This study focuses on the State of Delaware, the same study area evaluated geospatially in the companion analysis of NMTF exposure to coastal inundation [7]. A public survey was designed and administered by the University of Delaware research team under the SMARTER Center project. Distribution targeted Delaware and nearby communities that regularly use, or are affected by, active transportation and micromobility systems, including university channels at the University of Delaware. Participation in the survey was voluntary, and respondents were free to skip any question. A total of 200 valid responses were collected, providing a dataset suitable for both descriptive analysis and supervised machine learning. Eligibility was limited to adults aged 18 or older who live, work, or regularly travel within Delaware; no compensation was offered for participation. Recruitment relied primarily on University of Delaware email listservs and social media channels, supplemented by outreach through SMARTER Center community partners, so the sample constitutes a convenience sample of engaged residents and university-affiliated individuals rather than a probability sample of the statewide population; all descriptive and predictive results reported in this paper should accordingly be interpreted as perceptions within this convenience sample rather than as statewide estimates for Delaware. The online survey platform permitted only one submission per unique link/session, and the research team additionally reviewed all responses for duplicate or near-duplicate response patterns prior to analysis; no duplicate submissions were identified. Of the responses received, 200 were retained as valid after removing incomplete or disqualified entries; the analytic sample size for each classification model (n = 200 for both Model 1 and Model 2) reflects only respondents who provided a codable answer to the corresponding outcome question (Q12 for Model 1, Q11 for Model 2), and no missing-response imputation was performed for either model.

The instrument was organized into five thematic sections combining closed-ended multiple-choice, Likert-frequency, and rating-scale items with optional open-ended comment fields. Table 1 summarizes the structure of the instrument.

Table 1.Survey instrument structure.

Responses to the three environmental frequency items (Q5-Q7, coastal inundation, temperature, and flooding, respectively) were later aggregated into a composite Environmental Impact Score used throughout the modeling stage, reflecting the cumulative environmental disruption experienced by each respondent across all three stressor types.

Raw survey responses were transformed into a numerical feature set suitable for supervised classification. Ordinal items were encoded to preserve their natural ordering, binary items were dummy-coded, and the three environmental frequency items were summed into a composite score ranging from 0 to 12. Table 2 summarizes the resulting feature set.

Table 2.Engineered features used in the machine learning models.

Note (added in revision): as discussed in Section 3.4, listing Micromobility Availability (Q11) as a predictor for Model 2 constitutes target leakage, since Q11 also defines the Model 2 outcome; likewise, including the Environmental Impact Score alongside its three constituent items (Q5 - Q7) introduces collinearity in Model 1. Both issues are flagged as limitations in Section 5.3 pending a future re-analysis with corrected predictor sets.

Two binary classification tasks were defined. Model 1 predicts respondent-perceived safety of active transportation from a traffic standpoint (Q12); “Yes” responses were coded 1 (safe) and “No” or ambiguous/qualified responses were conservatively coded 0 (unsafe), yielding a 63%/37% class split. Model 2 predicts perceived micromobility availability (Q11), binarized into High availability (Somewhat or Very available, coded 1) versus Low availability (Not very available or Not available at all, coded 0), yielding a 30%/70% class split. Two clarifications and one methodological limitation, raised in peer review, are addressed here. First, regarding the Q12 safety outcome: the descriptive breakdown in Table 3 (62% safe/33% unsafe/~5% conditional or route-dependent) is computed over all 200 respondents using three original response categories, whereas the 63%/37% split reported above reflects the binary recoding used for modeling, in which conditional/route-dependent responses were merged into the “unsafe” category together with “No” responses and a small number of respondents whose answer could not be unambiguously classified were excluded from the modeling denominator; these differences in category definition and denominator, rather than a data error, account for the shift from 62/33/5 to 63/37, and both values are retained here for transparency. Second, reviewers correctly noted that Micromobility Availability (Q11) is used both as a predictor in Table 2 and as the basis for the Model 2 outcome, which constitutes direct target leakage; this means the 65.50% accuracy reported for Model 2 should be interpreted as an inflated upper bound rather than a valid, unbiased estimate of predictive performance. Similarly, reviewers noted that including the composite Environmental Impact Score alongside its three constituent items (Coastal Inundation, Temperature Impact, and Flooding Frequency; Q5 - Q7) introduces collinearity by construction, which likely inflates the apparent importance attributed to environmental exposure in Model 1’s feature-importance ranking presented later in Section 4.3. Correcting both issues requires retraining each classifier on the original dataset, which could not be completed within the scope of this revision; we have therefore not altered the reported values in Table 4 and Table 5, but flag both issues explicitly here and in the Limitations subsection (5.3), and we revise the interpretive language in Section 4.2 and 4.3 accordingly so that these results are not overstated. A corrected re-analysis excluding Q11 from the Model 2 predictor set and using the composite Environmental Impact Score alone in Model 1 is identified as a priority for future work.

