Ranking-guided coreset selection for NIRS-based nutrient prediction in liquid organic manures

Accurate nutrient estimation in liquid organic manures (LOM) is crucial to avoid over-fertilization and to support precision agriculture. While near-infrared spectroscopy (NIRS) provides a fast and non-destructive method for nutrient prediction, its practical adoption remains limited by the lack of efficient calibration strategies that balance model accuracy with resource constraints. This study proposes a relevance-guided coreset selection approach to improve model performance by selectively expanding the calibration set with the most informative samples. A dataset comprising 507 cattle and 439 pig manure samples was used to predict four key LOM properties: dry matter (DM), ammonium nitrogen (NH 4 -N), total nitrogen (TN), and phosphorus ( P 2 O 5 ). Relevance scores were computed via repeated model training using a Shapley-like attribution method. As a representative example for DM in cattle, training on the full calibration data (405 samples) yielded R 2 = 0.81 on the validation data (102 samples). Using only half the calibration data (211 samples) plus the top 10% most relevant samples (19 coreset samples) increased performance to R 2 = 0.83 , whereas adding the bottom 10% (19 anti-coreset samples) reduced it to R 2 = 0.78 . A feature-based learning-to-rank model, trained to generalize relevance estimation to unseen data, selected comparably informative samples and achieved R 2 = 0.82 under the same reduced-data configuration. This result shows that the ranking model effectively identifies informative samples and enables targeted calibration expansion with minimal laboratory effort, thus making NIRS scalable and resource-efficient in smart agriculture. • Results show that sample relevance varies in liquid organic manure for NIRS models. • Adding the coreset (most relevant samples) maintains or improves accuracy. • Adding the anti-coreset (least relevant samples) reduces predictive accuracy. • Feature-based learning-to-rank enables relevance-based selection for unseen samples. • Method enables efficient, scalable NIRS calibration for smart agriculture.