Computational Foundations of Nutritional Imaging in Agri‐Food Systems

Malnutrition affects over 820 million people globally, while improper fertilizer application annually degrades approximately 1.6 billion tonnes of soil. Noninvasive imaging techniques, such as hyperspectral, multispectral, and CT, appear promising: the number of scientific articles has increased from 45 published in 2015 to 158 published in 2023 (growth rate of 18% per year), while the imaging for precision agriculture market is expected to witness a surge from $1.5 billion in 2022 to $4.2 billion in 2030 (Compound Annual Growth Rate (CAGR) of 16.5%). However, almost 90% of these studies ignore the incorporation of multiscale imaging, and fewer than five repositories satisfy the FAIR criteria. This research aims at establishing a common computational basis for nutrition-related imaging in agri-food systems by (i) systematically comparing the state-of-the-art methods, (ii) quantifying their utilities and resource footprints, and (iii) finding critical missing links in the standardization of data, uncertainty quantification, and ecological cost. A meta-analysis of 35 highly impactful papers (2020–2025) retrieved from Scopus, Web of Science, IEEE, ACM Digital Library, PubMed, and the UGC-CARE List I databases was undertaken. Inclusion criteria involved 30 or more citations, standardized reporting of either R 2 or RMSE for nutrient estimation, and open access to either datasets or code. Keywords are “hyperspectral nutrient estimation,” “multiscale imaging,” and “physics-informed deep learning.” For comparative analysis, quantitative indicators such as mean R 2 , RMSE, sample size, training time, and energy use were mined from the literature and tabulated or diagrammed; in the form of a flowchart, they present our work plan, while bar charts display publication trends. The results show pure CNNs at mean R 2 = 0.79 (RMSE = 5.9 g) for wheat N; physics-informed networks improve R 2 to 0.81 (5.4 g) (12% bias reduction); and our spectral-spatial-physics model further improves R 2 to 0.85 (5.0 g; 18% lower RMSE) across five crops. Federated learning at 12 sites achieves 92% global accuracy with ε ≤ 1 privacy leakage. Only 4 of the 22 repositories surveyed completely adhere to FAIR, and training typically takes 12 ± 4 h and consumes 2.4 kWh per model. Unlike any other, our work puts forward the very first all-encompassing comparative framework that (i) covers micro-to-field-scale imaging; (ii) folds in uncertainty and ecological-cost criteria; and (iii) releases an open dataset with over 500 paired multispectral/CT samples with chemical assays. This foundation initiates addressing key neglected gaps in standardization, interpretability, and sustainability, and charts the roadmap for future research in agri-food imaging.