A Random Forest Classifier (100 trees, maximum depth 5) was selected as the primary algorithm for both tasks because ensemble tree methods reduce overfitting relative to a single decision tree and remain robust with modest sample sizes such as the 200-respondent dataset used here [13]. A Decision Tree Classifier (maximum depth 4) was trained on each task as a simpler interpretable baseline for comparison [14]. Both classifiers were evaluated using five-fold stratified cross-validation, which partitions the data into five equally sized, class-balanced folds, trains on four folds, and tests on the held-out fifth fold across five repetitions, reporting the mean accuracy and standard deviation across folds. Feature importance for the Random Forest models was computed as the mean decrease in Gini impurity attributable to each feature across all trees in the ensemble, providing a ranked, interpretable measure of each variable’s contribution to model predictions.

The survey sample skews young and higher-income: approximately 95.5% of respondents fall in the 18 - 34 age bracket, and 60% report household incomes of $75,000 or more, a pattern consistent with distribution through University of Delaware channels. In terms of travel mode, roughly 57% of respondents primarily combine car use with walking and 29% rely on the car alone, reflecting the largely suburban, car-oriented character of much of Delaware. Table 3 summarizes the key descriptive results across the environmental, urbanization, and infrastructure sections of the survey.

Table 3.Key descriptive findings from the 200-response survey.

Two patterns stand out from the descriptive results. First, although coastal inundation is the environmental stressor most likely to cause catastrophic, localized infrastructure loss, as shown by the geospatial companion analysis, it is temperature that most consistently disrupts residents’ day-to-day travel routines, reflecting its chronic, season-long character. Second, the survey data show a strong and consistent preference for physical infrastructure investment (bike lanes, drainage, protected facilities) over financial subsidies or awareness campaigns, cited by roughly three-quarters to over ninety percent of respondents depending on the question.

Table 4 reports the five-fold stratified cross-validation accuracy for both classification tasks and both algorithms.

Table 4.Model performance (mean cross-validation accuracy).

*This value is affected by the collinearity issue noted in Section 3.4 (composite Environmental Impact Score reported alongside its Q5 - Q7 components); see Section 5.3 for discussion. **This value is affected by the Q11 target-leakage issue noted in Section 3.4 and should be treated as an inflated upper bound rather than a valid, unbiased estimate of predictive performance; see Section 5.3 for discussion. A Decision Tree result for Model 2 is not reported, as this comparison was not performed in the original analysis (see Section 5.3).

For Model 1, the Random Forest outperforms the Decision Tree baseline by approximately 5.4 percentage points and is competitive with the majority-class baseline of roughly 63%, while retaining the ability to identify the minority “unsafe” class, which a naive majority-class classifier cannot do. The low cross-fold standard deviation (±3.98%) indicates a stable model that is not overfitting to any particular partition of the data. For Model 2, the Random Forest achieves a notably higher accuracy of 65.50%, suggesting that environmental and demographic variables explain geographic variation in infrastructure availability, an objective, largely supply-side outcome, somewhat more reliably than they explain the more subjective, individually variable outcome of perceived safety. These accuracy levels are consistent with the broader travel-behavior machine learning literature, in which ensemble tree methods applied to survey-scale datasets typically achieve moderate but meaningful gains over baseline and simpler-tree comparisons [14][15]. A more precise benchmarking comparison, added in response to peer review, qualifies the interpretation above: both classifiers’ raw cross-validated accuracies fall at or below their respective majority-class baselines (approximately 63% for Model 1 and 70% for Model 2), so raw accuracy alone does not demonstrate predictive utility over a naive majority-class classifier, and the statement that Random Forest is “competitive with” the Model 1 majority-class baseline is accordingly withdrawn. Reviewers additionally requested balanced accuracy and minority-class precision, recall, and F1-score for both tasks and both classifiers; these metrics were not computed in the original analysis and, absent access to the underlying predictions, cannot be added at this stage. We flag this as a limitation in Section 5.3 and note that the modest gains reported here over majority-class baselines should accordingly be interpreted cautiously, as reflecting exploratory association rather than demonstrated predictive utility, pending a fuller benchmarking analysis in future work.

Table 5 presents the ranked Random Forest feature importances for Model 1 (perceived safety), based on mean decrease in Gini impurity.

