Datasets:
File size: 9,513 Bytes
b0d8bdd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 | # Copyright 2026 Nikhil Raghavendra
# SPDX-License-Identifier: Apache-2.0
"""Tests of preprocessing, metadata, splits and the filter rules."""
from __future__ import annotations
from typing import cast
import numpy as np
import pandas as pd
import pytest
from sen_bridge_build import filters, meta, preprocess, splits
from sen_bridge_build.sample import Record, reject_reason
_SIZE = 264
_SHARE = {"A": 0.6, "B": 0.3, "C": 0.1} # the regional mix that _pool draws from
def test_sar_roundtrip_and_shapes() -> None:
rng = np.random.default_rng(0)
db = rng.uniform(-25, -6, (2, _SIZE, _SIZE)).astype(np.float32)
out = preprocess.sar_to_db128(db)
assert out.shape == (2, 128, 128)
u8 = preprocess.sar_to_uint8(out)
assert u8.shape == (3, 128, 128)
assert u8.dtype == np.uint8
assert (
np.abs(preprocess.sar_from_uint8(u8)[0] - out[0]).max()
<= 25 / 255 / 2 + 1e-3
)
def test_multilook_is_linear_power() -> None:
db = np.full((2, _SIZE, _SIZE), -10.0, np.float32)
db[:, 4, 4] = 0.0 # the crop starts at 4, so this is block (0, 0)
expected = 10 * np.log10((1 + 3 * 0.1) / 4)
assert preprocess.sar_to_db128(db)[0, 0, 0] == pytest.approx(expected, abs=1e-4)
def test_optical_roundtrip() -> None:
rho = np.linspace(0, 0.3, 3 * 128 * 128, dtype=np.float32).reshape(3, 128, 128)
back = preprocess.opt_from_uint8(preprocess.opt_to_uint8(rho))
assert np.abs(back - rho).max() < 0.002
ends = preprocess.to_model_range(np.array([0, 255], np.uint8))
assert ends.tolist() == [-1.0, 1.0]
def test_meta_vector() -> None:
vec = meta.meta_vector(100, 45.0, 10.0, 24.0, orbit_desc=True)
assert vec.shape == (7,)
assert vec.dtype == np.float32
assert vec[5] == pytest.approx(0.5)
assert vec[6] == 1
assert np.linalg.norm(vec[2:5]) == pytest.approx(1)
def test_metadata_filter() -> None:
ok = filters.metadata_ok([10, 60, 10, 10], [50, 0, 50, -50], [7, 7, 1, 1])
assert ok.tolist() == [True, False, False, True]
def test_latitude_band_uses_absolute_latitude_and_the_configured_edges() -> None:
lat = pd.Series([0.0, 19.9, 20.0, -29.9, 30.0, -45.0, 50.0, 80.0])
assert splits.latitude_band(lat).tolist() == [0, 0, 1, 1, 2, 3, 4, 4]
def test_region_of_and_other() -> None:
assert splits.region_of(50, 10) == "Europe"
assert splits.region_of(-80, 0) == "Other"
def test_splits_and_downsample() -> None:
rng = np.random.default_rng(1)
df = pd.DataFrame({
"lat": rng.uniform(-60, 60, 2000),
"lon": rng.uniform(-180, 180, 2000),
})
df["cluster_id"] = splits.assign_clusters(df, df.sample(50, random_state=0))
df["stratum"] = np.where(df["lat"] > 0, "N", "S")
df["split"] = splits.assign_splits(df, test_rows=300, val_rows=100)
assert (df.groupby("cluster_id")["split"].nunique() <= 2).all()
df["region"] = rng.choice(list("ABC"), len(df), p=[0.6, 0.3, 0.1])
assert len(splits.stratified_downsample(df, 500)) == 500
def test_assign_splits_matches_row_quotas_and_the_stratum_mix() -> None:
rng = np.random.default_rng(2)
n = 20_000
df = pd.DataFrame({"cluster_id": rng.integers(0, 400, n)})
df["stratum"] = np.where(df["cluster_id"] % 5 == 0, "small", "big")
split = splits.assign_splits(df, test_rows=3000, val_rows=1000)
counts = split.value_counts()
# quotas are met to within about one cluster (50 rows) per stratum
assert abs(counts["test"] - 3000) <= 150
assert abs(counts["val"] - 1000) <= 150
# a whole cluster is never split across two splits
first = df.assign(s=split).groupby("cluster_id")["s"].transform("first")
assert (first == split).all()
# each stratum holds its proportional share of the test pool
small = (split[df["stratum"] == "small"] == "test").sum() / counts["test"]
assert small == pytest.approx((df["stratum"] == "small").mean(), abs=0.03)
def test_drop_within_removes_only_rows_near_the_reference() -> None:
step = 3 / 111.195 # degrees of longitude per 3 km at the equator
reference = pd.DataFrame({"lat": [0.0], "lon": [0.0]})
rows = pd.DataFrame({"lat": [0.0, 0.0, 0.0], "lon": [step, 2 * step, 1.0]})
kept = splits.drop_within(rows, reference, km=5.0)
# 3 km is dropped; 6 km (beyond the rule) and 111 km stay.
assert kept["lon"].round(4).tolist() == [round(2 * step, 4), 1.0]
def test_drop_within_with_no_reference_returns_everything() -> None:
rows = pd.DataFrame({"lat": [0.0], "lon": [0.0]})
assert len(splits.drop_within(rows, rows.iloc[0:0])) == 1
def _pool(n: int, seed: int = 0) -> pd.DataFrame:
"""Build a pool with regions and water / bare-ground shares above the caps.
