Mapping Flood-Exposed Road Segments from Freely Available Satellite Imagery During The 2022 Flood Season

model, which differ in the per-segment elevation error the reference carries, with the study area held fixed. The block component is the spread of the same statistic across six spatial IIARD International Journal of Geography & Environmental Management blocks within one study area with the reference replicate held fixed. Neither is a between-basin variance: with two study areas there is no such estimate to be had, and the paper does not claim one. 5 Results 5.1 One area works modestly, the other not at all Table 3 and Figure 4 report overall skill. On the Kwara floodplain, segment-level detection reaches an F1 of 0.356, with precision 0.378, recall 0.337, a critical success index of 0.217 and an area under the receiver operating curve of0.746. That is modest. It is enough to screen a network for attention and not enough to dispatch on. Table 3. Detection skill at the operating point, with the two variance components reported separately. Area Precision Recall F1 CSI MCC κ Positives Kwara 0.378 0.337 0.356 0.217 0.336 0.335 5,709 replicate SD (n=20) 0.0135 0.0077 0.0097 0.0072 0.0101 0.0099 area SD (6 blocks) 0.1622 0.2066 0.1834 0.1193 0.1705 0.1726 Osun – 0.000 0.000 0.000 – 0.000 566 replicate SD (n=20) – 0.0000 0.0000 0.0000 – 0.0000 area SD (6 blocks) – 0.0000 0.0000 0.0000 – 0.0000 Note: The operating point is a split-based backscatter threshold selected on each date’s own imagery, after Martinis et al. (2009): the scene is tiled, each tile is tested for the bimodality that open water produces, and Otsu’s split is taken on the pooled histogram of the most bimodal tiles. A date with fewer than three bimodal tiles falls back to a fixed−20dB, which is conservative rather than empty: it fired on 2 of 15 Kwara dates and on all 15 Osun dates and still mapped a trace of water on each. The rest of the operating point is a 30 m buffer from the centerline and a decision cut of 25% of buffered cells classified as water. CSI is the critical success index and MCC is Matthews’ correlation coefficient. The replicate standard deviation is the spread across 20 independent realizations of the reference model with the area held fixed; the area standard deviation is the spread across spatial blocks with the reference replicate held fixed. The two are different kinds of uncertainty and must not be added. Precision, recall, F1, CSI, MCC andκare means over the 20 replicates; Positives is the count on the baseline replicate alone, so it does not average to the mean reference rate. Skill here is agreement with the hydrological reference of Section 3; Osun’s stage law within that reference is transferred from Kwara. IIARD International Journal of Geography & Environmental Management Figure 4. Segment-level agreement with the hydrological reference, and the diagnostic that governs it. (a) and (b) are the confusion tables on the baseline replicate, Kwara and Osun, in counts and as a percentage of all segment-days; Osun’s detector column is empty. (c) is the precision-recall curve for Kwara by road class, obtained by sweeping the decision cut on the water fraction, with average precision per class and the 3.36 percent reference base rate marked. (d) is the count of bimodal tiles per scene for both areas on a log axis, against the minimum of three that a data-driven threshold requires: Kwara crosses it on 13 of 15 dates and Osun on none. On the Osun basin the method returns no detection across all fifteen acquisitions. This is not allowed score. No tile of any scene passed the bimodality test, so no threshold was ever selected from the data. Panel (d) of Figure 4 shows the diagnostic directly: Kwara’s bimodal tile count rises from 4 in June to 216 at the September peak, while Osun’s is zero on every date. Two things about that null need stating precisely. First, the conservative−20 dB fallback didmap a trace of water in Osun, between 0.00 and 0.02 km2on seven of the fifteen dates, which is why the Osun water-extent agreement reported in Section 6 is a small number rather than an undefined one. What is exactly zero is the number of segments that reached the decision cut. Second, and more important, the null is a statement about the detector and not about the basin. The hydrological reference calls 566 Osun segment-days impassable over the same season, peaking at 177 segments on 18 October, on a catchment that received 1,045 mm of rain between June and November. The correct reading is that the method made no claim in Osun, not that Osun did not flood. Three mechanisms explain the blindness and they compound. At 76 percent canopy, most of the basin’s surface is not visible to the sensor in the first place(Tsyganskaya et al. 2018). Ianthe built-up parts, double-bounce returns raise the bright mode and fill the IIARD International Journal of Geography & Environmental Management histogram valley that thresholding depends on(Mason et al. 2014). And a 12-day revisit, with a 24-day gaping this series, cannot see flash flooding on a basin of this size, which recedes between acquisitions. This is a method with a domain, not a method with an accuracy. Reporting pooled figure across both areas would produce a number describing neither. 