Advanced chemometrics toward robust spectral analysis for fruit quality evaluation

Background: The application of visible/near-infrared (Vis/NIR) spectroscopy for fruit quality evaluation has garnered significant attention over the past few decades. Various chemometric techniques have been developed to predict fruit quality from spectral data. However, the broad applicability of existing chemometric models is limited by unpredictable data variability caused by various biological factors, instrumental settings, and measurement conditions. Deep learning has emerged as a leading methodology, offering substantial improvements in the accuracy and robustness of fruit quality assessments. Scope and approach: This review examines the challenges of model robustness in fruit spectral analysis, tracing the advancement from conventional chemometrics to deep learning approaches. Developments in chemometric methods to enhance model reliability are explored, encompassing dataset-level, variable-level, and model parameter-level strategies, while outlining their applicability and limitations. Recent advances in deep learningbased techniques, e.g., transfer learning, multi-task learning, multi-modal data fusion, and knowledge-guided model design, are further highlighted, providing prospective pathways for achieving superior model robustness. Key findings and conclusions: Current chemometric methods have enhanced model accuracy and proven effective in fruit spectral analysis. While the results are improved for certain research objectives, many analyses remain dependent on specific dataset characteristics and manual feature engineering, such as preprocessing, which limits their generalizability. Deep learning techniques with advanced feature extraction capabilities have shown promise in reducing the need for manually engineered features and expanding model robustness. However, further investigation into the applicability and limitations of these models is crucial for their successful integration into chemometric analysis.