Args:
n: Number of rows.
seed: Seed of the draw.
Returns:
Rows with the columns the selection uses.
"""
rng = np.random.default_rng(seed)
return pd.DataFrame({
"region": rng.choice(list("ABC"), n, p=[0.6, 0.3, 0.1]),
"water_frac_opt": np.where(rng.random(n) < 0.30, 0.95, 0.0),
"lulc_bare": np.where(rng.random(n) < 0.20, 0.9, 0.0),
})
def test_stratified_downsample_matches_the_target_mix() -> None:
pool = _pool(10_000)
target = {"A": 0.5, "B": 0.4, "C": 0.1} # differs from the pool's 60/30/10
chosen = splits.stratified_downsample(pool, 1000, target_share=target, seed=1)
shares = chosen["region"].value_counts(normalize=True)
for region, want in target.items():
assert shares[region] == pytest.approx(want, abs=0.002)
assert len(chosen) == 1000
def test_stratified_downsample_under_represents_a_short_region() -> None:
pool = _pool(10_000)
pool = pool[(pool["region"] != "C") | (pool.index % 50 == 0)] # C nearly absent
chosen = splits.stratified_downsample(pool, 1000, target_share=_SHARE)
assert len(chosen) == 1000
assert (chosen["region"] == "C").sum() <= (pool["region"] == "C").sum()
def test_select_capped_is_exact_in_size_and_caps() -> None:
pool = _pool(5000)
chosen = splits.select_capped(pool, 1000, _SHARE, seed=3)
assert len(chosen) == 1000
assert chosen.index.is_unique
assert splits.is_water(chosen).sum() <= 0.10 * 1000
assert splits.is_bare(chosen).sum() <= 0.08 * 1000
def test_select_capped_on_a_small_pool_obeys_the_caps_without_padding() -> None:
pool = _pool(50)
chosen = splits.select_capped(pool, 1000, _SHARE)
assert len(chosen) <= 50 # never more rows than the pool holds
assert splits.is_water(chosen).sum() <= 0.10 * len(chosen)
def test_cap_share_respects_cap_per_split() -> None:
rng = np.random.default_rng(0)
df = pd.DataFrame({
"split": rng.choice(["train", "test"], 1000),
"w": rng.random(1000),
})
out = splits.cap_share(df, df["w"] > 0.7, 0.10)
for _, group in out.groupby("split"):
assert (group["w"] > 0.7).mean() <= 0.10 + 1e-9
assert (out["w"] <= 0.7).sum() == (df["w"] <= 0.7).sum() # others untouched
def _clean_record() -> Record:
"""Build a record that passes every filter.
Returns:
A complete record with unremarkable values.
"""
return {
"id": "x_0",
"opt_mean": 0.5,
"opt_std": 0.1,
"opt_white_frac": 0.0,
"flat_bright": False,
"water_frac_sar": 0.0,
"vh_noise_frac": 0.0,
"slope_std": 2.0,
"elev_mean": 100.0,
"cloud_rule_frac": 0.0,
"snow_score": 0.0,
"haze_score": 0.8,
"sensei_cloud_frac": 0.0,
"sensei_nodata_frac": 0.0,
"water_frac_opt": 0.0,
"snow_frac_opt": 0.0,
"rfi_score": 0.0,
"sat_frac": 0.0,
"border_frac": 0.0,
"lulc_water": 0.0,
"lulc_trees": 0.2,
"lulc_flooded": 0.0,
"lulc_crops": 0.6,
"lulc_built": 0.2,
"lulc_bare": 0.0,
"lulc_snow": 0.0,
"lulc_rangeland": 0.0,
}
def test_clean_record_passes() -> None:
assert reject_reason(_clean_record(), invalid=False) is None
@pytest.mark.parametrize(
("change", "reason"),
[
({"sat_frac": 0.5}, "saturated"),
({"sensei_cloud_frac": 0.2}, "cloud_shadow"),
({"cloud_rule_frac": 0.9}, "cloud_opaque"),
({"opt_white_frac": 0.9, "opt_std": 0.01}, "overexposed"),
({"opt_mean": 0.03}, "too_dark"),
(
{"opt_mean": 0.95, "opt_std": 0.01, "lulc_bare": 0.9},
"featureless_arid",
),
({"water_frac_sar": 0.9, "lulc_water": 0.5}, "water_mismatch"),
({"slope_std": 12.0}, "steep"),
({"elev_mean": 3000.0}, "elevation"),
],
)
def test_each_filter_rejects_with_its_own_reason(
change: dict[str, float], reason: str
) -> None:
# The test builds a record from a dict literal; cast it to the TypedDict.
record = cast("Record", {**_clean_record(), **change})
assert reject_reason(record, invalid=False) == reason
def test_invalid_pixels_win_over_every_other_reason() -> None:
# The test builds a record from a dict literal; cast it to the TypedDict.
record = cast(
"Record", {**_clean_record(), "sat_frac": 0.5, "elev_mean": 3000.0}
)
assert reject_reason(record, invalid=True) == "invalid_pixels"
def test_arid_flat_bright_is_not_rejected_off_arid_land() -> None:
# The test builds a record from a dict literal; cast it to the TypedDict.
record = cast("Record", {**_clean_record(), "opt_mean": 0.95, "opt_std": 0.01})
assert reject_reason(record, invalid=False) is None
|