5.2 Where skill goes within the area that works Table 4 and Figure 5 report the stratification, and the gradients are steep. By road class, F1 falls from 0.368 on tracks to 0.344 on local roads, 0.233 on tertiary and 0.097on primary. Kwara’s trunk roads carry no reference positive at all on the baseline replicate and only 9across all twenty, and its secondary roads carry 154 positives across twenty replicates against zero detections, so their recall is 0.000 and their precision is undefined. Both are reported that way rather than as a number computed on a handful of cases. The ordering runs opposite to operational need, and the temptation is to conclude that the method is least reliable on the roads that matter most. That conclusion is not available from this design. The reference rate falls monotonically with road class, from 5.94 percent of segment-days on tracks to 0.04 percent on trunk roads, because embankment heights rise with class in the reference model and so do the depths at which a segment is called impassable, both set out in Section 3. F1 is sensitive to the base rate, so a monotone fall in F1 across stratified whose base rate falls monotonically tracks the base rate at least as much as the detector. What the road-class rows do establish is narrower: the detector found almost nothing on the higher classes, whether because those roads genuinely flood less, or because a formed embankment keeps the carriageway above the water the radar sees beside it, or because the reference model’s ladders say so. This design cannot separate the three. Table 4. Detection skill disaggregated by road class, tree cover, built-up cover, segment length, height above nearest drainage and the presence of mapped water within 30 m of the centerline. Stratum Segment s Ref. rate (%) Precision Recall F1 F1 replicate SD Kwara, Niger floodplain Road class trunk 67 0.04 – 0.000 0.00 0 – primary 496 0.40 0.092 0.102 0.09 7 0.0682 secondary 52 0.99 – 0.000 0.00 0 – tertiary 767 0.74 0.457 0.156 0.23 3 0.0633 local 5,694 2.12 0.338 0.350 0.34 4 0.0151 track 4,204 5.94 0.404 0.338 0.36 8 0.0098 Tree cover 0-10% 8,105 3.66 0.392 0.339 0.36 4 0.0109 IIARD International Journal of Geography & Environmental Management Stratum Segment s Ref. rate (%) Precision Recall F1 F1 replicate SD 10-30% 1,705 2.46 0.331 0.353 0.34 1 0.0162 30-60% 923 2.46 0.324 0.345 0.33 4 0.0311 >60% 547 3.12 0.364 0.254 0.29 9 0.0247 Built-up cover <5% 8,835 4.03 0.383 0.348 0.36 5 0.0094 5-20% 1,044 1.42 0.292 0.201 0.23 8 0.0352 >20% 1,401 0.51 0.152 0.089 0.11 2 0.0425 Segment length <150 m 750 3.43 0.380 0.331 0.35 4 0.0290 150-300 m 904 4.05 0.399 0.376 0.38 8 0.0189 300-450 m 8,941 3.15 0.377 0.330 0.35 2 0.0110 >450 m 685 5.06 0.363 0.363 0.36 3 0.0197 HAND <2 m 3,913 8.88 0.412 0.352 0.38 0 0.0099 2-5 m 2,513 1.23 0.131 0.174 0.15 0 0.0221 5-10 m 1,249 0.00 0.000 – – – >10 m 3,605 0.00 – – – – Mapped water within 30 m within 30 m 473 0.12 0.008 0.701 0.01 6 0.0095 beyond 30 m 10,807 3.50 0.438 0.337 0.38 1 0.0105 Osun, Osun River basin Road class trunk 401 0.00 – 0.000 0.00 0 – primary 868 0.01 – 0.000 0.00 0 – secondary 1,182 0.07 – 0.000 0.00 0 0.0000 IIARD International Journal of Geography & Environmental Management Stratum Segment s Ref. rate (%) Precision Recall F1 F1 replicate SD tertiary 1,985 0.04 – 0.000 0.00 0 0.0000 local 35,959 0.09 – 0.000 0.00 0 0.0000 track 1,483 0.28 – 0.000 0.00 0 0.0000 Tree cover 0-10% 17,932 0.07 – 0.000 0.00 0 0.0000 10-30% 7,659 0.12 – 0.000 0.00 0 0.0000 30-60% 6,182 0.16 – 0.000 0.00 0 0.0000 >60% 10,105 0.08 – 0.000 0.00 0 0.0000 Built-up cover <5% 11,317 0.11 – 0.000 0.00 0 0.0000 5-20% 3,746 0.12 – 0.000 0.00 0 0.0000 >20% 26,815 0.08 – 0.000 0.00 0 0.0000 Segment length <150 m 13,918 0.13 – 0.000 0.00 0 0.0000 150-300 m 9,406 0.09 – 0.000 0.00 0 0.0000 300-450 m 16,810 0.07 – 0.000 0.00 0 0.0000 >450 m 1,744 0.09 – 0.000 0.00 0 0.0000 HAND <2 m 1,588 2.21 – 0.000 0.00 0 0.0000 2-5 m 3,796 0.11 – 0.000 0.00 0 0.0000 5-10 m 7,587 0.00 – – – – >10 m 28,907 0.00 – – – – Mapped water within 30 m within 30 m 221 0.01 – 0.000 0.00 0 – beyond 30 m 41,657 0.09 – 0.000 0.00 0 0.0000 IIARD International Journal of Geography & Environmental Management Note: Reference rate is the share of segment-days the hydrological reference calls impassable in that stratum; precision and recall are not comparable across strata without it. F1 is base-rate sensitive, and the reference rate falls with road class by construction, because the reference model’s embankment heights and pass ability depths rise with class, so the road-class gradient has to be read against it. Metrics come from confusion cells pooled over the 20 reference replicates, and every quantity describing the positive class is blanked together where recall is undefined. Cover fractions are ESA World Cover 2021 within 30 m of the centerline; class 50 includes road surfaces, so the built-up stratified is partly the road itself. The water stratum is every segment whose 30 m context holds a mapped waterway or that carries an OSM bridge tag; only 8 of the 473 in Kwara and 13 of the 221 in Osun carry the tag. Osun’s stage law within the reference is transferred from Kwara. Figure 5. Skill by stratum, both study areas. (a) is the only receiver operating panel in the set: Kwara’s curves by canopy fraction, obtained by sweeping the decision cut, with the area under each curve. The remaining panels plot precision and recall at the operating point with their replicate standard deviations, against (b) road class, (d) segment length, (e) height above nearest drainage and (f) built-up cover; (c) plots F1 against tree cover. Osun sits on zero throughout, which is the flat lower series in every panel. There is no panel for the mapped-water stratum, which Table 4 carries instead. The canopy and built-up gradients are the detector’s own. By canopy, F1 falls from 0.364 to 0.299 and recall from 0.339 to 0.254 between the low and high strata. By built-up fraction, F1 falls from0.365 to 0.112, a drop of 25.4 percentage points, and recall from 0.348 to 0.089, a drop of 26.0percentage points. The comparable published figure, a roughly 15-point urban penalty inGiustarini et al. (2013), is an accuracy and recall penalty, so the recall drop is the like-for-like comparison and it is the larger of the two. Our built-up stratum is more mixed than a dedicated urban study’s would be, and the ESA World Cover caveat noted with Table 2 applies to it. IIARD International Journal of Geography & Environmental Management Above 2 m of height above nearest drainage, skill collapses, and the terrain refinement is the obvious reason. The reading has a limit, though. Between 2 and 5 m the detector still fires on 1.6percent of segment-days and precision is 0.131, so what happens at elevation is not that detections are suppressed but that the ones that survive are mostly wrong. Above 5 m the reference calls nothing impassable and the comparison stops being informative in either direction. 