Student engagement in science classroom is an essential element for delivering effective instruction. However, the popular method for measuring students' emotional learning engagement (ELE) relies on self-reporting, which has been criticized for possible bias and lacking fine-grained time solution needed to track the effects of short-term learning interactions. Recent research suggests that students' facial expressions may serve as an external representation of their emotions in learning. Accordingly, this study proposes a machine learning method to efficiently measure students' ELE in real classroom. Specifically, a facial expression recognition system based on a multiscale perception network (MP-FERS) was developed by combining the pleasure-displeasure, arousal-nonarousal, and dominance-submissiveness (PAD) emotion models. Data were collected from videos of six physics lessons with 108 students. Meanwhile, students' academic records and self-reported learning engagement were also collected. The results show that students' ELE measured by MP-FERS was a significant predictor of academic achievement and a better indicator of true learning status than self-reported ELE. Furthermore, MP-FERS can provide fine-grained time resolution on tracking the changes in students' ELE in response to different teaching environments such as teacher-centered or student-centered classroom activities. The results of this study demonstrate the validity and utility of MP-FERS in studying students' emotional learning engagement.