BikeshareGAN: Predicting dockless bike-sharing demand based on satellite image

Abstract

Understanding the drop-off demand of Dockless Bikeshare Systems (DBS) is crucial for efficient urban management but has long been challenging. Conventional prediction models are mostly regression-based, requiring multisource and fine-grained GIS data (e.g., socio-demographics, land use, POI), whose collection could be laborious and costly. Some data do not even exist for fast-growing cities in the developing world, largely hindering the application of the conventional models. Noting that high dimensional satellite images contain rich data about complex urban systems (e.g., density, land use, transportation network), we hypothesize that Generative Adversarial Networks (GAN) can embed inherent urban features as the latent space, to predict DBS demand directly from satellite images effectively. To test the hypothesis, we took Shenzhen - a city with diverse urban forms as a case study. Pairwise satellite image and DBS drop-off heatmap during AM/PM and non-peak hours on a random workday became the input and output images for Pix2Pix, a proven GAN framework, to train the image-to-image translation at the 200 m level. Fake heatmaps were generated and validated by ground truth using loss functions including L1, L2, and Structure Similarity Index Measure (SSIM). R² was also calculated to compare our pixelated results to conventional regression models. First, simply taking a satellite image as the input achieved ∼0.49 R² (82 % SSIM), outperforming many regression-based models that require a bunch of numeric/vector inputs. Moreover, pixelating vector maps (e.g., metro station, road network, office building) onto satellite images significantly improved the accuracy (∼0.56 R²/90 % SSIM), outperforming some machine learning or hybrid deep learning models in this regard (R² 0.18–0.76). Therefore, GAN is plausible to predict DBS demand from solely satellite images, while feeding more urban layers significantly improves the predictive power. Our raster-oriented framework can effectively aid the decision-making process for DBS implementation and operation in developing countries where up-to-date GIS data is less accessible.