5.3 Roads next to water The sharpest failure in this study is not statistical but geometric. The stratum is every segment whose 30 m context holds a mapped watercourse together with every segment carrying an OpenStreetMap bridge tag, 473 of them in Kwara. Across it precision is 0.008against 0.438 elsewhere, while recall over the same stratum is 0.701 against 0.337. Pooled over the twenty reference replicates the stratum holds 117 true positives against 14,043 false ones and 50 false negatives. The detector is not missing these segments; it is calling roughly one in ten of their segment-days cut, against a reference rate of 0.12 percent. Almost every positive is false, and the reason is that the water the sensor sees is really there and is there on every date. The two rates come from opposite sides of the same geometry. The reference treats a segment with a mapped crossing as bridged and holds it passable until the water reaches an assumed 3.5 m deck freeboard, which is why its positive rate in this stratum is 0.12 percent while the detector calls9.98 percent of the same segment-days cut. A different freeboard would move the printed precision of 0.008 up or down without touching the mechanism, which is that water inside the 30 m context ofa road is present flood or no flood. The stratum is not what the phrase 'bridge problem' would suggest. Of the 473 Kwara segments in it, 8 carry an OpenStreetMap bridge tag, and of the 221 in Osun, 13 do. For roughly 98 percent of the stratum the segment is not a deck over a river at all; it is a road whose 30 m context happens to contain a mapped watercourse, which in practice usually means a road running beside a stream. The failure mechanism is proximity to mapped water, and the small tagged-bridge subset isa special case of it rather than the whole of it. This is not fixable by better thresholding, because the water is really there. It requires masking segments with mapped water in their context out of the closure inference, or treating them under a separate rule, and any study that consumes a flood extent product to infer road closures without doing so will report a large number of confident false closures. Whether those segments are also the ones whose closure would matter most to the network is a question this paper does not answer: no connectivity, betweenness or cut analysis was performed, and Section 4.5 explains why the published closure set does not support one. 6 Robustness Checks 6.1 Sweeps Table 5 and Figure 6 report the sweeps. The backscatter threshold matters most. Across−22to−10 dB, F1 ranges from 0.130 to 0.365,so the headline figure sits at the top of a wide range and a poorly chosen threshold loses most of the available skill. The road buffer barely matters. Across 10 to 120 m, F1 ranges from 0.343 to 0.376. That is worth knowing because buffer width is the parameter practitioners most often argue about, and here it is close to irrelevant. The decision cut also matters less than its prominence suggests. The parameter is the water fraction cut, the share of buffered cells that must classify as water before a segment is called cut, and it is not a calibrated probability. Swept from 0.05 to 0.80 it moves F1 between 0.172 IIARD International Journal of Geography & Environmental Management and0.370, with the maximum of 0.370 at a cut of 0.10 against 0.356 at this paper’s operating value of0.25. Panel (c) of Figure 6 shows why the curve is flat over its useful range: precision and recall cross close to the operating point, so tightening the cut buys precision at almost exactly the recall it costs. Table 5. Sensitivity of detection skill to the three parameters the analyst chooses: the backscatter threshold, the buffer between water and centerline, and the fraction of buffered cells required before a segment is called cut. Parameter Swept over Operating value F1 at operating F1 range Best value Best F1 Kwara, Niger floodplain VV threshold −22to−10 dB split- based,−20.0to −12.4dB 0.356 0.130– 0.365 −11.0dB 0.365 Buffer distance 10 to 120 m 30 m 0.356 0.343– 0.376 120 m 0.376 Decision cut 0.05 to 0.80 0.25 0.356 0.172– 0.370 0.10 0.370 Osun, Osun River basin VV threshold −22to−10 dB split- based,−20.0to −20.0dB 0.000 0.000– 0.000 −22.0dB 0.000 Buffer distance 10 to 120 m 30 m 0.000 0.000– 0.000 10 m 0.000 Decision cut 0.05 to 0.80 0.25 0.000 0.000– 0.000 0.05 0.000 Note:F1 at operating is the mean over reference replicates at this paper’s operating point. F1 range is the full span of the sweep, so a narrow range means the result does not depend on the choice. Each parameter is swept with the other two held at their operating values. Figure 6. Sweeps over the three parameters the analyst chooses, F1 against each, with the operating value marked by a dotted vertical line and the replicate spread shaded. (a) the VV IIARD International Journal of Geography & Environmental Management backscatter threshold, with the shaded band giving the range the split-based selection actually chose across the season. (b) the buffer from the