Spaces:
Paused
Paused
| """ | |
| Virtual Sensor Profile Interpolation Validator | |
| =============================================== | |
| Evaluates spatial-depth interpolation fidelity against the | |
| Calibrated Physical Reference Model based on TEOS-10 and Southern Ocean climatology. | |
| Provides academic transparency regarding climatological baseline reference datasets. | |
| """ | |
| import numpy as np | |
| import xarray as xr | |
| from sklearn.metrics import mean_absolute_error | |
| from scipy.stats import pearsonr | |
| import os | |
| import sys | |
| sys.path.append(os.path.dirname(__file__)) | |
| from profile_interpolator import ProfileInterpolator | |
| def validate_interpolation_accuracy( | |
| dataset_path: str = os.path.join(os.path.dirname(__file__), "..", "data", "argo_southern_ocean.nc"), | |
| test_fraction: float = 0.2 | |
| ) -> dict: | |
| """ | |
| Splits profiles 80/20, trains interpolator on 80%, validates on 20%. | |
| Reports MAE, RMSE, and R-squared for each parameter against the | |
| Calibrated Physical Reference Model based on TEOS-10 and Southern Ocean climatology. | |
| This benchmark quantifies profile spline interpolation fidelity against | |
| the calibrated hydrographic reference model baselines. | |
| """ | |
| if not os.path.exists(dataset_path): | |
| print(f"Error: {dataset_path} not found. Please run datasets/fetch_argo.py first.") | |
| return {} | |
| ds = xr.open_dataset(dataset_path) | |
| n = len(ds.N_PROF) | |
| if n < 5: | |
| print("Not enough profiles in dataset to run validation split.") | |
| return {} | |
| test_idx = np.random.choice(n, max(1, int(n * test_fraction)), replace=False) | |
| train_idx = np.setdiff1d(np.arange(n), test_idx) | |
| interpolator = ProfileInterpolator.__new__(ProfileInterpolator) | |
| interpolator.ds = ds.isel(N_PROF=train_idx) | |
| interpolator._build_profile_index() | |
| results = {} | |
| params = ["DOXY", "CHLA", "PH_IN_SITU_TOTAL", "NITRATE"] | |
| for param in params: | |
| y_true, y_pred = [], [] | |
| for idx in test_idx[:50]: # limit for speed | |
| profile = ds.isel(N_PROF=idx) | |
| lat = float(profile.LATITUDE) | |
| lon = float(profile.LONGITUDE) | |
| nearest = interpolator.select_nearest_profile(lat, lon) | |
| pressures = profile.PRES.values | |
| values = profile[param].values | |
| valid = ~(np.isnan(pressures) | np.isnan(values)) | |
| for p, v in zip(pressures[valid], values[valid]): | |
| pred = interpolator.get_value_at_depth(nearest, p, param) | |
| if pred is not None: | |
| y_true.append(v) | |
| y_pred.append(pred) | |
| if len(y_true) > 10: | |
| y_true = np.array(y_true) | |
| y_pred = np.array(y_pred) | |
| rmse = np.sqrt(np.mean((y_true - y_pred) ** 2)) | |
| mae = mean_absolute_error(y_true, y_pred) | |
| # handle case where predictions might be constant yielding nan R2 | |
| std_true = np.std(y_true) | |
| std_pred = np.std(y_pred) | |
| if std_true > 0 and std_pred > 0: | |
| r2 = pearsonr(y_true, y_pred)[0] ** 2 | |
| else: | |
| r2 = 0.0 | |
| results[param] = { | |
| "MAE": round(mae, 4), | |
| "RMSE": round(rmse, 4), | |
| "R2": round(r2, 4), | |
| "n_points": len(y_true) | |
| } | |
| return results | |
| if __name__ == "__main__": | |
| results = validate_interpolation_accuracy() | |
| if results: | |
| print("\nVirtual Sensor Validation Results") | |
| print("Dataset: Calibrated Physical Reference Model based on TEOS-10 and Southern Ocean climatology") | |
| print("Method: 80/20 profile split, cubic interpolation\n") | |
| for param, metrics in results.items(): | |
| print(f"{param}") | |
| print(f" MAE: {metrics['MAE']}") | |
| print(f" RMSE: {metrics['RMSE']}") | |
| print(f" R2: {metrics['R2']}") | |
| print(f" N: {metrics['n_points']}\n") | |