Training and test dataset for: Choosing an operational inference pipeline for internal solitary wave detection in Sentinel-1 SAR imagery: EVA02-Large+XGBoost versus SAR_CNN v2 (Lux.jl)

This record contains the training and evaluation dataset used in the manuscript "Choosing an operational inference pipeline for internal solitary wave detection in Sentinel-1 SAR imagery: EVA02-Large+XGBoost versus SAR_CNN v2 (Lux.jl)" (Pinelo et al., submitted to Geoscientific Model Development, 2026; preprint egusphere-2026-1798). The archive comprises Sentinel-1 WV mode vignettes drawn from the AIR Centre's Internal Waves Service operational annotation set, split into a training partition ([N_TRAIN] images) and a held-out test partition (5,860 images, evenly balanced between 2,930 ISW-positive and 2,930 ISW-negative). Each partition is accompanied by a CSV table mapping images to their expert-assigned class labels. The test partition is the same benchmark set used to compare the two inference pipelines described in the paper — a Python pipeline based on EVA02-Large (305M parameters) with an XGBoost classification head, and a Julia SAR_CNN v2 model implemented in Lux.jl (283,329 parameters). The underlying Sentinel-1 imagery is derived from ESA Copernicus data. Class labels were produced by domain experts at the AIR Centre. The corresponding source code, trained model weights, benchmark scripts, and output metrics are archived separately at https://doi.org/10.5281/zenodo.19322369.