centerline. (c) the water fraction decision cut, with precision and recall added as light dotted lines to show where they cross. Kwara is the upper series in each panel and Osun the flat lower one. 6.2 Replicate variance against block variance The two are a factor of 18.8 apart. Holding the area fixed and varying the reference across its twenty replicates gives a standard deviation on F1 of 0.0097. Holding the reference replicate fixed and varying across six spatial blocks within the Kwara study area gives 0.1834. The comparison is between two sources of uncertainty that a reader might otherwise assume are of the same size, and they are not: which part of a basin is examined moves the answer far more than which replicate of the reference it is scored against. It follows that a skill figure should be reported with the surface conditions it was measured under, and that published accuracy figures for flood detection are better read as properties of a validation site than of a method. What does not follow is a number for how much skill varies between basins. The block component is within-area spread, not a between-area one, and with two study areas there is no between-area variance to estimate. Our two areas do differ by more than the difference between modest performance and none, which points the same way, but it is one contrast and not a variance component. 6.3 Benchmarks Eighteen quantities are recorded against published values in Table 6. Seven of them have a published numeric range to sit inside or outside, and four of those fall inside it and three below. Four more are checked against a published claim that states a direction and attaches no level to it, so the most that can be said is that the direction holds: the recall losses across built-up land and across closed canopy in Kwara, and the mean clear-sky fraction in each area. Placing any of the four inside an interval would be reporting an interval the source does not give. The remaining seven are marked not comparable, and they divide into three kinds. Four are road-pass ability scores, for which no published envelope exists at all. One is Osun’s automatic water threshold, which was never selected. Two are Osun degradation statistics, differences between detection rates in an area where the detector made no detection, so they are zero by arithmetic rather than by measurement and grading them in either direction would read a null as a result. The Kwara water-extent critical success index of 0.404 sits inside the 0.30 to 0.90 envelope Bernhofen et al. (2018)reports for flood extent methods generally, but below the0.45 to 0.70 that source gives for Nigerian floodplains specifically, which is the tighter and more relevant of the two. The comparison also needs a caveat the number does not carry on its face. The stage law underlying the reference is calibrated to the radar’s own total flooded area, so any agreement on area is circular. What is not circular is where that area falls, which the elevation model decides alone, and the 0.404 is a cell-level agreement rather than anareal one. That is the part of the check that carries information. Three quantities fall below their envelope. Osun’s water-extent index of 0.039 is one of them, and it is a score rather than a null: the fallback threshold classified a trace of water and the comparison is real, it is simply very poor. The other two are the modeled seasonal stage ranges,1.33 m in Kwara and 1.10 m in Osun, both smaller than the 3.91 m floodplain vertical error Hawker et al. (2019)reports for elevation models of this generation. That failure was acted on rather than noted. The vertical uncertainty was carried into the reference as a per-segment error varied across replicates, and swept from zero to the full published value. At IIARD International Journal of Geography & Environmental Management the operating value of 1.2 m, F1 is 0.358; at the full 3.91 m it is 0.296, a fall of 0.062, or 17percent of the operating value, and with no elevation error at all it would be 0.405. The loss falls almost entirely on recall while precision barely moves, which is what a vertical error should do: it scatters segments across the pass ability threshold rather than relocating the flood. The sweep also moves the reference positive rate from 3.09 percent to 4.82 percent of segment-days, so the reference model’s base rate is itself a function of an assumption about elevation error, and the absolute skill figures should be read with that in mind. The road pass ability rows carry no verdict at all, and that absence is the paper’s point restated. No published envelope exists for road closure detection skill, so there is nothing to compare against. This paper is an attempt to start one. Table 6. Benchmarking the quantities this study produces against published values for comparable systems. Quantity This study Published range Source Verdict Water-extent CSI at the seasonal peak, Kwara 0.404 0.30–0.90 overall; 0.45–0.70 for Nigerian floodplains Bernhofen et al. 2018 inside Automatic VV water threshold, Kwara −13.1 dB (−14.8 to−12.4) −22to−12dB for C- band VV open water Twele et al. 2016; Martinis et al. 2009 inside Dry-season median VV over land, Kwara −11.5 dB −12to−5dB for vegetated and cropped land Twele et al. 2016 inside Water-extent CSI at the seasonal peak, Osun† 0.039 0.30–0.90 overall; 0.45–0.70 for Nigerian floodplains Bernhofen et al. 2018 below Automatic VV water threshold, Osun never selected −22to−12dB for C- band VV open water Twele et al. 2016 not comparabl e Dry-season median VV over land, Osun −7.8 dB −12to−5dB for vegetated and cropped land Twele et al. 2016 inside Recall loss, open to built-up, Kwara 26.0 pp about 15 pp (rural 90% to urban 75%) Giustarini et al. 2013 direction only