#712: Exploring Object Detection Approaches to Analyze Students’ Drawn Scientific Models

With the increasing emphasis of knowledge-in-use, students are expected to apply knowledge of three-dimensional learning. Validly assessing this complexity necessitates the use of learning progressions, which provide developmental learning pathways for student growth. Although AI-based analysis of student-generated models shows promise, current classification methods often lack the transparency needed to align AI-generated scores with feature-based rubrics that reflect learning progressions. Specifically, they fail to provide interpretable information on specific symbols and conceptual components of student models. To address this, we propose an image recognition scoring system with object detection and scoring model to evaluate student-generated model. This dual-model architecture precisely identifies and localizes specific symbols within students’ models to derive student performance. This system is designed to enhance the accuracy (range from 0.96 to 0.99), consistency (Kappa values range from 0.78 to 0.92), and, critically, the interpretability of scoring, thus building a robust validity argument for complex three-dimensional learning performance.