Speaker
Description
Automated quality control (QC) of rainfall measurements from sparse networks, such as the Trans-African Hydro-Meteorological Observatory (TAHMO), is challenging due to environmental faults like persistent clogging and mechanical spikes. Traditional QC algorithms rely heavily on spatial cross-validation with neighboring gauges; however, this approach fails in these regions where spatial correlation is weak. While recent studies have demonstrated the viability of Commercial Microwave Links (CMLs) as opportunistic sensors for rainfall estimation, their integration into automated gauge auditing frameworks remains underexplored.
To bridge this gap, this study introduces an observer-based fault detection framework utilizing a decoupled, two-stage architecture to validate ground stations. In the first stage, a constrained Luenberger Observer acts as a dynamic state estimator to generate a highly stable virtual sensor. Operating within a 3-month rainy season window dictated by current data availability, the system fuses pre-processed, localized gridded CML rainfall maps with hourly multivariate ERA5 weather variables (Total Column Water Vapour, surface solar radiation downwards, 2m temperature, 2m dew temperature, Total Cloud Cover, and Boundary Layer Height) downscaled to 5-minute resolution via shape-preserving PCHIP interpolation in Ouagadougou, Burkina Faso, using bilinear interpolation to resolve spatial mismatches. To mitigate the CML Wet Antenna Effect, radome drying is modeled as a first-order lag system (τ=14.52 minutes) to build an inverse lead compensator. Using system identification via multiple linear regression, we capture highly accurate atmospheric dynamics (R²=0.793), deriving convective persistence (A≈0.621) and environmental forcings, while the optimal Kalman gain (L≈0.228) is extracted via the steady-state Riccati equation.
Because linear observers struggle with non-stationary environmental variance, the second stage introduces a diagnostic decision engine. Operating on the observer’s innovation residuals, this state-machine uses a 5x baseline noise gate and an asymmetric decay memory to isolate faults, while autonomously self-healing during natural correlation gaps.
Validation via synthetic fault injection over a 3-month rainy season demonstrated fail-safe resilience. The optimally tuned model achieved 91% overall accuracy, maintaining perfect detection for transient spikes (Precision and Recall of 1.00) while successfully isolating 90% of persistent clogs. By drastically reducing false positives, this framework proves the viability of utilizing opportunistic sensing as a highly reliable, automated auditing tool, paving the way for scalable QC across Africa.