This study presents a novel approach to bicycle volume estimation using data fusion techniques that integrate crowdsourced data (e.g., Strava) with traditional count data. We aim to improve the accuracy and robustness of bicycle volume estimates in different regions by using deep learning algorithms, especially Long Short-Term Memory networks and Deep Neural Networks. Our method involves combining static variables, such as network characteristics, demographics, and land use dynamic crowdsourced data, count data, and static data from different regions. The results show that our models, particularly when using Monthly Average Daily Bicyclists (MADB) as the target variable, provide superior generalization and transferability compared to traditional Annual Average Daily Bicyclists (AADB) models. This increased accuracy has significant implications for urban planning and infrastructure development, enabling more effective allocation of resources to improve bicycle infrastructure.