Stacked Models and Conformal Prediction to Classify Exotic Hardwoods by Laser Induced Breakdown Spectroscopy

Wednesday, October 8, 2025

2:50 PM - 3:10 PM EDT

Location: Meeting Room #7

Presenting Author(s)

KB

Karl S. Booksh, PhD

Professor
University of Delaware
Newark, Delaware, United States

Non-Presenting Author(s)

HC

Helder V. Carneiro

PhD Candidate
University of Delaware
Newark, Delaware, United States
EM

Erin McClure-Price

US Forest Service International Programs Wood Identification and Screening Center
Ashland, Oregon, United States
KC

Kierra Cano

US Forest Service International Programs Wood Identification and Screening Center
Ashland, Oregon, United States
CC

Caelin Celani, PhD (he/him/his)

Postdoctoral Fellow
University of Delaware
Claymont, Delaware, United States
JJ

James Jordan

United States Geological Survey
Reston, Virginia, United States
KE

Kent Elliott

US Forest Service International Programs Wood Identification and Screening Center
Ashland, Oregon, United States

In fifteen words or less, explain the significance of this contribution (Novel Aspect).: Stacked models, conformal prediction, and applications with more than 15 classes are each seldom seen.

Abstract Text: We demonstrate the ability differentiate among 18 classes of exotic hardwood timber with stacked models and placement of statistically valid confidence intervals on the results with conformal prediction. Laser Induced Breakdown Spectroscopy (LIBS) spectra were collected from 700 timber samples spanning 18 species. Using flat and one-vs-all models, the best prediction performance was a 0.56 kappa. Stacked models aggregate poorly performing models and create a meta-learner, that is a classification model built on classification model outputs in this case. With a stacked model, the prediction performance was increased to 0.87 kappa. Conformal prediction was then employed to place statistically meaning full confidence intervals on each prediction. A conformal prediction analyses returns a set of classes that are guaranteed to contain the true class at a certain confidence level. Here we show the prediction set size following stacked models as a metric of overall model performance. Conformal predictions will also indicate which samples, and classes, the models struggles the most to predict, regardless of local model accuracy.