simulated listening station · 1/3-octave × time (newest at top)
connecting…
—
25632501k4k16k Hz
quietloud (dB)
Live compliance events
0
a card appears when the ML engine detects a vessel pass
Calibration & the significance gate
Calibration · Assumption A1
—
Ambient floor · Background P95
—
Significance gate · 1.645σ
—
Model scorecards
Task A — per-class F1 & average precision (test)
Task A — auxiliary heads (grouped CV)
Task B — forecast error vs horizon
Task B — skill vs persistence
Task C — received-level regression
Task C — range-class partial dependence
Limitations
Site & calibration — 1 hydrophone, 1 lake, uncalibrated. Absolute dB is illustrative.
Model error — ≈3–8 dB; flagged only above 1.645σ by default. σ is one global per-band figure, not per vessel class — UUV and KaiYuan's own average band errors (~9–11 dB) both run above it, so their "confidence" labels are less trustworthy than the same label for SpeedBoat or QianDao under the default (legacy Gaussian) uncertainty source. The Conformal (Mondrian) option fixes this measurably — held-out coverage check: KaiYuan 91.1%→96.2%, WorkShip 88.4%→93.1% (target 95%) — but is opt-in, not the default, since it needs a per-vessel-class sample size the rarest classes don't have (falls back to the global margin below n=30).
Received, not source, level — not ISO 17208.
Mock ordinance — project-defined limit, not a legal instrument.
⚠ Detectability screening, not an impact assessment — detection ≠ harm. Sources per species below.
Loading regulatory context…
Vessel footprint vs. species hearing thresholds
dB re 1 µPa, same calibration as the compliance engine
Species focus — how shipping activity relates to this species
Computed live from every real recorded clip…
–
Margin above hearing threshold, by vessel class
% of real passes detectable, by vessel class
Spread of real per-clip levels, by vessel class
Species evidence table
artifacts/fish_species.parquet · – rows
Drag a box on the spectrogram to select a time x frequency region — it's logged to the
table below with its own extracted stats (mean/max dB, duration, mel-bin range). This is a
manual inspection tool, not an automated event detector or a scorer against ground truth —
this dataset has no manually-annotated event start/end times to score against, so that part
of the Raven Pro precedent genuinely doesn't apply here; what transfers is treating the
spectrogram as something you select on, not just a picture.
Vessel passes
Mel spectrogram · select a clip
drag to select · box/lasso tool in the toolbar
Selection log (0)
#
clip_id
t_lo (s)
t_hi (s)
mel bin lo
mel bin hi
mean dB
max dB
No selections yet — drag a box on the spectrogram above.