Spectral fusion platform to identify ethylene vinyl alcohol (EVOH) and facilitate feedstock sorting

Abstract

The present disclosure relates to techniques including real-time prediction of a presence and / or an amount of Ethylene Vinyl Alcohol (EVOH) based on one or more hyperspectral images of an object or feedstock being processed in real-time (e.g., in a recycling workflow). The feedstock includes e.g., one or more EVOH-free and EVOH-based items, potentially comprising EVOH as a barrier layer. In some aspects, the disclosed techniques include accessing the one or more hyperspectral images associated with a three-dimensional space comprising one spectral dimension and two spatial dimensions acquired via one or more hyperspectral sensors. At least part of the feedstock may be processed via an EVOH detection model predicting a presence of EVOH and an EVOH quantification model to predict a variable corresponding to the amount of EVOH. Based on the predicted variable, the feedstock being automatically moved and processed as part of a recycling handling process, may be routed.

Description

Ally. Ref. No.: 60BN033-PCTSPECTRAL FUSION PLATFORM TO IDENTIFY ETHYLENE VINYL ALCOHOL (EVOH) AND FACILITATE FEEDSTOCK SORTINGCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the priority to and the benefit of U.S. Provisional Application Number 63 / 685,916, filed on August 22, 2024, entitled “Spectral Fusion Platform to Identify EVOH and Facilitate Feedstock Sorting”, which is hereby incorporated by reference in its entirety for all purposes.BACKGROUND

[0002] A significant portion of plastic products are designed for single use and may often not be recovered through recycling systems. Global plastic production is estimated to exceed 350 million tons annually. Of this, approximately 9-10% is recycled, around 12% is incinerated, and the remaining (e.g., 75-80%) accumulates in landfills or the natural environment. In such environments, many plastics may take hundreds to over a thousand years to degrade, depending on the polymer type and environmental conditions. Global plastic production is projected to double by 2030 and may triple by 2050 if current trends continue. Efficient recycling processes may rely on accurate material characterization, effective sorting, and reliable prediction of decomposition behavior and yield.

[0003] In some recycling infrastructures or industries, mixed streams of objects or materials are typically sorted using automated equipment or machinery that may distinguish materials by size, weight, and / or density. Some approaches may take advantage of computer vision-based methods that may incorporate various spectroscopy methods (e.g., UV-Vis, X-ray, Infrared). Currently, industrial grade optical sorting equipment, found in modem MRFs (material recovery facilities), typically uses select bands in the near infrared (NIR) range (700-1000 nm) to determine base polymer composition of plastics. However, every material or element may not be characterized by these select regions in the near-infrared frequency bands.

[0004] Meanwhile, as there is continued pressure on the recycling industry to improve recycling efficiency, identification as to what composes plastic waste at a more granular, molecular level - i.e., the polymer composition or blends, formulated additives and contaminants, may also be a concern. With the knowledge of the molecular composition ofAtty. Ref. No.: 60BN033-PCT feedstock, a quality, and thus the value, of recycled output (for both mechanical and chemical recycling) may be effectively improved.

[0005] The quality of the recycling output may also depend on whether feedstocks used to produce the recycling output included Ethylene Vinyl Alcohol (EVOH) and / or an amount of EVOH in the feedstocks. EVOH is a copolymer that is often used to form a barrier layer sandwiched between layers of plastics, which can include (for example) polyethylene terephthalate (PET), polyethylene (PE), or polypropylene (PP). The EVOH barrier layer can prevent or reduce gas or moisture transfer between packaged products and the outside environment. While important to maintain freshness of packaged products, EVOH may have adverse impacts on recycling processes.

[0006] In mechanical recycling, EVOH may cause discoloration and other issues which can lower the quality and value of recycled plastics. The Association of Plastic Recyclers’ APR Design® Guide indicates that <6.0 wt.% EVOH is compatible with high-density polyethylene (HDPE) and polypropylene (PP) recycling; however, it is considered a contaminant that requires additional testing for polyethylene (PE) film and polyethylene terephthalate (PET) recycling. EVOH may also be considered a contaminant for higher end HDPE and PP recycling applications where color or appearance is a concern. Similarly, in chemical recycling, EVOH can reduce the quality and yield of products. For example, EVOH barrier layers are a source of oxygen in pyrolysis feedstock and result in the formation of non-targeted oxygenated molecules and reduced yield of pyrolysis oil. Though often viewed as a contaminant, EVOH can also unlock opportunities in recycling. Since EVOH is typically used for food packaging, the presence of EVOH can be used as a selective signal for sorting food grade plastics.SUMMARY

