Spaces:
Paused
Paused
File size: 8,415 Bytes
a358495 | 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 | from __future__ import annotations
from dataclasses import dataclass
from enum import IntEnum
from typing import Any
import numpy as np
from scipy import ndimage as ndi
class ShadowType(IntEnum):
NONE = 0
BUILDING_SHADOW = 1
TERRAIN_SHADOW = 2
CLOUD_SHADOW = 3
VEGETATION_SHADOW = 4
UNKNOWN_SHADOW = 5
@dataclass(frozen=True)
class ShadowEvidence:
probability: np.ndarray
mask: np.ndarray
shadow_type: np.ndarray
features: dict[str, np.ndarray]
details: dict[str, Any]
def rgb_shadow_evidence(
r: np.ndarray,
g: np.ndarray,
b: np.ndarray,
valid: np.ndarray,
*,
water_probability: np.ndarray | None = None,
building_mask: np.ndarray | None = None,
vegetation_probability: np.ndarray | None = None,
) -> ShadowEvidence:
"""Estimate illumination shadow independently from surface identity.
Darkness contributes evidence, but cannot by itself create a shadow mask.
Local illumination contrast, chromatic continuation, texture/edges, and
object adjacency provide the additional evidence.
"""
r, g, b = (np.asarray(channel, dtype="float32") for channel in (r, g, b))
valid = np.asarray(valid, dtype=bool)
luminance = 0.299 * r + 0.587 * g + 0.114 * b
local_mean = ndi.uniform_filter(np.where(valid, luminance, 0.0), size=21)
valid_density = ndi.uniform_filter(valid.astype("float32"), size=21)
local_mean = local_mean / np.maximum(valid_density, 1e-4)
illumination_drop = np.clip((local_mean - luminance) / np.maximum(local_mean, 0.04), 0.0, 1.0)
maximum = np.maximum.reduce([r, g, b])
minimum = np.minimum.reduce([r, g, b])
saturation = (maximum - minimum) / np.maximum(maximum, 1e-5)
chromatic_neutrality = 1.0 - np.clip(saturation / 0.35, 0.0, 1.0)
darkness = np.clip((0.34 - luminance) / 0.34, 0.0, 1.0)
gradient = np.hypot(ndi.sobel(luminance, axis=0), ndi.sobel(luminance, axis=1)) / 4.0
edge_support = np.clip(ndi.uniform_filter((gradient > 0.035).astype("float32"), 9) / 0.18, 0.0, 1.0)
building_adj = np.zeros_like(luminance, dtype="float32")
if building_mask is not None:
objects = np.asarray(building_mask, dtype=bool)
if objects.shape == valid.shape:
building_adj = (ndi.binary_dilation(objects, iterations=8) & ~objects).astype("float32")
vegetation_adj = np.zeros_like(luminance, dtype="float32")
if vegetation_probability is not None:
vegetation = np.asarray(vegetation_probability, dtype="float32") > 0.55
if vegetation.shape == valid.shape:
vegetation_adj = (ndi.binary_dilation(vegetation, iterations=5) & ~vegetation).astype("float32")
bright_object = (luminance > 0.82) & valid
bright_adjacency = (ndi.binary_dilation(bright_object, iterations=18) & ~bright_object).astype("float32")
water_support = np.zeros_like(valid)
within_water_drop = np.zeros_like(luminance, dtype="float32")
if water_probability is not None:
water_support = np.asarray(water_probability, dtype="float32") >= 0.55
water_density = ndi.uniform_filter(water_support.astype("float32"), size=31)
water_local = ndi.uniform_filter(np.where(water_support, luminance, 0.0), size=31) / np.maximum(water_density, 1e-4)
within_water_drop = np.clip((water_local - luminance) / np.maximum(water_local, 0.02), 0.0, 1.0).astype("float32")
# At least two independent families are required: darkness/local contrast
# plus geometry/chromatic evidence. A dark pixel alone therefore stays low.
probability = (
0.25 * darkness
+ 0.32 * illumination_drop
+ 0.13 * chromatic_neutrality
+ 0.10 * edge_support
+ 0.14 * building_adj
+ 0.06 * vegetation_adj
+ 0.10 * bright_adjacency
)
contextual_support = (building_adj > 0) | (vegetation_adj > 0) | (bright_adjacency > 0) | ((edge_support > 0.55) & (chromatic_neutrality > 0.50))
corroborated = (illumination_drop > 0.20) & contextual_support
probability = np.where(corroborated, probability, probability * 0.35)
