# Gurara Reservoir Research Implementation Report

## Study

**AI-Driven Forecasting of Reservoir Surface Water Dynamics Using Multi-Sensor Earth Observation, Climate Variables, and Machine Learning: Gurara Reservoir, Nigeria**

## Implementation Status

The reproducible implementation covers reservoir-boundary derivation, Landsat
and Sentinel-2 preprocessing, NDWI/MNDWI/AWEI extraction, monthly surface-water
area, climate integration, trend analysis, statistical and machine-learning
forecasting, explainability, uncertainty, cross-sensor validation, and
decision-support reporting.

## Reservoir Boundary

Water extraction is constrained to a `55.837 km2` multi-year recurring-water
footprint derived from JRC Global Surface Water. The original `715.865 km2`
rectangle is retained only as the broader study region.

## Data Quality

- Monthly Landsat records: `304`
- Quality-approved water-area months: `196`
- Missing/failed fraction: `35.5%`
- Surface-water range: `0.000` to `55.462 km2`
- Latest valid month: `2026-04-01`

## Water-Index Validation

Cross-sensor validation recommends **MNDWI**:

- Valid Landsat-Sentinel-2 overlap months: `73`
- Correlation: `0.540`
- MAE: `3.463 km2`
- Bias: `-0.967 km2`

This is cross-sensor validation, not independent field ground truth. The
Dynamic World threshold-validation workflow is included for an additional
satellite-reference assessment.

## Trend Results

- Analysis period: `2001-01-01` to `2026-04-01`
- Mann-Kendall tau: `0.431`
- Mann-Kendall p-value: `3.19e-19`
- Sen slope: `1.817 km2/year`

The positive long-term trend includes reservoir formation after 2001 and must
not be interpreted solely as recent climate-driven expansion.

## Forecast Model Comparison

|   horizon_months | model                |   rmse |   mae |    nse |   kge |   observation_count |
|-----------------:|:---------------------|-------:|------:|-------:|------:|--------------------:|
|                1 | sarima               |  8.014 | 5.393 |  0.075 | 0.516 |                  52 |
|                3 | sarima               |  9.648 | 7.011 | -0.341 | 0.317 |                  52 |
|                6 | seasonal_persistence |  9.836 | 5.029 | -0.292 | 0.088 |                  42 |
|               12 | persistence          |  9.836 | 5.029 | -0.292 | 0.088 |                  42 |

Longer-horizon negative NSE values indicate limited predictive skill relative
to a mean benchmark. These results should be reported directly rather than
overstating AI performance.

## Operational Forecast

Forecast origin: `2026-04-01`. Values are constrained to
the reservoir footprint and include 95% SARIMA prediction intervals.

|   horizon_months | date       |   forecast_water_area_km2 |   lower_95_km2 |   upper_95_km2 |
|-----------------:|:-----------|--------------------------:|---------------:|---------------:|
|                1 | 2026-05-01 |                    41.89  |         29.449 |         54.331 |
|                3 | 2026-07-01 |                    42.992 |         26.009 |         55.837 |
|                6 | 2026-10-01 |                    53.645 |         34.536 |         55.837 |
|               12 | 2027-04-01 |                    40.936 |         19.735 |         55.837 |

## Advanced Model Decision

Do not claim LSTM/GRU/Transformer results as reliable. The validated monthly sample is too sparse for high-capacity sequence models; retain SARIMA, persistence, Random Forest, and XGBoost as the defensible comparison.

- Valid months: `196`
- Missing fraction: `35.5%`
- Longest contiguous valid run: `23` months

This data-driven gate follows the proposal's warning that Transformer models
may overfit a short or incomplete monthly series.

## Main Limitations

1. Surface-water area is not storage volume.
2. Dynamic World and cross-sensor validation are not field ground truth.
3. Missing optical observations required quality filtering and limited interpolation.
4. Dam operations, abstraction, discharge, sedimentation, and measured depth are unavailable.
5. Forecast skill is weak at some longer horizons.
6. SWOT remains supplementary because its record is recent and coverage must be assessed separately.

## Reproducible Outputs

- Processed datasets: `outputs/processed`
- Validation: `outputs/validation`
- Trend and climate analysis: `outputs/analysis`
- Models and comparison: `outputs/models`
- Operational forecasts: `outputs/forecasts`
- Figures: `outputs/figures`
- Decision-support report: `outputs/reports/gurara_decision_support_report.md`

## Conclusion

The implementation delivers a complete, reproducible first research system for
Gurara Reservoir surface-water monitoring and forecasting. The strongest
scientific result is the validated long-term Earth Observation time series and
its trend/seasonal characterization. Forecasts provide decision-support value,
but model uncertainty and weak longer-horizon skill must remain explicit.