Identification of land-use characteristics using bicycle sharing data: A deep learning approach

Extensive research has shown that urban land-use characteristics, including resident, work, consumption, transit, etc., are significantly interrelated with travel behaviors and travel demands. Many research efforts have been made to evaluate the impact of land use planning or policies on travel behavior, however, few studies are able to quantitatively measure the land-use characteristics based on the data of travel behaviors or travel demand. In this paper, a new hybrid model that combines time series feature extraction and deep neural network is proposed to identify regional land use characteristics and quantify land use intensity using ridership data of bicycle sharing. This method consists of four main parts: (i) A set of land-use characteristic labels are evaluated based on planning and Geographic Information System (GIS) data. (ii) An ensemble clustering method is used to determine the segmentation points of ridership time series. (iii) The statistical characteristics of the segmented time series are extracted and used as input to the neural network. (iv) A deep neural network is established and trained based on the processed ridership features and land-use labels. In terms of data collection, ridership data of the bicycle-sharing parking spots and land-use planning data are obtained from bicycle-sharing system and planning department in San Francisco Bay Area, California U.S.A., respectively. The test results show that this approach has high accuracy for identifying land-use characteristics based on several standard evaluation measures and that the identification distribution can be well explained. The extension results further prove that the model can be applied to effectively analyze the main land-use characteristics of the region although the identification results may become unstable after 3–4 months.