Many strong peaks in Regression coefficients reflected the complexity of the calibration model

"To clarify the calibration structure, the PLS regression
coefficient plots, which are the regression coefficients
plotted against wavelength, are shown in Fig. 4. Many
strong peaks could be observed, indicating the complexity of the calibration model."

Suthiluk, P., Saranwong, S., Kawano, S., Numthuam, S., & Satake, T. (2008). Possibility of using near infrared spectroscopy for evaluation of bacterial contamination in shredded cabbage. International Journal of Food Science and Technology, 43(1), 160-165.

Good argument on the difficulty to indicate which are the molecules responsible for spectral information

"By analysing the regression coefficient of the four PLS models
(Fig. 6) it can be seen that the region between 5000 and
4000 cm 1 is the most important. This spectral region is mainly
due to C–H and C–C combination bands and is the one that contains
most information regarding the chemical composition of the defective
and non-defective beans, although, as already stated, coffee is a
combination of numerous organic molecules and therefore it is difficult
to indicate which are the molecules responsible for the information
contained in the spectral region referred to previously."

Ana Paula Craig, Adriana S. Franca, Leandro S. Oliveira, Evaluation of the potential of FTIR and chemometrics for separation between defective and non-defective coffees, Food Chemistry, Volume 132, Issue 3, 1 June 2012, Pages 1368-1374, ISSN 0308-8146,

Reference for the PLS regression

"The PLS regression model is commonly used in chemometrics
with the objective of establishing regression models (for physical
or chemical properties) from spectral data (Geladi & Kowalski,
1986; Miller, 2000)."

Geladi, P., & Kowalski, B. R. (1986). Partial least-squares regression – a tutorial.
Analytica Chimica Acta, 185, 1–17.

Reference for leave-one-out cross validation for the optimal number of latent variables for PLSR model

"The coffee bean qualities mass fraction values were modelled
from the NIR spectra with partial least squares (PLS) regression.
The optimal number of latent variables for each model was estimated
with the leave-one-out cross-validation method (Tormod,
Tomas, Fearn, & Tony, 2002)."

Tormod, N., Tomas, I., Fearn, T., & Tony, D. (2002). A User-Friendly Guide to
Multivariate Calibration and Classification. Chichester: NIR Publications.