Physics-Informed Neural Network (PINN) is a powerful, mesh-free tool for predicting contamination dispersion, but they often struggle with the unknown boundary conditions common in real-world scenarios. This work presents a novel framework that overcomes this limitation by integrating sparse sensor data with a perceptually-aware evaluation metric. We introduce a systematic sensor placement strategy that significantly reduces the required instrumentation while maintaining high predictive accuracy. To better assess performance, we utilize a perceptual LAB error metric that more closely aligns with visual fidelity than traditional MSE. Critically, we validate the framework’s geometric robustness across six distinct obstacle configurations, including shifts and rotations, proving its adaptability with a single network architecture. Together, these innovations provide a unified and efficient method for real-world dispersion prediction using minimal field instrumentation.