Table 5. Top predictors of perceived active-transportation safety (Model 1).

Note (added in revision): as discussed in Section 3.4, rows 2, 4, and 6 above (Flooding, Temperature, and Coastal Inundation Frequency) are individual components of row 1’s Environmental Impact Score, so their importance values are not independent of one another; the ranking should be read as showing that cumulative and flooding-related environmental exposure are jointly influential, rather than as four independently ranked environmental predictors. This collinearity is flagged as a limitation in Section 5.3.

The remaining features (Noticed Environmental Changes, Car-Only User, Population Growth, and Age Group) contributed smaller but non-negligible importance (approximately 0.02 - 0.09 each), with car-only users systematically more likely to report feeling unsafe, consistent with a self-reinforcing cycle in which perceived risk discourages a shift away from car dependency. For Model 2 (micromobility availability), the leading predictors were income level, with higher-income respondents significantly more likely to report available shared bikes and scooters, and the Environmental Impact Score, which was negatively associated with availability, suggesting that environmentally exposed areas tend to receive less micromobility investment. Car-only use and flooding frequency were also significant negative predictors, the latter plausibly reflecting the physical impracticality of deploying docked or dockless vehicles in flood-prone locations.

Table 6 synthesizes the descriptive and predictive findings into planning-relevant implications.

Table 6.Integrated findings and planning implications.

The survey-based results reinforce and extend the geospatial vulnerability findings of the companion study. The geospatial analysis identified extensive intersections between NMTF corridors and projected flood zones, particularly in Sussex County, and quantified the length and depth of anticipated inundation under low, medium, and high coastal inundation scenarios [7]. The present survey results independently confirm that flooding is the single most influential environmental predictor of perceived safety, a convergence between physically measured exposure and human-perceived risk that is not guaranteed a priori, since perception can lag or diverge from physical reality. The fact that both analyses converge on flooding as the primary driver of vulnerability strengthens the case for prioritizing flood-resilient design in Delaware’s active transportation investment strategy. At the same time, the survey reveals a dimension not captured by the geospatial model alone: temperature’s outsized role in day-to-day disruption, and income’s strong association with micromobility availability, both of which point to social and behavioral vulnerabilities that a purely physical exposure model would miss.

Several limitations qualify these findings. The survey sample is heavily skewed toward respondents aged 18 - 34 and toward higher-income households, largely a function of distribution through university channels, which limits generalizability to older and lower-income populations who may have different risk perceptions and mobility constraints. The classification models are exploratory and identify statistical association rather than causal relationships; the Environmental Impact Score captures frequency of disruption but not its severity or duration, and the safety outcome reflects perceived rather than objectively measured risk. The reported cross-validation accuracies (61% - 65%) are consistent with the inherent noise of survey-based behavioral outcomes and should be interpreted as realistic benchmarks rather than shortcomings; comparable ensemble tree-based studies of travel behavior report accuracies in a similar range when working from survey-scale samples [14][15]. Future work should incorporate stratified sampling to improve demographic representativeness and should fuse geospatial variables, such as proximity to mapped flood zones, bike-lane density, and heat-island intensity, directly into the feature set to create a richer, multi-source predictive model. Three additional limitations, raised in peer review, are noted here. First, Micromobility Availability (Q11) is listed as a predictor in Table 2 and is also used to construct the Model 2 outcome (Section 3.4), constituting direct target leakage; the reported 65.50% cross-validated accuracy for Model 2 should therefore be interpreted as an inflated upper bound rather than a valid, unbiased estimate of predictive performance, pending a re-analysis that excludes Q11 from the Model 2 predictor set. Second, Model 1 includes the composite Environmental Impact Score alongside its three constituent items (Q5 - Q7), which are collinear by construction; this likely inflates the combined apparent importance of environmental exposure in Table 5 and should be corrected in future work by using the composite score alone. Third, model comparisons in this study rely on raw cross-validated accuracy rather than majority-class-baseline-normalized metrics; both reported Random Forest accuracies (61.29% and 65.50%) are at or below their respective majority-class baselines (approximately 63% and 70%), and a Decision Tree result for Model 2 was not computed, so predictive utility beyond the majority-class baseline is not yet demonstrated for either task. None of these three issues could be corrected within the scope of the present revision, as doing so requires retraining the classifiers on the original survey dataset; we report them transparently here so that Table 4 and Table 5 are interpreted appropriately, and we identify a corrected re-analysis, including balanced accuracy and minority-class precision, recall, and F1-score, as a priority for a follow-up study.