[0007] Certain aspects of the present disclosure relate to techniques including real-time prediction of a presence and / or an amount of Ethylene Vinyl Alcohol (EVOH) in an object or feedstock based on one or more hyperspectral images. The object or feedstock may include, for example, EVOH-free or EVOH-based items that potentially comprise EVOH as a barrier layer. In some aspects, the disclosed techniques may include accessing the one or more hyperspectral images that may be acquired via one or more hyperspectral sensors (or cameras) configured to capture the object or feedstock being processed in real-time (e.g., in a recycling workflow). EachAtty. Ref. No.: 60BN033-PCT of the one or more hyperspectral images is associated with a three-dimensional space comprising a spectral dimension and two spatial dimensions.

[0008] In some instances, each of the one or more hyperspectral sensors is sensitive to a set of wavelengths (or frequencies). For example, the one or more hyperspectral sensors may be configured to capture the one or more hyperspectral images across spectral bands (e.g., nearinfrared (NIR, -700-1000 nm), shortwave infrared (SWIR, -1000-2500 nm), midwave infrared (MWIR, -3-5 pm), and / or longwave infrared (LWIR, -8-14 pm)).

[0009] At least part of the object or feedstock may be processed via an EVOH detection model that is configured to predict whether the EVOH is present in the object or feedstock based on the one or more hyperspectral images. In response to predicting that the at least part of the object or feedstock includes the EVOH, a variable corresponding to the amount of EVOH may be predicted in the at least part of the object or feedstock using an EVOH quantification model. In some instances, the EVOH detection model corresponds to a supervised machine-learning classifier that is configured to generate, based on at least part of the one or more hyperspectral images, a predicted probability representing a likelihood of presence of the EVOH within the object or feedstock. In various instances, the EVOH quantification model corresponds to a supervised machine-learning regressor.

[0010] In some aspects, the EVOH detection model and quantification model may be separately trained, where the EVOH quantification model may be configured to generate the predicted variable based on the prediction of the EVOH detection model and the at least part of the one or more hyperspectral images. Alternatively, the EVOH detection model and EVOH quantification model are jointly trained as a single machine-learning model in an end-to-end manner. Accordingly, the EVOH quantification model may be configured to generate the predicted variable based on the prediction of the EVOH detection model and a compressed spectral representation associated with the at least part of the one or more hyperspectral images. For example, the compressed spectral representation may be generated by averaging spectral values across all spatial pixels within the object (e.g., generating a one-dimensional spectral vector), dimensionality-reduced features such as principal components, spectral embeddings, or learned latent vectors (e.g., feature embeddings extracted from deeper or final layers of the EVOH detection model).Ally. Ref. No.: 60BN033-PCT

[0011] In some instances, the feedstock corresponds to one or more objects that are automatically moved as part of a material handling or inspection process. The prediction(s) as to the presence of and / or the amount of EVOH in the object or feedstock may be used to automatically route the feedstock. For example, the feedstock may be routed to a particular bin or conveyor belt, where the particular bin or conveyor belt is used to collect multiple feedstocks that are to be handled in a given manner. One exemplary type of handling is to route the feedstocks to begin a given type of recycling process (e.g., to be transformed into a pyrolysis oil). The recycling process may be one that is performed with predefined parameters and / or with at least one parameter that is defined based on the predicted cumulative amount of EVOH and / or of one or more other substances in the feedstocks being processed. Another exemplary type of handling is to route the feedstocks to a landfill. In some instances, the handling may further be influenced by one or more other predictions or measurements pertaining to the feedstock.

[0012] In some instances, the EVOH detection model and the EVOH quantification model were trained using a dataset comprising multiple samples (e.g., EVOH-based and EVOH-free) each associated with a corresponding enthalpy value derived from an associated set of differential scanning calorimetry (DSC) curves. To produce the training dataset, a DSC analysis can be run on each sample to determine the amount of EVOH. When running the DSC analysis, different polymers — including EVOH — exhibit distinct phase transition temperatures. These transitions can be identified by subjecting the sample to a thermal cycle comprising an initial heating, followed by cooling, and then a second heating phase. A characteristic feature of a DSC curve of the set of DSC curves that corresponds to EVOH is the presence of an area under a dip or depression within a temperature range, typically with a lower bound of 170°C to 182°C and an upper bound of 186°C to 200°C. Subsequently, EVOH can be identified by this dip in the second heat of the DSC curve using one or more thresholds and / or ranges disclosed herein (e.g., around 180-187°C).