Recall loss, open to closed canopy, Kwara 8.5 pp a detection floor, not a small percentage penalty Chapman et al. 2015; Tsyganskaya et al. 2018 direction only Recall loss, open to built-up, Osun† 0.0 pp about 15 pp (rural 90% to urban 75%) Giustarini et al. 2013 not comparabl e Recall loss, open to closed canopy, Osun† 0.0 pp a detection floor, not a small percentage penalty Chapman et al. 2015; Tsyganskaya et al. 2018 not comparabl e IIARD International Journal of Geography & Environmental Management Quantity This study Published range Source Verdict Seasonal stage range against DEM error, Kwara 1.33 m SRTM vertical RMSE 3.91 m on floodplains Hawker et al. 2019 below Seasonal stage range against DEM error, Osun† 1.10 m SRTM vertical RMSE 3.91 m on floodplains Hawker et al. 2019 below Sentinel-2 mean clear-sky fraction, Kwara 33.1% cloud is the binding constraint on optical flood mapping in the wet tropics Wilson and Jetz 2016; Landuyt et al. 2019 direction only Sentinel-2 mean clear-sky fraction, Osun 19.5% cloud is the binding constraint on optical flood mapping in the wet tropics Wilson and Jetz 2016; Landuyt et al. 2019 direction only Road-passability CSI, Kwara 0.217 no published envelope exists – not comparabl e Road-passability F1, Kwara 0.356 no published envelope exists – not comparabl e Road-passability CSI, Osun† 0.000 no published envelope exists – not comparabl e Road-passability F1, Osun† 0.000 no published envelope exists – not comparabl e Note:Every published range in this table was read from the source named beside it rather than from a secondary summary, and where a quantity falls outside its range the text states what was done about it. A verdict of inside or below is available only where the source states a numeric range; where it states a direction and attaches no level, the check is directional and the verdict is direction only. The published literature validates flood maps against water extent; this study validates against road passability, so the comparison is indicative rather than like for like, and that is the point the paper makes.†marks a row whose absolute level rests on a transferred calibration: Osun’s stage law is taken from Kwara because the Osun radar detects no standing water all season and cannot calibrate its own reference. Water-extent agreement is cell-level rather than areal, for the reason given in the text. Degradation rows for an area in which the detector made no detection at all are marked not comparable, because a difference between two zero detection rates is zero by arithmetic and is not a measured penalty. Road-passability rows rest on the hydrological reference of Section 3. IIARD International Journal of Geography & Environmental Management Figure 7. The 2022 season in both areas. (a) and (b) give CHIRPS daily catchment rainfall on the left axis and the modeled flood stage that drives the reference on the right, for Kwara and Osun; the markers on the stage line are the Sentinel-1 acquisition dates. (c) and (d) give the number of segments called cut on the left axis, by the reference as a replicate mean with its range over 20 replicates shaded and by the Sentinel-1 detector, with the classified flood area on the right. Kwara’s detected and reference series track each other through the September peak. Osun’s reference rises twice while its detected series stays at zero all season, which is the null this paper reports. Figure 7 places both areas on the same time base and makes the shape of the season visible. In Kwara the detector and the reference rise and fall together through the September peak, which is the agreement the F1 of 0.356 summarizes. In Osun the reference rises twice, in early July and again through October, while the detected series never leaves zero. The gap between those towlines is the whole of the Osun result. 7 Discussion 7.1 What follows for anyone using flood products on roads The audience here is anyone building a road closure layer from satellite flood mapping, which increasingly includes national emergency agencies, and in Nigeria the National Emergency Management Agency and the state emergency management agencies of Kwara and Osun. First, mask segments with mapped water inside their buffer out of the closure inference, or handle them under a separate rule. This is a one-line fix and it removes the study’s largest source of confident false positives. Note that it is a broader class of road than the word bridge suggests:473 Kwara segments qualify and 8 of them carry a bridge tag, so the rule has to key on proximity to mapped water rather than on bridge tagging, which is sparse in both areas. Second, do not apply a single accuracy figure across a region. Within Kwara alone, the spread of F1across six spatial blocks is 18.8 times its spread across twenty reference replicates. A skill IIARD International Journal of Geography & Environmental Management figure should be reported with the surface conditions it was measured under. Third, publish probabilities rather than binary maps. Our Kwara closure set carries calibrated confidence with a Brier skill score of 0.153, and a downstream user can threshold it for their own tolerance for false alarms. The Osun set has a Brier skill score of 0.000, which is itself the honest signal that there is nothing in it to threshold. A binary map hides both facts. Fourth, treat canopy as a domain boundary rather than as a difficulty. In the Osun basin the correct output was no claim, and a method that had forced a threshold anyway would have produced a confident and entirely spurious map. The Osun null measures the detector’s blindness under canopy, not the incidence of flooding in the Osun basin, where 1,045 mm of rain fell between June and November and the hydrological reference calls 566 segment-days impassable. Fifth, do not treat a published segment table as a network without checking that it is one. Ours was not, and said it was; Section 4.5 gives the measured consequence and the repair. 