Explainability in vis-NIR and XRF-based modeling for soil fertility analysis: A comprehensive review

Sustaining soil health is paramount for global food security, carbon sequestration, and ecosystem resilience, while traditional wet-chemistry analyses remain time-consuming and resource-intensive. Visible-Near InfraRed (vis-NIR) and X-ray Fluorescence (XRF) have emerged as rapid, cost-effective alternatives for monitoring key soil health parameters, especially fertility attributes. In parallel, supervised Machine Learning (ML) has revolutionized soil modeling by transforming complex spectral signatures into high-level soil features. Despite their outstanding performance, such AI-driven approaches often operate as ”black boxes”. In this context, this review situates the current landscape of soil fertility modeling via integrating ML with vis-NIR and XRF, particularly focusing on adopting explainability methods to make models more explainable and transparent. By systematically surveying studies from 2019 to June 2025 (315 articles), we found that carbon, organic matter, clay, pH, nitrogen, and sand are the most prevalent predicted parameters, reflecting their pivotal role in soil health assessment. Vis-NIR dominates as the sensing technology (especially in multi-country investigations) while XRF, despite its elemental sensitivity, requires testing in large-scale contexts. In modeling, there has been a transition from traditional linear regressors ( e.g. , MLR and PLS) to more advanced, non-linear learners such as neural networks, which impact models’ explainability. Explainability techniques have primarily been implemented at the global level (notably through informative coefficients and dimensionality reduction), whereas local explainers (sample-specific assignments) are scarcely employed. To offer a practical guide for selecting methods aligned with research goals and data characteristics, we leveraged eXplainable AI (XAI) core concepts to propose a taxonomy for categorizing explainability tools according to agnosticism, scope, timing, and transparency. In conclusion, achieving a balance among predictive performance, complementarity between data sources, and effective communication of model behavior—accomplished through close collaboration with domain experts and robust XAI practices—is essential for developing more practical and reliable soil fertility modeling strategies. • A review of ML and adopted explainability tools in soil fertility monitoring was conducted. • A shift toward employing advanced nonlinear models, especially deep learning, was noticed. • A balance between performance and explainability needs XAI solutions. • Enhancing model communication through explainability can make soil fertility modeling more trustworthy. • Explainability tools was arranged by nature, scope, workflow stage, and transparency by a new taxonomy proposal.