Integrating historical archives and geospatial data to revise flood estimation equations for Philippine rivers

Abstract. Flood magnitude and frequency estimation are essential for the design of structural and nature-based flood risk management interventions and water resources planning. However, the global geography of hydrological observations is uneven, with many regions, especially in the Global South, having spatially and temporally sparse data that limit the choice of statistical methods for flood estimation. To address this data scarcity, we pool all available annual maximum flood data for the Philippines to estimate flood magnitudes at the national scale. Available river discharge data were collected from publications covering 842 sites, with data spanning from 1908 to 2018. Of these, 466 sites met criteria for reliable estimation of the annual maximum flood. Using the index flood approach, a range of controls was assessed at both national and regional scales using modern land cover and rainfall data sets, as well as geospatial catchment characteristics. Predictive equations for 2 to 100 year recurrence interval floods using only catchment area as a predictor have R2≤0.59. Adding a rainfall variable, the median annual maximum 1 d rainfall, increases R2 to between 0.56 for Q100 and 0.66 for Q2. Very few other topographic or land use variables were significant when added to multiple regression equations. Relatively low R2 values in flood predictions are typical of studies from tropical regions. Although the Philippines exhibits regional climate variability, residuals from national predictive equations show limited spatial structure, and region-specific equations do not significantly outperform the national equations. The predictive equations are suitable for use as design equations in ungauged catchments for the Philippines, but statistical uncertainties must be reported. Our approach demonstrates how combining individually short historical records, after careful screening and exclusion of unreliable data, can generate large data sets that can produce consistent results. Extension of continuous flood records by continuous and rated monitoring is required to reduce uncertainties. However, the national-scale consistency in our results suggests that extrapolation from a small number of carefully selected catchments could provide nationally reliable predictive equations with reduced uncertainties.