7.2 Transferability Whether any of this carry to a third basin is settled by the paper’s own numbers rather than by an argument. The chain needs nothing that is not free and nothing that is not global, so it can be run wherever Sentinel-1 acquires, which for the operational land modes is most of the populated world. What it returns there is the open question. Inside the Kwara study area alone, F1 moves 18.8 times more across six spatial blocks than across twenty realizations of the reference, and across the two basins it moves between a usable screening tool and no claim at all. A figure that unstable within one basin of 6,305 km2is a description of a surface, and carrying the 0.356elsewhere would mean carrying the Niger floodplain with it. The design does travel, even where the numbers do not. Any study reporting flood detection skill for roads can and should report it by canopy, by built-up fraction and by proximity to mapped water, because those are the physical mechanisms and they are all available from open data. Road class is worth reporting too, but only alongside the reference rate in that class, because otherwise the gradients uninterpretable. That costs nothing and it converts a pooled number into a statement about where the method can be believed. 7.3 Limitations Osun’s stage law is transferred from Kwara rather than fitted, for the reason given in Section 3, so the absolute level of Osun’s reference follows from Kwara’s fit; every Osun number derived from it, including the water-extent index of 0.039 and the seasonal stage range of 1.10 m, is marked in Table 6. Three parameters of the reference model are set rather than fitted here: embankment height and pass ability depth, both assigned by road class, and the 3.5 m deck freeboard at watercourse crossings. The revisit is 12 days, with one 24-day gap between 19 August and 12 September 2022 that falls on the rising limb of the Niger flood, so this study speaks to sustained inundation only, and it uses ascending relative orbit 103 alone, which holds incidence angle and look direction constant and leaves their effects unmeasured. 8 Conclusion We asked what free satellite radar is worth for inferring road closures, using Sentinel-1 across two contrasting Nigerian basins in the 2022 season with 53,158 road segments. The answer is that the method has a domain rather than an accuracy. On the open Niger floodplain it reaches an F1 of 0.356 against the hydrological reference, with an area under the IIARD International Journal of Geography & Environmental Management curve of 0.746, enough for screening. On the canopy-heavy Osun basin it makes no claim at all across fifteen acquisitions, because no scene is bimodal and no threshold is ever selected. That null describes the detector and not the basin: the same season put 1,045 mm of rain on the Osun catchment and the hydrological reference calls 566 segment-days impassable there. Reporting an average of the two areas would describe neither. Within the working domain, skill is stratified by surface condition. F1 falls from 0.364 to 0.299across canopy and from 0.365 to 0.112 across built-up land, both declines being properties of the detector. Skill also falls across road class, from 0.368 on tracks to 0.097 on primary roads, but the reference rate falls with class as the embankment and pass ability thresholds rise, so this design cannot separate that ordering from a real effect. The most actionable result concerns roads next to water. Across the 473 Kwara segments whose 30 context holds a mapped watercourse or that carry a bridge tag, precision is 0.008 against 0.438elsewhere while recall rises to 0.701: the detector fires on roughly one segment-day in ten there, against a reference rate of 0.12 percent. Only 8 of those segments carry the tag, so the mechanisms proximity to mapped water rather than a deck over a river. Any pipeline inferring closures from flood extent will generate confident false closures on that class of road unless it masks them or handles them separately. Whether those are also the network’s most important links is a question we did not measure and do not claim. Within one basin, the spread of F1 across six spatial blocks is 18.8 times its spread across twenty reference replicates, which means published flood detection accuracies are better read as properties of their validation sites than of their methods. We therefore publish calibrated closure probabilities rather than a binary map, with the measured limits of the published file stated in its own metadata so that a study consuming it inherits the uncertainty rather than a claim of truth. What would improve on this is ground observation. A modest record of which segments were actually impassable, on a few dates, in both basins, would convert every skill figure here from internal consistency into accuracy, and it is the one input that no amount of imagery can substitute for. IIARD International Journal of Geography & Environmental Management References Adelekan, Ibidun O. 2010.“Vulnerability of Poor Urban Coastal Communities to Flooding inLagos,Nigeria.”Environment and Urbanization22 (2): 433– 50.https://doi.org/10.1177/0956247810380141. Agbola, Babatunde S., Owolabi Ajayi, Olalekan J. Taiwo, and Bolanle W. Wahab. 2012.“TheAugust2011 Flood inIbadan,Nigeria: Anthropogenic Causes and Consequences.”International Journal of Disaster Risk Science3 (4): 207– 17.https://doi.org/10.1007/s13753-012-0021-3. Alfieri, L., P. Burek, E. Dutra, et al. 2013.