The rationality and the effectiveness of classroom teaching behavior directly influence the quality of classroom instruction. Analyzing teaching behavior intelligently can provide robust data support for teacher development and teaching supervision. By observing verbal and nonverbal behaviors of teachers in the classroom, valuable data on teacher-student interaction, classroom atmosphere, and teacher-student rapport can be obtained. However, traditional approaches of teaching behavior analysis primarily focus on student groups in the classroom, neglecting intelligent analysis and intervention of teacher behavior. Moreover, these traditional methods often rely on manual annotation and decision making, which are time consuming and labor intensive, and cannot efficiently facilitate analysis. To address these limitations, this article proposes an innovative automated multimode teaching behavior analysis framework, known as AMTBA. First, a model for segmenting classroom events is introduced, which separates teacher behavior sequences logically. Next, this article utilizes deep learning strategies with optimal performance to conduct multimode analysis and identification of split classroom events, enabling the fine-grained measurement of teacher's behavior in terms of verbal interaction, emotion, gaze, and position. Overall, we establish a uniform description framework. The AMTBA framework is utilized to analyze eight classrooms, and the obtained teacher behavior data are used to analyze differences. The empirical results reveal the differences of teacher behavior in different types of teachers, different teaching modes, and different classes. These findings provide an efficient solution for large-scale and multidisciplinary educational analysis and demonstrate the practical value of AMTBA in educational analytics.