fuxicfd-model/inference_example/utils/preprocessing.py

import numpy as np
from scipy.ndimage import zoom

def build_model_input(input_dict: dict, in_stats: dict):
    # Raw inputs
    dem = input_dict["dem"]
    rough = input_dict["roughness"]
    u100 = input_dict["u_100m"]
    v100 = input_dict["v_100m"]

    dem_rough = np.stack([dem, rough], axis=0)          # (2,301,301)
    uv_100m   = np.stack([u100, v100], axis=0)          # (2,  9,  9)

    # Normalize
    high_mean = in_stats["high_mean"][:, None, None]
    high_std  = in_stats["high_std"][:, None, None]
    low_mean  = in_stats["low_mean"][:, None, None]
    low_std   = in_stats["low_std"][:, None, None]

    dem_rough = (dem_rough - high_mean) / high_std
    uv_100m   = (uv_100m   - low_mean)  / low_std

    # Resample to (300,300) like the original script
    dem_rough = zoom(dem_rough, (1, 300/301, 300/301), order=1)   # -> (2,300,300)
    uv_100m   = zoom(uv_100m,   (1, 300/9,   300/9),   order=1)   # -> (2,300,300)

    x = np.concatenate([uv_100m, dem_rough], axis=0).astype(np.float32)  # (4,300,300)
    return x