“GloFAS— Global Ensemble Streamflow Forecasting and Flood Early Warning.”Hydrology and Earth System Sciences17 (3): 1161– 75.https://doi.org/10.5194/hess-17-1161-2013. Arrighi, C., M. Pregnolato, R. J. Dawson, and F. Castelli. 2019.“Preparedness Against Mobility Disruption by Floods.”Science of The Total Environment654: 1010– 22.https://doi.org/10.1016/j.scitotenv.2018.11.191. Barrington-Leigh, Christopher, and Adam Millard-Ball. 2017.“The World’s User-Generated Road Map Is More Than 80% Complete.”PLOS ONE12 (8): e0180698.https://doi.org/10.1371/journal.pone.0180698. Bernhofen, Mark V., Charlie Whyman, Mark A. Trigg, et al. 2018.“A First Collective Validation of Global Fluvial Flood Models for Major Floods inNigeriaandMozambique.”Environmental Research Letters13 (10): 104007.https://doi.org/10.1088/1748-9326/aae014. Camboim, Silvana Philippi, João Vitor Meza Bravo, and Claudia Robbi Sluter. 2015.“An Investigation into the Completeness of, and the Updates to,OpenStreetMapData in a Heterogeneous Area inBrazil.”ISPRS International Journal of Geo-Information4 (3): 1366–88.https://doi.org/10.3390/ijgi4031366. Chapman, Bruce, Kyle McDonald, Masanobu Shimada, Ake Rosenqvist, Ronny Schroeder, and Laura Hess. 2015.“Mapping Regional Inundation with SpaceborneL-BandSAR.”Remote Sensing7 (5): 5440–70.https://doi.org/10.3390/rs70505440. Chini, Marco, Renaud Hostache, Laura Giustarini, and Patrick Matgen. 2017.“A Hierarchical Split-Based Approach for Parametric Thresholding ofSARImages: Flood Inundation as a Test Case.”IEEE Transactions on Geoscience and Remote Sensing55 (12): 6975– 88.https://doi.org/10.1109/TGRS.2017.2737664. Cohen, Jacob. 1960.“A Coefficient of Agreement for Nominal Scales.”Educational and Psychological Measurement20 (1): 37–46.https://doi.org/10.1177/001316446002000104. Congalton, Russell G. 1991.“A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data.”Remote Sensing of Environment37 (1): 35– 46.https://doi.org/10.1016/0034-4257(91)90048-B. DeVries, Ben, Chengquan Huang, John Armston, Wenli Huang, John W. Jones, and Megan W. Lang. 2020.“Rapid and Robust Monitoring of Flood Events UsingSentinel- 1andLandsatData on theGoogle Earth Engine.”Remote Sensing of Environment240: 111664.https://doi.org/10.1016/j.rse.2020.111664. Douglas, Ian, Kurshid Alam, Maryanne Maghenda, Yasmin McDonnell, Louise McLean, and Jack Campbell. 2008.“Unjust Waters: Climate Change, Flooding and the Urban Poor inAfrica.”Environment and Urbanization20 (1): 187– 205.https://doi.org/10.1177/0956247808089156. Drusch, M., U. Del Bello, S. Carlier, et al. 2012.“Sentinel-2:ESA’s Optical High-Resolution IIARD International Journal of Geography & Environmental Management Mission forGMESOperational Services.”Remote Sensing of Environment120: 25– 36.https://doi.org/10.1016/j.rse.2011.11.026. Echendu, Adaku Jane. 2020.“The Impact of Flooding onNigeria’s Sustainable Development Goals .”Ecosystem Health and Sustainability6 (1): 1791735.https://doi.org/10.1080/20964129.2020.1791735. Farr, Tom G., Paul A. Rosen, Edward Caro, et al. 2007.“TheShuttle Radar Topography Mission.”Reviews of Geophysics45 (2).https://doi.org/10.1029/2005RG000183. Feyisa, Gudina L., Henrik Meilby, Rasmus Fensholt, and Simon R. Proud. 2014.“Automated Water Extraction Index: A New Technique for Surface Water Mapping UsingLandsatImagery.”Remote Sensing of Environment140: 23– 35.https://doi.org/10.1016/j.rse.2013.08.029. Giustarini, Laura, Renaud Hostache, Patrick Matgen, Guy J.-P. Schumann, Paul D. Bates, and David C. Mason. 2013.“A Change Detection Approach to Flood Mapping in Urban Areas UsingTerraSAR-X.”IEEE Transactions on Geoscience and Remote Sensing51 (4): 2417– 30.https://doi.org/10.1109/TGRS.2012.2210901. Hawker, Laurence, Jeffrey Neal, and Paul Bates. 2019.“Accuracy Assessment of theTanDEM- X90Digital Elevation Modelfor Selected Floodplain Sites.”Remote Sensing of Environment232: 111319.https://doi.org/10.1016/j.rse.2019.111319. Herfort, Benjamin, Sven Lautenbach, João Porto de Albuquerque, Jennings Anderson, and Alexander Zipf. 2021.“The Evolution of Humanitarian Mapping Within theOpenStreetMapCommunity.”Scientific Reports11 (1): 3037.https://doi.org/10.1038/s41598-021-82404-z. Kermanshah, Amirhassan, and Sybil Derrible. 2017.“Robustness of Road Systems to Extreme Flooding: Using Elements ofGIS, Travel Demand, and Network Science.”Natural Hazards86 (1): 151–64.https://doi.org/10.1007/s11069-016-2678-1. Kittler, J., and J. Illingworth. 1986.“Minimum Error Thresholding.”Pattern Recognition19 (1): 41–47.https://doi.org/10.1016/0031-3203(86)90030-0. Koks, E. E., J. Rozenberg, C. Zorn, et al. 2019.“A Global Multi-Hazard Risk Analysis of Road and Railway Infrastructure Assets.”Nature Communications10 (1): 2677.https://doi.org/10.1038/s41467-019-10442-3. Landuyt, Lisa, Alexandra Van Wesemael, Guy J.-P. Schumann, Renaud Hostache, Niko E. C. Verhoest, and Frieke M. B. Van Coillie. 2019.“Flood Mapping Based on Synthetic Aperture Radar: An Assessment of Established Approaches.”IEEE Transactions on Geoscience and Remote Sensing57 (2): 722– 39.https://doi.org/10.1109/TGRS.2018.2860054. Lee, Jong-Sen. 1980.“Digital Image Enhancement and Noise Filtering by Use of Local Statistics.”IEEE Transactions on Pattern Analysis and Machine IntelligencePAMI-2 (2): 165–68.https://doi.org/10.1109/TPAMI.1980.4766994. Martinis, S., A. Twele, and S. Voigt. 2009.“Towards Operational Near Real-Time Flood Detection Using a Split-Based Automatic Thresholding Procedure on High ResolutionTerraSAR-XData.”Natural Hazards and Earth System Sciences9 (2): 303– 14.https://doi.org/10.5194/nhess-9-303-2009. Mason, D. C., L. Giustarini, J. Garcia-Pintado, and H. L. Cloke. 2014.“Detection of Flooded Urban Areas in High ResolutionSynthetic Aperture RadarImages Using Double Scattering.”International Journal of Applied Earth Observation and Geoinformation28: 150–59.https://doi.org/10.1016/j.jag.2013.12.002. IIARD International Journal of Geography & Environmental Management Mason, D. C., R. Speck, B. Devereux, G. J.