Great Ruaha River — A1/A2 Low-Flow Hindcast
Reconstruction using information available through 06 August 2026
RETROSPECTIVE HINDCAST — NOT AN ORIGINAL 06 AUG ISSUED FORECAST
Later observations are shown only in the separate evaluation section and were not allowed into the hindcast calculations.
FC/PRE recent-decline projection
PRE_ORIG_V3_RECENT_LINEAR used the six most recent valid Sentinel FC/NDWI water-area observations available by the cutoff and fitted a straight line against actual acquisition date. The endpoint was the mathematical extrapolation of classified A2 residual-channel water area to zero.
At 05 August the classified A2 area was 0.000942 km² (942 m²). The fitted recent decline was approximately 0.000043 km²/day (43 m²/day), giving about 22.1 days to the zero-area endpoint.
JOINT-population 2024 shape analog
JOINT_POP_2024_NORMALIZED_SHAPE_ANALOG compared the normalized 2026 joint-population decline with the available 2024 decline shape. It produced a zero-core estimate near 11 September.
Status: provisional experiment only. At the time of this reconstruction the analog sample was effectively N=1, so this result must not be presented as an established forecast method.
JOINT-population recent linear extrapolation
JOINT_POP_RECENT_LINEAR linearly extrapolated the recent core-count trajectory. The result was far later than the subsequent observed zero-core state and demonstrates that a simple linear fit to the early joint-population decline was unsuitable for the accelerating 2026 recession.
Status: negative/failed hindcast experiment — retained as scientific evidence, not as an operational forecast method.
| Acquisition | A2 classified water area | Relative to calibrated reference |
|---|---|---|
| 06 Jul 2026 | 0.002427 km² | 95.52% |
| 16 Jul 2026 | 0.001953 km² | 76.89% |
| 23 Jul 2026 | 0.001945 km² | 76.55% |
| 26 Jul 2026 | 0.001732 km² | 68.16% |
| 02 Aug 2026 | 0.001473 km² | 58.00% |
| 05 Aug 2026 | 0.000942 km² | 37.10% |
The FC/PRE signal belongs to the older Sentinel false-colour/NDWI water-area family. The active SR water classification used B8 < 2300, B4 < 2000, B3 > 350 and NDWI > −0.15, followed by physical pixel-area summation. The later PRE lineage applied that classifier to the smaller A2 residual-channel/ZF geometry. The 27 August endpoint itself did not use the older 0.005/0.006 km² Pipeline collapse threshold.
“Core” and “transition” are spectral population classes. They should not be read automatically as “deep” and “shallow” water. The 89.8% value is relative to the 2026 early-dry-season reference core population of 118 cells, not to historical zero flow.
Observed A2 zero-core state: 30 August 2026
On 30 August the joint-population observation reached core = 0 with transition = 12, classified ZF_LIKE. This is an observation/state endpoint; residual transition signatures mean it is not equivalent to saying that every wet signature had disappeared.
| Hindcast experiment | Predicted endpoint | Compared with 30 Aug zero-core | Assessment |
|---|---|---|---|
| FC/PRE recent-six linear | 27 Aug | −3 days | Promising single-year result; validation still required |
| JOINT 2024 shape analog | 11 Sep | about +12 days | Provisional N=1 analog |
| JOINT recent linear | 19 Nov | about +81 days | Poor / rejected as simple linear approach |
Earlier warning remains the objective
The 27 August FC/PRE result is encouraging because it reconstructed the terminal A2 decline closely, but it became fully specified only after the six recent observations through 05 August. That makes it a short-range A2 projection, not the final goal of the forecasting programme.
The broader research programme is testing whether rainfall, A1 timing, historical FC decline behaviour, Landsat evidence and joint-population trajectories can provide reliable warnings progressively earlier in the season. Rainfall/A1 work has a complex multi-experiment history and is being reconstructed separately; it is therefore not used here to manufacture or strengthen the A2 hindcast result.
A1/A2 only: this bulletin does not derive or present Park HQ/Msembe ZFO as an A1/A2 forecast output. Alert 3 and Park/Msembe are separate stages in the wider river chronology.
No future leakage: the three hindcast calculations above were restricted to information available by 06 August 2026; later observations appear only in the retrospective evaluation box.
Methods are not independent votes: Pipeline and PRE belong to the same broad Sentinel FC/NDWI water-area lineage. JOINT-pop uses a different fixed-point/RF population representation but still uses Sentinel imagery. Landsat, rainfall and visual/field evidence are separate evidence families and require dependency-aware comparison.
Forecast status: none of the experimental A2 methods on this page should be interpreted as already validated from a single 2026 case. Historical pseudo-operational backtesting and future prospective tests remain necessary.
| Family | Role in this bulletin | Status |
|---|---|---|
| FC/NDWI Pipeline → PRE | Classified A2 water-area trajectory; recent-six extrapolation to zero | Short-range candidate; historical lineage/backtest being reconstructed |
| JOINT-pop observation | Fixed 255-point RF spectral population; core/transition/state | Observation/state method |
| JOINT recent linear | Experimental core-count extrapolation | 2026 hindcast failed |
| JOINT 2024 analog | Experimental normalized decline-shape transfer | Provisional N=1 |
| Rainfall → A1 / ZFO warning | Earlier-warning research context | Complex lineage under reconstruction; not used to select A2 endpoint here |