Investigating the Potential of Nighttime Light Data to Estimate Travel Demand

Taking the bike‐sharing travel demand (BSTD) as an example, this study investigates the potential of Nighttime Light (NTL) data to optimize forecasting performance and replace the land use factors. Stepwise regression is trained with the travel demand in each unit as the dependent variable, and land use factors are introduced as the independent variable one by one, which finds the set of independent variables. Five machine learning algorithms driven by ensemble learning and decision trees including the GBDT, Random Forecast, Adaboost, Extratrees, and Catboost, are employed and evaluated to achieve comparative analysis of “before considering‐after considering NTL data”. The methodological verification of Beijing city shows: (1) Adaboost and GBDT are superior to all other algorithms, since they generally have the highest R2, lowest RMSE, and lowest absolute MAPE. (2) All methods by employing NTL data obviously optimize the performance of BSTD forecast with decreased RMSE, decreased MAPE, etc. In particular, GBDT performs the best in reducing MSE, with a percentage of −99.99% in the training set and −86.985% in the test set, which AdaBoost, Extratrees, and Catboost follow. (3) Land use factors no longer make sense in predicting BSTD after employing NTL data, and NTL data has covered the roles of land use factors to ensure accuracy. The conclusions presented here enrich our understanding of the relative roles of land use factors and NTL data in travel demand and boost our optimization in traffic prediction in the future.