-P. Schumann, J. C. Neal, and P. D. Bates. 2010.“Flood Detection in Urban Areas UsingTerraSAR-X.”IEEE Transactions on Geoscience and Remote Sensing48 (2): 882– 94.https://doi.org/10.1109/TGRS.2009.2029236. McFeeters, S. K. 1996.“The Use of theNormalized Difference Water Index in the Delineation of Open Water Features.”International Journal of Remote Sensing17 (7): 1425–32.https://doi.org/10.1080/01431169608948714. Nkeki, Felix Ndidi, Philip John Henah, and Vincent Nduka Ojeh. 2013.“Geospatial Techniques for the Assessment and Analysis of Flood Risk Along theNiger-BenueBasin inNigeria.”Journal of Geographic Information System5 (2): 123– 35.https://doi.org/10.4236/jgis.2013.52013. Nkwunonwo, U. C., M. Whitworth, and B. Baily. 2020.“A Review of the Current Status of Flood Modelling for Urban Flood Risk Management in the Developing Countries.”Scientific African7: e00269.https://doi.org/10.1016/j.sciaf.2020.e00269. Nobre, A. D., L. A. Cuartas, M. Hodnett, et al. 2011.“HeightAbovetheNearest Drainage— a Hydrologically Relevant New Terrain Model.”Journal of Hydrology404 (1-2): 13– 29.https://doi.org/10.1016/j.jhydrol.2011.03.051. Nobre, Antonio Donato, Luz Adriana Cuartas, Marcos Rodrigo Momo, Dirceu Luı́s Severo, Adilson Pinheiro, and Carlos Afonso Nobre. 2016.“HANDContour: A New Proxy Predictor of Inundation Extent.”Hydrological Processes30 (2): 320– 33.https://doi.org/10.1002/hyp.10581. Olofsson, Pontus, Giles M. Foody, Martin Herold, Stephen V. Stehman, Curtis E. Woodcock, and Michael A. Wulder. 2014.“Good Practices for Estimating Area and Assessing Accuracy of Land Change.”Remote Sensing of Environment148: 42– 57.https://doi.org/10.1016/j.rse.2014.02.015. Otsu, Nobuyuki. 1979.“A Threshold Selection Method from Gray-Level Histograms.”IEEE Transactions on Systems, Man, and Cybernetics9 (1): 62– 66.https://doi.org/10.1109/TSMC.1979.4310076. Pekel, Jean-François, Andrew Cottam, Noel Gorelick, and Alan S. Belward. 2016.“High- Resolution Mapping of Global Surface Water and Its Long-Term Changes.”Nature540 (7633): 418–22.https://doi.org/10.1038/nature20584. Pontius, Robert Gilmore, and Marco Millones. 2011.“Death toKappa: Birth of Quantity Disagreement and Allocation Disagreement for Accuracy Assessment.”International Journal of Remote Sensing32 (15): 4407– 29.https://doi.org/10.1080/01431161.2011.552923. Pregnolato, Maria, Alistair Ford, Craig Robson, Vassilis Glenis, Stuart Barr, and Richard Dawson. 2016.“Assessing Urban Strategies for Reducing the Impacts of Extreme Weather on Infrastructure Networks.”Royal Society Open Science3 (5): 160023.https://doi.org/10.1098/rsos.160023. Pregnolato, Maria, Alistair Ford, Sean M. Wilkinson, and Richard J. Dawson. 2017.“The Impact of Flooding on Road Transport: A Depth-Disruption Function.”Transportation Research Part D: Transport and Environment55: 67–81.https://doi.org/10.1016/j.trd.2017.06.020. Rennó, Camilo Daleles, Antonio Donato Nobre, Luz Adriana Cuartas, et al. 2008.“HAND, a New Terrain Descriptor UsingSRTM-DEM: Mapping Terra-Firme Rainforest Environments inAmazonia.”Remote Sensing of Environment112 (9): 3469– 81.https://doi.org/10.1016/j.rse.2008.03.018. IIARD International Journal of Geography & Environmental Management Shen, Xinyi, Emmanouil N. Anagnostou, George H. Allen, G. Robert Brakenridge, and Albert J. Kettner. 2019.“Near-Real-Time Non-Obstructed Flood Inundation Mapping Using Synthetic Aperture Radar.”Remote Sensing of Environment221: 302– 15.https://doi.org/10.1016/j.rse.2018.11.008. Tellman, B., J. A. Sullivan, C. Kuhn, et al. 2021.“Satellite Imaging Reveals Increased Proportion of Population Exposed to Floods.”Nature596 (7870): 80– 86.https://doi.org/10.1038/s41586-021-03695-w. Trigg, M. A., C. E. Birch, J. C. Neal, et al. 2016.“The Credibility Challenge for Global Fluvial Flood Risk Analysis.”Environmental Research Letters11 (9): 094014.https://doi.org/10.1088/1748-9326/11/9/094014. Tsyganskaya, Viktoriya, Sandro Martinis, Philip Marzahn, and Ralf Ludwig. 2018.“SAR-Based Detection of Flooded Vegetation — a Review of Characteristics and Approaches.”International Journal of Remote Sensing39 (8): 2255– 93.https://doi.org/10.1080/01431161.2017.1420938. Twele, André, Wenxi Cao, Simon Plank, and Sandro Martinis. 2016.“Sentinel-1-Based Flood Mapping: A Fully Automated Processing Chain.”International Journal of Remote Sensing37 (13): 2990–3004.https://doi.org/10.1080/01431161.2016.1192304. Wilson, Adam M., and Walter Jetz. 2016.“Remotely Sensed High-Resolution Global Cloud Dynamics for Predicting Ecosystem and Biodiversity Distributions.”PLOS Biology14 (3): e1002415.https://doi.org/10.1371/journal.pbio.1002415. Wing, Oliver E. J., Paul D. Bates, Christopher C. Sampson, Andrew M. Smith, Kris A. Johnson, and Tyler A. Erickson. 2017.“Validation of a 30 m Resolution Flood Hazard Model of the ConterminousUnited States.”Water Resources Research53 (9): 7968– 86.https://doi.org/10.1002/2017WR020917. Xu, Hanqiu. 2006.“Modification of Normalised Difference Water Index to Enhance Open Water Features in Remotely Sensed Imagery.”International Journal of Remote Sensing27 (14): 3025–33.https://doi.org/10.1080/01431160600589179. Yamazaki, Dai, Daiki Ikeshima, Ryunosuke Tawatari, et al. 2017.“A High-Accuracy Map of Global Terrain Elevations.”Geophysical Research Letters44 (11): 5844– 53.https://doi.org/10.1002/2017GL072874. Yin, Jie, Dapeng Yu, Zhane Yin, Min Liu, and Qing He. 2016.“Evaluating the Impact and Risk of Pluvial Flash Flood on Intra-Urban Road Network: A Case Study in the City Center ofShanghai,China.”Journal of Hydrology537: 138– 45.https://doi.org/10.1016/j.jhydrol.2016.03.037.