In recent years, the study of university students’ classroom learning behaviors has gained significant attention due to the growing importance of personalized education and effective teaching strategies. Traditional methods of behavior observation and analysis have limitations in capturing diverse and nuanced behaviors in the classroom setting. To address this, advanced techniques such as deep learning and data fusion have emerged as promising approaches for behavior classification. The research aims to collect and analyze multiple sources of classroom videos to gain a comprehensive understanding of students’ behaviors during class. The collected data is then processed and classified using deep learning techniques. In this study, an approach is proposed for categorizing how college students behave when studying based on multi-source data. The first contribution of this study is a more effective algorithm for sampling video. The program evaluates the incoming video sequence to provide a baseline for comparison, from which a sampling strategy may be developed. Second, this research provides a technique of classifying student behavior in the classroom in light of the issue of limited spatial feature extraction capacity. To improve the model’s capability to extract spatial discriminative features, a residual network is constructed and inserted into the convolutional attention module during the spatial feature extraction phase. Given the issues with convolutional attention module training, this research additionally enhances the channel attention component.