# Uniform dark water is not a cast shadow. Water may still be shadowed when
# it has a clear illumination drop relative to neighboring water pixels.
water_shadow_support = water_support & (within_water_drop > 0.28) & contextual_support
probability = np.where(water_support & ~water_shadow_support, probability * 0.20, probability)
probability = np.where(water_shadow_support, np.maximum(probability, 0.42 + 0.45 * within_water_drop), probability)
probability = np.clip(probability, 0.0, 1.0).astype("float32")
probability[~valid] = 0.0
shadow_type = np.full(valid.shape, ShadowType.NONE, dtype="uint8")
likely = probability >= 0.52
shadow_type[likely] = ShadowType.UNKNOWN_SHADOW
shadow_type[likely & (bright_adjacency > 0)] = ShadowType.CLOUD_SHADOW
shadow_type[likely & (vegetation_adj > 0)] = ShadowType.VEGETATION_SHADOW
shadow_type[likely & (building_adj > 0)] = ShadowType.BUILDING_SHADOW
# Water and shadow are allowed to overlap. Strong independent water evidence
# is never removed from the illumination product.
shaded_water_pixels = int((likely & (np.asarray(water_probability) >= 0.60)).sum()) if water_probability is not None else 0
return ShadowEvidence(
probability=probability,
mask=likely & valid,
shadow_type=shadow_type,
features={
"darkness": darkness.astype("float32"),
"illumination_drop": illumination_drop.astype("float32"),
"chromatic_neutrality": chromatic_neutrality.astype("float32"),
"edge_support": edge_support.astype("float32"),
"building_adjacency": building_adj,
"vegetation_adjacency": vegetation_adj,
"bright_object_adjacency": bright_adjacency,
"within_water_illumination_drop": within_water_drop,
},
details={"method":"rgb_local_illumination_geometry_v1", "shaded_water_pixels": shaded_water_pixels},
)
def multispectral_shadow_evidence(
brightness: np.ndarray,
valid: np.ndarray,
spectral_water_probability: np.ndarray,
*,
vegetation_probability: np.ndarray | None = None,
builtup_probability: np.ndarray | None = None,
) -> ShadowEvidence:
brightness = np.asarray(brightness, dtype="float32")
valid = np.asarray(valid, dtype=bool)
local = ndi.uniform_filter(np.where(valid, brightness, 0.0), 21)
density = ndi.uniform_filter(valid.astype("float32"), 21)
local = local / np.maximum(density, 1e-4)
illumination_drop = np.clip((local - brightness) / np.maximum(local, 0.01), 0.0, 1.0)
finite_values = brightness[valid]
dark_scale = float(np.percentile(finite_values, 35)) if finite_values.size else 0.0
darkness = np.clip((dark_scale - brightness) / max(dark_scale, 1e-4), 0.0, 1.0)
water = np.asarray(spectral_water_probability, dtype="float32")
nonwater_support = 1.0 - water
vegetation = np.asarray(vegetation_probability, dtype="float32") if vegetation_probability is not None else np.zeros_like(water)
builtup = np.asarray(builtup_probability, dtype="float32") if builtup_probability is not None else np.zeros_like(water)
context = np.maximum(vegetation, builtup)
probability = np.clip(0.40 * illumination_drop + 0.20 * darkness + 0.22 * nonwater_support + 0.18 * context, 0.0, 1.0)
corroborated = (illumination_drop > 0.22) & ((nonwater_support > 0.55) | (context > 0.45))
probability = np.where(corroborated, probability, probability * 0.30).astype("float32")
probability[~valid] = 0.0
mask = (probability >= 0.52) & valid
shadow_type = np.full(valid.shape, ShadowType.NONE, dtype="uint8")
shadow_type[mask] = ShadowType.UNKNOWN_SHADOW
shadow_type[mask & (vegetation >= builtup) & (vegetation > 0.45)] = ShadowType.VEGETATION_SHADOW
shadow_type[mask & (builtup > vegetation) & (builtup > 0.45)] = ShadowType.BUILDING_SHADOW
return ShadowEvidence(
probability=probability, mask=mask, shadow_type=shadow_type,
features={"darkness":darkness.astype("float32"), "illumination_drop":illumination_drop.astype("float32"), "nonwater_support":nonwater_support.astype("float32")},
details={"method":"multispectral_illumination_spectral_disagreement_v1", "shaded_water_pixels":int((mask & (water >= 0.60)).sum())},
)
|