and published in IRJMETS, March 2025. There is no public repository for this one. The write-up is the paper.
How it is put together
Left to right, the path a request or a record takes through the system.
01
Sources
- Ground sensor networks: accurate, sparse
- Weather radar: dense, biased
- Meteosat imagery: total coverage, coarse
Each accurate where the others are not.
02
Align
- Spatiotemporal alignment
- Common grid and time base
Nothing fuses until it shares a grid and a clock.
03
Base estimators
- Random Forest per source
- Source-specific error characteristics
04
Meta-learner
- Stacking layer over base outputs
- Learns which source to trust, and where
The fusion step proper.
05
Evaluate
- Probability of detection
- Against Kriging with External Drift
What happens, step by step
In plain language, in the order it happens.
Take three sources that disagree
Gauges, radar and satellite imagery, each with a different and well-understood failure mode.
Put them on one grid and one clock
Heterogeneous fusion is mostly alignment. Sources with different resolutions and cadences cannot be combined until they share a spatiotemporal frame.
Estimate from each source separately
Random Forest base estimators, so each source's characteristic error is learned rather than assumed away.
Learn who to believe
A meta-learner stacking layer sits over the base estimators and resolves their disagreement, which is the actual fusion, and the reason this outperforms averaging.
Test where it matters
Evaluated by probability of detection against Kriging with External Drift, the industry-standard interpolation baseline, on the same set.
Decisions, and what they cost
Every choice worth recording has a road not taken. Both are here.
A stacked meta-learner
instead ofWeighted averaging of the three sources
Fixed weights assume each source's reliability is constant across space and time, which is precisely what is not true. Radar bias and gauge sparsity are both local. Learning the arbitration is the difference between fusion and blending.
Probability of detection as the headline metric
instead ofRMSE
For flood early warning the operational question is did it catch the event, not how close the millimetre estimate was. Optimising the wrong metric produces a model that looks better and warns worse.
Compare against Kriging with External Drift
instead ofA naive baseline
Beating a weak baseline proves nothing. KED is what the field actually uses, so it is the only comparison that carries information.
What came out of it
- Probability of detection of 0.58, outperforming Kriging with External Drift on the same evaluation set.
- Peer-reviewed and published in IRJMETS, March 2025.
- Holds up between gauges, which is operationally worth more than a marginally better model that only works where a sensor already sits.
Still open
- No public repository: the paper is the artefact.
- Evaluation is retrospective. Operational value would need testing against live warning thresholds, not a held-out set.
Built with
- Python
- Random Forest
- Meta-learner stacking
- Spatiotemporal analysis
- Meteosat