A Cross-Modal Multi-Task Learning Framework for Soil Property Estimation Based on Teacher-Student Distillation from VNIR, XRF, and LIBS under Small-Sample Conditions

Soil fertility information is essential for precision agriculture, yet conventional wet-chemistry assays are costly and difficult to scale for dense sampling. Proximal spectroscopies such as visible–near-infrared (VNIR), X-ray fluorescence (XRF), and laser-induced breakdown spectroscopy (LIBS) enable rapid “dry-chemistry” measurements, but existing multi-sensor fusion is often shallow, relies on single-task modeling, and is difficult to train reliably with limited labeled samples. To accelerate the broader adoption and application of multispectral techniques in precision agriculture, a cross-modal multi-task learning framework via teacher-student distillation that enables efficient integration of multispectral data and joint modeling of multi attributes was developed in this study. First, aiming for stable and reliable modeling under small-sample conditions, this framework transfers the structured knowledge and fine-grained supervisory signals derived from traditional chemometric teacher models to the student network through a multi-level teacher-student distillation strategy, thereby enhancing its knowledge assimilation and generalization capability. Furthermore, by incorporating modality-specific branches, Dirichlet prior-aware gating, multi-task learning, and Mixture-of-Experts, this framework achieves substantial improvements in both feature extraction and predictive performance. To assess the feasibility and effectiveness of the proposed framework, we systematically evaluated it on a multi-source spectral dataset composed of VNIR, XRF, and LIBS measurements, with only 71 training samples. Even under such limited data conditions, the student model achieves consistent gains across all nine soil properties, raising the macro-averaged R² to 0.81 ± 0.03, notably higher than the teacher model’s 0.77. Finally, the ablation study, insertion-deletion diagnostic, and interpretability analysis collectively verified the necessity and effectiveness of the framework’s components and strategies.