Abstract
<jats:p>Recurrent inundation in metropolitan Accra is, on the preponderance of the evidence, an anthropogenic rather than a climatological phenomenon: its proximate determinants are the obstruction of drainage by solid waste and the encroachment of structures onto watercourses. These determinants are metre-scale and therefore invisible to the moderate-resolution satellite imagery on which prior work has relied; municipal authorities presently enumerate them by pedestrian survey conducted after floods. This work develops a six-stage, fully reproducible pipeline that couples deep-learning semantic segmentation of centimetre-resolution aerial orthoimagery with terrain hydrology and open-channel hydraulics, using only openly licensed data, across five informal settlements of the Odaw catchment. Its scope is deliberately bounded: the system maps the built fabric and the drainage network, rates whether that network's measured geometry suffices, and tests the rating against independently documented flood history. Five results are reported. First, segmentation accuracy in this setting is bounded by label noise rather than model capacity; a cross-source consensus verification scheme - trusting only pixels on which two independent footprint sources agree, and excluding disputed pixels from loss and metric alike - establishes that the incumbent model was never deficient but mis-benchmarked, scoring 0.77 on verified pixels where it scored 0.58 against raw labels, and retraining under the corrected supervision reaches 0.80. Second, height above nearest drainage separates flood-reported localities from elevated controls at p = 0.0074 in a basin for which no flood extent has ever been published. Third, Manning conveyance analysis of 652 field-surveyed drain cross-sections finds the network is not undersized: 2.3% fails a conservative reference storm when clean, and siltation to three-quarters depth would multiply the failing length elevenfold to 25.8%. Fourth, that scenario is then measured rather than assumed - hand interpretation of 99 readable segments returns a mean surface obstruction of 0.29, so the elevenfold figure describes the network's sensitivity to siltation and not its present condition, under which one segment in ninety-nine fails. Fifth, cross-site degradation of the segmentation model is shown by leave-one-site-out testing to be a limitation of corpus coverage rather than a ceiling, recovering a mean of +0.287 IoU on unseen settlements. Five negative results are documented alongside, among them that reanalysis rainfall cannot supply a design storm for Accra, and that obstruction is not separable from overhead imagery at this resolution - a building-segmentation encoder discriminates blocked from clear channel at chance. Automated obstruction sensing is therefore scoped out of the present system and set out as future work, with the reasons quantified.</jats:p>