ted-downstream-eo / examples /classification_cube.py
ZiyunPOLYU's picture
Add downstream readout examples
cc74c5e verified
Raw History Blame Contribute Delete
3.62 kB
#!/usr/bin/env python3
"""Read a task-ready classification cube and fit a temporal-mean linear probe.
``data`` is float32 with shape (time, band, sample). The first seven bands are
Blue, Green, Red, NIR, SWIR1, SWIR2, NDVI, in DN x10000 except NDVI. Empty
steps are NaN. ``sensor`` is 0 empty, 1 S30, 2 L30.
CityRep and LUCAS store ``split``. The five older classification cubes do not;
their published splits are spatial blocks and are not in the file. This example
uses the stored split and does not rescale features.
python classification_cube.py --dataset cityrep --limit 4000
"""
from __future__ import annotations
import argparse
import numpy as np
REPO = "ZiyunPOLYU/ted-downstream-eo"
CUBES = {
"cityrep": "ready_3day/cityrep_fullband_3day.nc",
"lucas": "ready_3day/lucas_fullband_3day.nc",
"cdl": "ready_3day/cdl_fullband_3day.nc",
"cropharvest": "ready_3day/cropharvest_fullband_3day.nc",
"globaltree": "ready_3day/globaltree_fullband_3day.nc",
"glance": "ready_3day/glance_fullband_3day.nc",
"lcmap": "ready_3day/lcmap_fullband_3day.nc",
}
COMMON = 7
def _open(dataset: str, path: str):
import netCDF4
from huggingface_hub import hf_hub_download
local = path or hf_hub_download(REPO, CUBES[dataset], repo_type="dataset")
return netCDF4.Dataset(local)
def _mean_features(ds, limit: int) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
n = ds.dimensions["sample"].size if limit <= 0 else min(limit, ds.dimensions["sample"].size)
cube = np.asarray(ds.variables["data"][:, :COMMON, :n], dtype=np.float32)
with np.errstate(all="ignore"):
features = np.nanmean(cube, axis=0).T
labels = np.asarray(ds.variables["label"][:n])
split = np.asarray(ds.variables["split"][:n]).astype(str) if "split" in ds.variables else np.array([""] * n)
return features, labels, split
def _linear(x_train: np.ndarray, y_train: np.ndarray, x_test: np.ndarray) -> np.ndarray:
classes = np.unique(y_train)
design = np.concatenate([x_train, np.ones((len(x_train), 1), dtype=np.float32)], axis=1)
targets = (y_train[:, None] == classes[None, :]).astype(np.float64)
coef, *_ = np.linalg.lstsq(design, targets, rcond=None)
test = np.concatenate([x_test, np.ones((len(x_test), 1), dtype=np.float32)], axis=1)
return classes[(test @ coef).argmax(axis=1)]
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--dataset", choices=tuple(CUBES), default="cityrep")
parser.add_argument("--cube", default="")
parser.add_argument("--limit", type=int, default=0)
args = parser.parse_args()
with _open(args.dataset, args.cube) as ds:
print("time", [str(item) for item in ds.variables["time"][:4]], "...", ds.dimensions["time"].size)
print("bands", [str(item) for item in ds.variables["band"][:]])
print("note", ds.note)
features, labels, split = _mean_features(ds, args.limit)
finite = np.isfinite(features).all(axis=1)
print(f"samples={len(features)} finite={int(finite.sum())} labeled_splits={sorted(set(split))}")
train = finite & (split == "train")
test = finite & (split == "test")
if not train.any() or not test.any():
print("This cube has no train/test split column to score. CityRep and LUCAS do.")
return
pred = _linear(features[train], labels[train].astype(np.int64), features[test])
accuracy = float(np.mean(pred == labels[test].astype(np.int64)))
print(f"temporal-mean linear accuracy={accuracy:.4f} n_train={int(train.sum())} n_test={int(test.sum())}")
if __name__ == "__main__":
main()