Accurate prediction of receiver state is vital for optimizing network performance in urban settings, where rapid spatial variations in channel conditions pose significant challenges to communication quality. This paper presents a Machine Learning-based framework for predicting channel states in Unmanned Aerial Vehicle-assisted mmWave communication networks. Given that mmWave signals are susceptible to blockage by buildings and other urban structures, predicting receiver conditions at a specific location can be determined by directly deploying the geometric features describing the built-up environment surrounding the receiver. A set of geometrical features is extracted and used as input to train the adopted learning models, namely Decision Tree, Linear Decision Tree (LDT), Random Forest, Support Vector Machine, and Deep Neural Network (DNN), to estimate the probability of three distinct receiver states: Line-of-Sight, Non-Line-of-Sight, and Blocked. Experimental results indicate that the DNN-based model achieves the highest prediction accuracy and robustness, while the LDT provides computational efficiency and straightforward explainability. To improve the interpretability of the black-box DNN model, we employ the SHapley Additive exPlanations (SHAP) method, which identifies the most influential environmental features in state probability prediction. Furthermore, we enrich the standard 3GPP model by incorporating the top SHAP-ranked features, leading to notable performance improvements.