TABLE III. Mean adjusted R? of GAMs and RF models, averaged over all 
37 frequency bands, for different combinations of parameters. 
Mean adjusted R? for: 
Zombination of parameters 
Speed + MMSI + distance + propulsion 0.74 
Speed + length + distance + propulsion 0.72 
Speed + length + distance 0.69 
Speed + length + MMSI 0.62 
Speed + length + propulsion 0.60 
Length + distance 2.57 
Speed + length 455 
Speed + distance 3.48 
GAMSs RF models 
0.78 
0.75 
0.69 
0.61 
0.53 
0.60 
0.60 
057 
frequency bands. In contrast to the bivariate analyses, the 
multivariate GAMs and RF models show substantially 
higher explanatory power. For both GAM and RF, the high- 
est adjusted R? values were obtained when including MMSI 
as a categorical variable, reaching 0.74 for GAM and 0.78 
for RF. This suggests that vessel-specific characteristics, 
which are not explicitly captured by the other parameters, 
have a substantial influence on RNLs, contributing to varı- 
ability within the CTV class. 
Although including MMSI yielded the best perfor- 
mance, it is not usable for predictive modelling beyond the 
observed vessels. Including ship length instead provides 
nearly comparable adjusted R? values (0.72 for GAM and 
3.75 for RF) when combined with speed, distance, and pro- 
pulsion type. This indicates that length may serve as a proxy 
for ship-specific factors such as operational state or engine 
configuration that are not explicitly considered in the GAMs 
ar RF models. The consistently strong contribution of length 
across frequencies further supports this interpretation. 
Among the individual predictors, speed and length were 
‘he most influential variables in the multivariate models, as 
reflected by their contribution to the adjusted R? values 
across frequency bands. This refers to their relative contri- 
bution within the GAM and RF frameworks, rather than 
‘heir isolated bivariate relationships with RNL. The combi- 
nation of speed, length, and distance still yielded a high 
explanatory power (0.72 for GAM and 0.75 for RF), rein- 
forcing the importance of these three parameters when con- 
sidered Jjointly. Notably, models that included only speed 
and length showed a noticeable drop in adjusted R? (0.57 for 
GAM, 0.60 for RP), highlighting that while these variables 
are important contributors, they alone do not fully explain 
‘he varlations in RNL. The overall contribution of the pro- 
pulsion type to the models was lower than expected. This 
suggests that the influence of propulsion type is secondary 
zompared to ship length and speed or that its effects are 
masked by other vessel-specific parameters. GAM and RF 
produced comparable adjusted R? across most combinations 
af parameters, with no consistent advantage for either 
method. In some frequency bands and for some parameter 
combinations, GAMs performed slightly better, while in 
others, RFs provided higher explanatory power. 
3412 J. Acoust. Soc. Am. 159 (4), April 2026 
https:/doi.org/10.1121/10.0043324 
JASA 
{80 
Median RNL by Propulsion Type (95% Cl) 
—- Azimuth Pods (N= 96 ) 
—CPP (N= 174) 
I— FPP (N= 259) 
‚= = bulker (200 m: 14 kn) 
170 
#160 
D 
+ 
» 15 
DD 
z 
147 
13r 
120 
101 
10° 
Frequency (Hz) 
104 
10' 
FIG. 8. Median RNLs for the different propulsion types with 95% confi- 
dence intervals. 
The strong performance of models including MMSI fur- 
ther indicates that individual vessel characteristics, not 
explicitly captured by speed or length alone, play a crucial 
role in shaping the observed variability. 
To further investigate the role of the propulsion type, 
we examined both the median RNLs and the proportion of 
passages exceeding the bulker reference spectrum per pro- 
pulsion type (Figs. 8 and 9). 
The characteristics of the RNL for the three propulsion 
types are distinct and statistically significantly different over 
all frequency bands above 30 Hz. Although the spectral dif- 
ferences are moderate in magnitude, they show that propul- 
sion configuration can shift the CTV spectra closer to or 
further away from the bulker reference spectrum at specific 
frequencies. Notably ships with CPP more frequently 
remained below the bulker reference curve above 1kHz 
than vessels with azimuth pods or FPPs. Interestingly, they 
also exhibited slightly higher average speeds (Fig. 9, right 
panel), further indicating that quieter performance does not 
necessarily result from lower operating speeds. However, 
the sample size (13 vessels) in this study is too low to draw 
general conclusions about this dependence. It is further note- 
worthy that CTVs with APs tend to exhibit the highest 
median RNLs between 1 and 20 kHz, while FPP-driven ves 
sels show the highest RNLs above 20kHz. The elevated lev- 
els above 20kHz are consistent with the emission 
characteristics of ultrasonic antifouling devices for two 
FPP-equipped CTVs (CTVs 9 and 11 in Fig. S1 in the sup- 
plementary material), although this interpretation requires 
further verification. 
IV. DISCUSSION 
This study analyzed a comprehensive dataset comprising 
529 CTV passages and calculated the respective RNLs of 13 
individual vessels. Although vessel speed, length, and propul- 
sion type do exert frequency-dependent effects on the RNLs, 
the dominant source of variability lies in vessel-specific 
Basan et al.