Coffee is a globally important product mainly cultivated in regions such as South America, Africa, and Southeast Asia. The total production of coffee exceeds 7 million tons annually [1]. Coffee is a globally traded commodity with a complex supply chain that includes harvesting, post-harvest processing, storage, transport, roasting, and grinding. Prior to roasting, green coffee beans typically undergo cleaning and sorting steps to remove foreign materials and non-compliant fractions that can compromise downstream processing and final product quality. Defects and contaminants such as stones, husks, broken beans, and visually defective kernels may negatively affect roasting performance, cause mechanical damage to equipment, and deteriorate sensory characteristics of the beverage [2].

Before roasting, the selection of coffee beans is a fundamental step to ensure the final product is of high quality. This selection process involves both mechanical and manual methods.

From a sustainability perspective, optimizing the sorting process helps reduce food waste and resource loss in coffee value chains. Improved sorting also minimizes the need for redundant processing stages and contributes to a more efficient, low-impact production system that aligns with global goals for sustainable agriculture and industrial processes.

Mechanically, systems are used to classify beans based on size, shape, and to remove foreign materials like stones, metal fragments, and other contaminants [3]. These systems help streamline the process and improve consistency. Manual selection, on the other hand, involves human operators inspecting the beans to eliminate defective or damaged ones, ensuring only the best beans proceed to roasting.

Industrial sorting is traditionally performed using mechanical separation coupled with optical systems. For example, artificial vision systems analyze surface features like colour, shape, size, and visible defects to identify and remove defective beans. These systems are quite effective in external defect detection, but have limitations; they cannot detect internal defects or contaminants hidden inside the beans [4]. Their effectiveness improves when the final product is more homogeneous, but they still serve as partial solutions. RGB-based optical sorters classify products mainly by surface colour, size, and shape. While effective for many visible defects, RGB approaches may be less reliable when contaminants share a similar visual appearance with sound beans or when discrimination would benefit from information linked to chemical composition rather than colour alone [5, 6].

In recent years, the food industry has been increasingly adopting automated process control systems to ensure consistent quality and reduce processing times [7–10].

Hyperspectral imaging (HSI) has emerged as a powerful non-destructive technique for food quality and safety assessment because it combines imaging and spectroscopy to acquire spatially resolved spectra over multiple wavelengths, often in the visible and near-infrared (VIS/NIR) region [11–13]. HSI has been extensively investigated for classification, grading, and defect detection across many agri-food products, demonstrating that spectral signatures can enhance discrimination beyond conventional imaging [11–13]. In the coffee sector, spectroscopic approaches (including NIR) have been used for quality evaluation and defect-related analysis, largely under laboratory or offline conditions [14, 15].

Despite this progress, transferring hyperspectral methods from laboratory studies to industrial environments [16] remains challenging due to constraints on acquisition speed, data processing, real-time classification, and integration within existing process-control architectures. Industrial deployment requires robust and computationally efficient models that can operate at high throughput and interface with programmable logic controller (PLC)-controlled rejection systems. Many studies have shown how useful it is for different products like vegetables [17], oil extraction [18], cereal processing [19, 20], and legumes [18]. Plus, it’s quite cost-effective to manage [21–26].

The novelty of the present study lies in the industrial-scale validation of a VIS/NIR HSI system integrated into a commercial optical sorter for green coffee processing. Rather than introducing a new sensor design, the manuscript focuses on feasibility and performance under real operating conditions, quantifying contaminant removal and yield across sorting passes. The specific objectives are to: (i) evaluate discrimination of green coffee beans from typical organic and inorganic contaminants using VIS/NIR hyperspectral sensing; (ii) quantify sorting performance (contaminant removal, product loss, yield) under industrial conditions; and (iii) discuss industrial implications and positioning relative to conventional colour sorting.

The classification models were implemented on an industrial optical calibration machine equipped with VIS/NIR sensors in the wavelength range 400–1,700 nm in order to discriminate the contaminants present in the coffee samples, according to the experimental plan prepared.

Spectral data processing and model development were performed offline in MATLAB^® (MathWorks, Natick, MA, USA). Standard hyperspectral preprocessing procedures (commonly used in VIS/NIR food analysis) were applied to reduce acquisition variability and improve robustness, including dark/white reference correction and spectral normalization [14, 15].

The classification strategy adopted in this work is a statistical/chemometric approach, selected for computational efficiency, interpretability, and compatibility with real-time industrial constraints. Statistical classifiers remain widely used for hyperspectral food applications and for industrial implementations where robustness and fast inference are required [14, 15]. Although deep learning methods have become increasingly prominent in hyperspectral analysis, their deployment in real-time industrial sorting can be limited by computational load and integration constraints, particularly when strict timing and PLC interfacing are required [27, 28]. For these reasons, a statistical classifier was preferred for the present industrial validation.

The classifier was trained using representative samples of sound coffee beans and each contaminant class (peel, defective beans, stones). Following offline verification, the resulting decision logic was translated into a PLC-compatible set of classification rules and embedded into the sorter control architecture for real-time rejection.

This study provides an industrial-scale validation of VIS/NIR hyperspectral sorting applied to green coffee, addressing a key gap between laboratory demonstrations and real-time deployment. The complete removal of stones after the first pass is particularly relevant for industrial processing, as foreign inorganic materials can cause mechanical damage and safety issues. Compared with conventional RGB-based sorting, which relies mainly on surface colour information, hyperspectral sensing can exploit spectral features related to material composition, potentially improving discrimination when visual appearance overlaps.

The progressive reduction of organic contaminants (peel and defective beans) across sorting passes indicates that the chosen multi-pass strategy is effective for refining product quality (Figure 3). Figure 4a and 4b show the results of the selection, respectively, the selected coffee product and the rejected product with the contaminants. Multi-pass operation is consistent with industrial practice for optical sorting systems when the objective is to maximize contaminant removal while balancing product retention. In our case, most contaminants were removed in the first pass, while the second pass primarily acted as a “polishing” step, further reducing residual fractions.

A key industrial trade-off is the loss of a compliant product. The cumulative yield of approximately 84% reflects the balance between strict rejection criteria and product retention (Figure 5). In high-value processing lines, moderate product losses can be acceptable when removal of harmful contaminants and improvement of batch uniformity reduce downstream risks and enhance final quality. Importantly, the primary contribution of this work is not proposing a new algorithm but demonstrating that a computationally efficient statistical classifier can be embedded into a PLC-based sorter architecture and operate in real time under industrial constraints. This positioning is aligned with the broader literature, highlighting that industrial HSI adoption depends strongly on robust preprocessing, efficient models, and system integration rather than accuracy alone.

While deep learning approaches are increasingly reported for hyperspectral analysis and have shown strong performance in several food applications, their use in industrial sorting remains constrained by computational requirements, model deployment complexity, and the need for deterministic real-time execution. In this context, the statistical approach adopted here represents a pragmatic solution for industrial deployment, while deep learning remains a promising direction for future upgrades when hardware/software constraints permit.

From a sustainability perspective, improved sorting accuracy can reduce waste and enhance resource utilization by limiting reprocessing and preventing the progression of contaminated material into downstream steps. More broadly, advanced physical technologies and automation are increasingly recognized as key enablers of sustainable food processing chains.