Spaces:
Sleeping
Sleeping
File size: 7,700 Bytes
41d98e2 | 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 | """AsteroidNET Two-Stage Classifier (RF → CNN)."""
from __future__ import annotations
import logging
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
import numpy as np
from asteroidnet.tracklet_linker.linker import Tracklet
logger = logging.getLogger(__name__)
@dataclass
class Classification:
tracklet: Tracklet
rf_score: float
cnn_score: float
is_asteroid: bool
priority: str # 'ROUTINE' | 'HIGH' | 'HAZARDOUS'
def classify_tracklets(
tracklets: list[Tracklet],
data_frames: list[np.ndarray],
config: Optional[dict] = None,
) -> list[Classification]:
"""
Two-stage classification: Random Forest (fast) → CNN (high precision).
RF stage uses kinematic features of the tracklet.
CNN stage uses 63×63 pixel cutouts from the detection positions.
"""
cfg = (config or {}).get("classifier", {})
rf_thresh = float(cfg.get("rf_threshold", 0.7))
cnn_thresh = float(cfg.get("cnn_threshold", 0.9))
rf_model = _load_rf(cfg)
cnn_model = _load_cnn(cfg)
results: list[Classification] = []
for tracklet in tracklets:
# ── RF stage ─────────────────────────────────────────────────────
features = _extract_features(tracklet)
rf_score = _rf_predict(rf_model, features)
if rf_score < rf_thresh:
continue
# ── Satellite filter ──────────────────────────────────────────────
if _is_satellite(tracklet):
logger.debug("Satellite filter rejected tracklet (vel=%.3f, pa=%.1f)",
tracklet.velocity_arcsec_s, tracklet.position_angle_deg)
continue
# ── CNN stage ─────────────────────────────────────────────────────
cutouts = _extract_cutouts(tracklet, data_frames)
cnn_score = _cnn_predict(cnn_model, cutouts)
is_asteroid = cnn_score >= cnn_thresh
if not is_asteroid:
continue
priority = _assign_priority(tracklet, cfg)
results.append(Classification(
tracklet=tracklet,
rf_score=rf_score,
cnn_score=cnn_score,
is_asteroid=True,
priority=priority,
))
results.sort(key=lambda c: c.cnn_score, reverse=True)
logger.info("Classification: %d/%d tracklets confirmed", len(results), len(tracklets))
return results
# ── Feature extraction ────────────────────────────────────────────────────────
def _extract_features(t: Tracklet) -> np.ndarray:
"""12-dimensional kinematic feature vector for RF classifier."""
snrs = [d["snr"] for d in t.detections]
mags = [d["mag"] for d in t.detections if d["mag"] < 90]
return np.array([
t.velocity_arcsec_s,
t.velocity_ra_arcsec_s,
t.velocity_dec_arcsec_s,
t.position_angle_deg / 360.0,
t.rms_residual_arcsec,
t.time_span_min,
len(t.detections),
float(np.mean(snrs)) if snrs else 0.0,
float(np.std(snrs)) if len(snrs) > 1 else 0.0,
float(np.mean(mags)) if mags else 25.0,
float(np.ptp(mags)) if len(mags) > 1 else 0.0,
len(set(t.frame_ids)),
], dtype=np.float32)
def _extract_cutouts(
t: Tracklet,
data_frames: list[np.ndarray],
size: int = 63,
) -> Optional[np.ndarray]:
"""Extract stacked cutouts from detection positions."""
if not data_frames:
return None
half = size // 2
cutouts = []
for det in t.detections:
fid = det.get("frame_id", 0)
if fid >= len(data_frames):
continue
data = data_frames[fid]
x, y = int(round(det.get("x", 0))), int(round(det.get("y", 0)))
h, w = data.shape
if x - half < 0 or y - half < 0 or x + half >= w or y + half >= h:
continue
cutout = data[y-half:y+half+1, x-half:x+half+1].copy()
if cutout.shape == (size, size):
finite = cutout[np.isfinite(cutout)]
if len(finite) > 0:
med = np.median(finite); mad = max(np.median(np.abs(finite-med)), 1e-10)
cutout = np.clip((cutout - med) / (3*mad), -3, 3)
cutouts.append(np.nan_to_num(cutout.astype(np.float32)))
return np.stack(cutouts) if cutouts else None
# ── Model loading ─────────────────────────────────────────────────────────────
def _load_rf(cfg: dict):
"""Load RF model if available, else return None (heuristic fallback)."""
path = cfg.get("rf_model_path", "models/rf_classifier.pkl")
if Path(path).exists():
try:
import joblib
return joblib.load(path)
except Exception as exc:
logger.warning("Could not load RF model %s: %s", path, exc)
return None
def _load_cnn(cfg: dict):
"""Load CNN model if available, else return None (heuristic fallback)."""
path = cfg.get("cnn_model_path", "models/cnn_classifier.pth")
if Path(path).exists():
try:
import torch
model = torch.load(path, map_location="cpu")
model.eval()
return model
except Exception as exc:
logger.warning("Could not load CNN model %s: %s", path, exc)
return None
def _rf_predict(model, features: np.ndarray) -> float:
"""RF prediction — heuristic if model not trained yet."""
if model is not None:
try:
p = model.predict_proba(features.reshape(1, -1))[0, 1]
return float(p)
except Exception:
pass
# Heuristic: based on velocity, SNR, residual
vel = features[0]
snr = features[7]
rms = features[4]
score = 0.3
if 0.01 <= vel <= 5.0: score += 0.3
if snr >= 5.0: score += 0.2
if rms <= 0.8: score += 0.2
return min(score, 0.99)
def _cnn_predict(model, cutouts: Optional[np.ndarray]) -> float:
"""CNN prediction — heuristic if model not trained yet."""
if model is not None and cutouts is not None:
try:
import torch
x = torch.from_numpy(cutouts).unsqueeze(0).float()
with torch.no_grad():
out = model(x)
return float(torch.sigmoid(out).mean().item())
except Exception:
pass
# Heuristic: check if any cutout has a point source at center
if cutouts is not None and len(cutouts) > 0:
peaks = [float(np.max(c[28:35, 28:35])) if c.shape == (63, 63)
else float(np.nanmax(c)) for c in cutouts]
return min(0.95, max(0.0, float(np.mean(peaks)) / 3.0 + 0.5))
return 0.5
def _is_satellite(t: Tracklet) -> bool:
"""Simple satellite/aircraft filter."""
if t.velocity_arcsec_s > 8.0:
return True
if t.rms_residual_arcsec < 0.01 and t.velocity_arcsec_s > 5.0:
return True
return False
def _assign_priority(t: Tracklet, cfg: dict) -> str:
high_v = float(cfg.get("high_velocity_threshold", 1.0))
haz_v = float(cfg.get("hazardous_velocity_threshold", 3.0))
if t.velocity_arcsec_s >= haz_v:
return "HAZARDOUS"
if t.velocity_arcsec_s >= high_v:
return "HIGH"
return "ROUTINE"
|