Reconciling discrepancies in Oklahoma field study observation counts

Determine the cause of the discrepancies between the observation counts for FM1, FM10, FM100, and FM1000 reported in Carlson (2007) and the counts present in the Oklahoma field study dataset used in this thesis (1,237 vs 1,233 for FM1; 1,237 vs 1,232 for FM10; 874 vs 871 for FM100; 877 vs 874 for FM1000), and reconcile the data if possible to ensure consistent and reproducible analyses.

Background

The thesis evaluates transfer learning methods for predicting fuel moisture content using data from the 1996–1997 Oklahoma field study that was used to calibrate the Nelson model. A copy of the data was obtained via personal correspondence, but the original hardware copy was reportedly destroyed.

The observation counts in the received dataset differ slightly from those reported in the original publication, and the authors explicitly state they were unable to reconcile these discrepancies. As the dataset is a benchmark for evaluation and calibration, clarifying these differences is important for reproducibility and cross-study comparisons.

References

The paper reported 1,237 observations of FM1, 1,237 observations of FM10, 874 observations of FM100, and 877 observations of FM1000. In the data received from Dr. Van der Kamp, there were 1,233 observations of FM1, 1,232 observations of FM10, 871 observations of FM100, and 874 observations of FM1000. We are unable to reconcile the slight discrepancies in the observation counts, and no other backup of the original data exists to our knowledge.

— Time-Warping Recurrent Neural Networks for Transfer Learning  (2604.02474 - Hirschi, 2 Apr 2026) in Chapter 3, Section: Overview of Empirical Analysis with Fuel Moisture Classes