Improving Neural ODEs Explainability: A Two-Stage Training Method for Modeling Constrained Natural Systems