Comparative Regression Modeling of Sewer and Stormwater Pipe Lengths and Manhole Counts: Cross-Validated Performance and Sensitivity Analysis

Existing residential communities may have readily available basic property characteristics but incomplete drainage-asset inventories, limiting record completeness checks and prioritization of field verification. This study compared multiple linear regression (MLR), support vector regression (SVR), and random forest regression (RFR) for estimating existing total sanitary sewer length (m), total stormwater pipeline length (m), sewage manhole count, and stormwater manhole count in 120 Shanghai communities. Models were evaluated using shuffled 10-fold cross-validation and pooled out-of-fold predictions, with preprocessing confined to training folds. Metric uncertainty was estimated from 10,000 paired case-bootstrap resamples, and pairwise differences were tested using randomization tests with Holm-adjusted p values. In the primary seed-42 partition, RFR had the lowest sanitary sewer RMSE (939.47 m), but pairwise differences were not statistically supported and the ranking was partition sensitive. MLR had the lowest RMSE for the other three targets and ranked first for all three in all 10 repeated partitions. Replacing observed sanitary sewer length with nested cross-fitted estimates increased stormwater pipeline RMSE by 99.9%, while manhole results depended on the error metric. The models support only provisional inventory screening and record-review prioritization; they do not provide design quantities and require independent validation before wider use.