Landscape Associations with Flood Response to Extreme Rainfall: A Remote-Sensing, Multi-Event Analysis Across Sumatra, Indonesia
Abstract
Tropical flooding emerges from interactions among rainfall forcing, drainage structure, land-surface condition, and evolving land cover, yet these components are often analyzed in separate remote-sensing workflows. This study develops a multi-event, sub-basin-scale association framework linking extreme-rainfall coverage, vegetation cover, built-up cover, and Sentinel-1-derived flood response across HydroBASINS Level 9 units in Sumatra, Indonesia. Nine annual area-mean daily rainfall maxima from 2017–2025 were selected using CHIRPS. Vegetation fraction was defined as the proportion of valid Sentinel-2 pixels with NDVI ≥ 0.40, built-up fraction as the proportion with NDBI ≥ 0.00, flood fraction as the proportion of Sentinel-1 pixels with Random-Forest flood probability ≥ 0.50, and extreme fraction as the proportion of CHIRPS pixels whose 7-day accumulated rainfall exceeded the local monthly 95th percentile. The full panel contained 14,409 polygon-event observations, of which 11,818 were complete across the four core parameters. Fractional logit models with event fixed effects showed a negative association between vegetation fraction and flood fraction (β = −0.073, p = 0.004) and a positive association for built-up fraction (β = 0.082, p < 0.001), whereas extreme-rainfall coverage showed no direct association after event-level heterogeneity was absorbed. The corresponding landscape effects were modest in magnitude and should be interpreted as associations rather than causal effects. Cross-scale analysis at HydroBASINS Level 8 preserved the direction of the vegetation and built-up effects. The contribution therefore lies not in a new satellite sensor or classifier, but in combining multi-sensor fractions, hydrologically aligned sub-basin units, and repeated-event statistical inference. Because the Sentinel-1 flood product is pseudo-label based and was not independently validated for all nine events, flood_fraction is treated as a SAR-derived flood indicator rather than a fully validated inundation product, and no operational warning threshold is inferred.