Abstract. Real-time, accurate taxi demand prediction plays an important role in intelligent traffic system. It can help manage taxi patching and minimize the time and energy waste caused by waiting. In the era of big data, a diversity of urban data and increasingly complex traffic data have been collected and published. Traditional forecasting methods have been unable to cope with the heterogeneous massive traffic data, whereas deep learning, as a new data-oriented technique, has been widely used in the field of traffic prediction. This paper aims to utilize multisource data and deep learning techniques to improve the accuracy of taxi demand prediction. In this paper, a joint guidance residual network JG-Net is proposed for city-scale taxi demand prediction. Taxi order data and multiple urban geospatial data POI, road network and population distribution data) are integrated into the JG-Net. Regional features are considered in the prediction process by three guidance branches composed of pixel-adaptive convolutional networks, each of which applies one type of urban data. JG-Net assigns learnable weights to different branches and regions to combine the output of the branches, then further aggregates weather and time information to forecast the taxi demand. Extensive experiments and analyses are conducted, which show our method outperforms traditional methods. The mean square error of the prediction result on the testing set is 1.868, which is 12.3% lower than related models. The positive influence of combining multiple geospatial data is also validated by ablation experiments.