Sparse, noisy labels and inputs
Drought labels depend on indices that are themselves estimates. The pipeline had to treat missingness as information rather than silently interpolating everything away.
RESEARCH / AI / DATA
Drought Prediction
Spatiotemporal drought forecasting from climate, satellite and soil-moisture data.
Academic research project.
- Year
- 2024
- Status
- Research
- Role
- Designed the data pipeline, model experiments and evaluation for spatiotemporal drought prediction.
- Categories
- ResearchTime SeriesDeep LearningGeospatial ML
01
The problem
Drought develops slowly and unevenly across space. Early warning matters for agriculture and water planning, but ground observations in Northern Nigeria are sparse, irregularly updated and noisy.
The modelling problem is to predict moderate and severe drought conditions ahead of time from heterogeneous inputs — climate variables, satellite-derived indices and soil moisture — when the observation network itself is unreliable.
02
Context
Academic research focused on spatiotemporal drought prediction for Northern Nigeria, using climate, satellite and soil-moisture features under sparse-data conditions.
03
My role
- Built the feature and label pipeline from climate, satellite and soil-moisture sources into model-ready spatiotemporal tensors.
- Designed and ran model experiments for drought class prediction, with emphasis on robustness to missing and noisy observations.
- Defined evaluation focused on operational usefulness: discrimination of moderate and severe drought, not just aggregate accuracy.
04
The solution
A spatiotemporal modelling approach that learns spatial structure and temporal evolution of drought indicators rather than treating grid cells independently.
Feature design that tolerates sparse inputs: masked observations, satellite proxies and lagged climate variables combined so the model degrades gracefully where ground data is thin.
Evaluation reporting class-wise discrimination — the operational question is whether moderate/severe drought is separable enough to warn on.
06
Architecture
- 01Data sources
- 02Feature pipeline
- 03Labels
- 04Models
- 05Evaluation
07
Implementation
Data alignment across sources with different resolutions and cadences was the first substantial engineering task — spatial joins, temporal interpolation and explicit missing-data masks.
Models were experimentally compared rather than assumed; the selection criterion was performance on moderate and severe drought classes under sparse conditions.
Results are reported with dataset and split context. Headline discrimination for moderate and severe cases reached approximately 0.85 AUC under sparse, noisy conditions — a research result, not an operational forecast service.
08 / Hard parts
The hard engineering.
Spatiotemporal structure vs. independent cells
Predicting each location independently ignores how drought propagates. Incorporating spatial context without overfitting a noisy observation network was the central modelling trade-off.
Honest evaluation under class imbalance
Severe drought is rarer than normal conditions. Aggregate accuracy is misleading; class-wise metrics and AUC on the specified split were used so the reported number means something.
09
Results
Research demonstrated discrimination of moderate and severe drought conditions at approximately 0.85 AUC on the specified evaluation split under sparse, noisy observational conditions. Work remains research-stage: no operational forecasting service, no client deployment.
≈0.85
AUC (moderate & severe)
Specified holdout, sparse/noisy conditions
3
Source families
Climate · satellite · soil moisture
10
Proof & technology
Evidence
Research
Evaluation metrics and methodology
Case study
Pipeline and modelling write-up
Technology
PythonDeep learningTime seriesGeospatial MLRemote sensing featuresPyTorch
Start a project
Building something difficult?
If this kind of system is close to your problem, tell us what you are trying to ship.