pnit_data

The pnit_data contains preprocessed terrain profiling datasets for propagation loss prediction, organized into 12 compressed NumPy archive files covering 6 distinct geographic measurement campaigns, each split into training and testing subsets. The six datasets are 1_MT, 2_OH, 3_PL, R1, and R2 from measurement set 1, plus xbp from an independent measurement set 2. Each dataset has a train and a test file following the naming pattern {dataset}_{split}_comprehensive_feature.npz, e.g., 1_MT_train_comprehensive_feature.npz and 1_MT_test_comprehensive_feature.npz. Each .npz file stores three components. The feature array is a 5-channel terrain feature tensor of shape (500, 5, N), where 500 denotes the number of profile points sampled along the propagation path, 5 denotes the feature channels, and N is the number of samples in that split. The Loss array of shape (N,) holds the excess propagation loss values serving as the prediction targets. A metadata string feature_info documents the channel composition as "5ch: main + dist + freq + th + rh". The five feature channels encode the following information: channels 0 and 1 capture implicit frequency-encoded terrain features for the main and distance propagation paths, respectively, extracted using Fresnel-zone-based profiling; channel 2 stores the transmitter antenna height (th); channel 3 stores the receiver antenna height (rh); and channel 4 stores the communication frequency (f). These comprehensive features are designed to feed the PNIT (Physics-Normalized Terrain Profiling with Implicit Frequency Encoding) model for learning the mapping from terrain profiles to excess propagation loss.