Predicting soil total nitrogen and organic matter with hybrid models on small laser-induced breakdown spectroscopy datasets

The total nitrogen (TN) and organic matter (OM) content of the soil is crucial to improve crop yields and reduce environmental impacts in precision agriculture. Recently, Laser-Induced Breakdown Spectroscopy (LIBS) has become a popular method for predicting soil nutrients because of its rapid, nondestructive, and multielement analytical capabilities. However, the high-dimensionality and complex peak features of LIBS spectra, combined with often limited sample sizes, pose challenges for previous deep learning methods, such as over fitting, feature redundancy, and poor generalization. To address these challenges, we propose a hybrid model tailored for small LIBS datasets to predict soil TN and OM content. This model integrates Partial Least Squares (PLS) for dimensionality reduction and key feature extraction, Convolutional Neural Networks (CNN) for capturing local spectral patterns, and Self-Attention mechanisms for modeling global dependencies. By combining these components with weighted integration, the hybrid model significantly improves prediction accuracy and robustness. Experiments show that the hybrid model outperforms other machine learning and standalone deep learning methods in small LIBS datasets, achieving superior performance with RMSE of 0.39 g/kg and R 2 of 0.75 for the prediction of TN (compared to the second-best method with 0.42 g/kg and 0.71), and RMSE of 8.26 g/kg and R 2 of 0.77 for the prediction of OM (compared to the second-best method with 8.77 g/kg and 0.74). This study presents an effective solution for analyzing high-dimensional spectral data with small datasets, supporting soil health management and sustainable precision agriculture. • Predicting soil TN and OM content using LIBS spectroscopy with hybrid model approach. • Achieves effective prediction on small LIBS datasets with only 483 soil spectral data. • Innovative framework integrating dimensionality reduction with local–global feature extraction. • Outperforms other methods with R 2 = 0.754 for TN and R 2 = 0.766 for OM.