Real-time prediction of spatial information such as road-traffic-related information has attracted much attention. Mobile crowdsensing (MCS), in which mobile user devices such as smartphones equipped with sensors work as distributed mobile sensors, is an effective way of collecting sensor data for real-time prediction of spatial information. Since user devices contributing to MCS incur various costs including energy cost and privacy risk, using incentive mechanisms is one approach to compensate for these costs. However, since, in general, the budget for incentive rewarding is limited, rewards should be effectively allocated with considering the contribution of sensor data to the accuracy in real-time prediction of spatial information, which has not been considered in any prior work. This paper presents a scheme to maximize the accuracy of real-time prediction when allocating incentive rewards to user devices. The proposed scheme estimates the contribution of each user device collecting and sending sensor data to the prediction accuracy. Then, the incentive reward received by a user device collecting and sending data increases in proportion to the contribution the data makes to prediction accuracy. Feature selection is used to extract the contribution of each input data point from a prediction model created by machine learning. Evaluation using a real road-traffic-related dataset demonstrated that the proposed scheme works better in terms of prediction accuracy for various cost conditions than a benchmark scheme.