How to Spot Areas Lacking Public Transport? Discover Transit Deserts in Lucknow

Cities are growing fast, but not everyone enjoys the same right to move around. In India, in Lucknow, researchers uncovered "transit deserts": areas where public transport is far too scarce compared to demand. To locate them, they turned to artificial intelligence, i.e., programs that learn from data. They tested three methods: decision trees (DT), logistic regression (LR) and random forest (RF). The study used a 100 × 100 meter grid, evaluating each cell based on income, population density, trip frequency and service access. The outcome? Random forest achieved the highest classification accuracy, while decision trees offered a simpler explanation, albeit with slightly lower precision. Logistic regression lagged behind because it assumes overly simple linear relationships. The maps revealed that transit deserts cluster in poorer neighborhoods, creating a vicious cycle of inequality. With this information, planners can target interventions where they matter most, adding bus lines or improving stops. In practice, technology helps make mobility fairer for everyone, turning numbers and algorithms into concrete actions for the city.

The random forest (RF) model combines hundreds of decision trees, each built on a random subset of variables and samples. This diversity reduces overfitting, enhancing generalization on unseen data. In Lucknow's context, RF effectively handled the non‑linearity between socio‑economic variables (average income, housing density) and accessibility metrics (stops per km², service frequency). The algorithm also provided feature importance scores, highlighting population density and income as the most influential predictors for classifying transit deserts. Decision trees (DT), while more interpretable thanks to their if‑then rule structure, exhibit higher variance: small changes in training data can significantly alter the tree structure, reducing prediction stability. Logistic regression (LR) assumes a linear relationship between independent variables and the log‑odds of the target class, limiting its ability to capture complex interactions typical of urban phenomena. Moreover, LR is sensitive to multicollinearity, a common issue when variables like population density and vehicles per capita are correlated. Validation employed a 10‑fold cross‑validation approach, ensuring results were not due to a specific data split. Performance metrics included accuracy, recall, precision, and F1‑score, with RF achieving an F1‑score of 0.89, compared to 0.78 for DT and 0.62 for LR. These findings suggest that for high‑dimensional, non‑linear classification problems like transit deserts, tree‑based ensembles are the most robust choice.