Wide and Deep network-based spectral fusion of LIBS-NIRS for enhancing soil organic carbon estimation accuracy

Abstract

Rapid and accurate soil organic carbon (SOC) estimation is essential for assessing soil fertility and health. Although near-infrared spectroscopy (NIRS) has demonstrated tremendous application potential in SOC estimation, its inherent characteristics of weak absorption signals and overlapping spectral bands limit prediction accuracy and reliability. Here, we propose a spectral fusion technique that enhances estimation accuracy by integrating elemental information from laser-induced breakdown spectroscopy (LIBS) with molecular information from NIRS. We designed a Wide and Deep (W&D) dual-branch spectral fusion network for LIBS and NIRS information integration: the wide branch processes LIBS elemental features through a shallow linear layer, the deep branch extracts NIRS molecular features through a convolutional neural network, and both branches are collaboratively trained to achieve complementary feature fusion. Experimental results demonstrate that element-molecule feature fusion significantly improves SOC prediction accuracy. Compared to the optimal single-spectrum model, the W&D model reduces prediction errors by 0.168 g/kg (29.07%) and 0.187 g/kg (29.00%) on the test and external test sets, respectively. SHAP interpretability analysis reveals that effectively integrating elemental features (C, CN, K) from LIBS with functional group features (C-H, O-H) from NIRS is the key driver of accuracy improvement. This study provides a novel solution for accurate SOC estimation and highlights the tremendous potential of spectral fusion techniques in soil property sensing.