[0013] In some instances, a (unified) hyperspectral image is generated based on multiple hyperspectral images collected from multiple hyperspectral cameras, each of which detects signals from a set or range of wavelengths different than the set(s) or range(s) of wavelengths that are detectable by other(s) of the multiple hyperspectral sensors. In some examples, the multiple hyperspectral images may undergo one or more operations prior to generating the unified hyperspectral image. For example, the one or more operations may include imageAtty. Ref. No.: 60BN033-PCT registration (e g., spatial alignment for pixel-level correspondence), radiometric correction (e g., normalizing pixel intensities and accounting for variations in illumination conditions or other environmental effects), and spectral calibration (e.g., aligning the spectral axis across all image bands).

[0014] The unified hyperspectral image may be generated by fusing the multiple hyperspectral images along the spectral dimension, thereby representing spatial and spectral information along multiple sets or ranges of wavelengths. Subsequently, the processing using the EVOH detection model or the prediction of the variable is performed using the unified hyperspectral image.

[0015] In some instances, one or more hyperspectral images may capture multiple objects within a single frame. To enable the processing using the EVOH detection model or the prediction of the variable for each individual object, the one or more hyperspectral images may be segmented via a parsing model. The parsing model may be configured to isolate or separate the multiple objects from one another by generating one or more masks that align with the spatial dimensions of the one or more hyperspectral images. For example, the one or more masks may include a labeled mask that assigns an object identifier or a background label to each pixel, or a set of binary masks where each binary mask corresponds to an individual object detected within t hyperspectral images.

[0016] In some embodiments, a system is provided that includes one or more data processors and a non-transitory computer readable storage medium containing instructions which, when executed on the one or more data processors, cause the one or more data processors to perform part or all of one or more methods disclosed herein.

[0017] In some embodiments, a computer-program product tangibly embodied in a non- transitory machine-readable storage medium, including instructions configured to cause one or more data processors to perform part or all of one or more methods or processes disclosed herein.

[0018] In some embodiments, a system is provided that includes one or more means to perform part or all of one or more methods or processes disclosed herein.

[0019] The terms and expressions which have been employed are used as terms of description and not of limitation, and there is no intention in the use of such terms and expressions of excluding any equivalents of the features shown and described or portions thereof,Atty. Ref. No.: 60BN033-PCT but it is recognized that various modifications are possible within the scope of the invention claimed. Thus, it should be understood that although the present invention as claimed has been specifically disclosed by embodiments and optional features, modification and variation of the concepts herein disclosed may be resorted to by those skilled in the art, and that such modifications and variations are considered to be within the scope of this invention as defined by the appended claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee. The present disclosure is described in conjunction with the appended figures:

[0021] FIG. 1 shows an exemplary architecture of disclosed techniques including real-time automated detection and quantification of Ethylene Vinyl Alcohol (EVOH) in one or more objects or feedstock based on hyperspectral imaging, in accordance with some aspects of the present disclosure.

[0022] FIG. 2 shows an exemplary illustration for spectral fusion of multiple hyperspectral images in accordance with some aspects of the present disclosure.

[0023] FIG. 3 illustrates an exemplary network for constructing labeled training dataset for machine-learning (ML)-based EVOH detection and quantification in accordance with some aspects of the present disclosure.

[0024] FIG. 4 illustrates a block diagram of various models associated with an EVOH assessment engine, in accordance with some aspects of the present disclosure.

[0025] FIG. 5A illustrates an exemplary differential scanning calorimetry (DSC) graph for an EVOH-based sample, in accordance with some aspects of the present disclosure.

[0026] FIG. 5B illustrates a magnified view of the exemplary DSC graph from the FIG. 5A highlighting a thermal event corresponding to EVOH melting.Atty. Ref. No.: 60BN033-PCT

[0027] FIG. 6A illustrates a first exemplary composite diagram depicting hyperspectral and visual data acquired from multiple cameras for an EVOH-based plastic object or feedstock.

[0028] FIG. 6B illustrates a second exemplary composite diagram depicting the hyperspectral and visual data acquired from the multiple cameras for an EVOH-free plastic object or feedstock.

[0029] FIG. 7A shows an exemplary confusion matrix for an EVOH detection model for multiple samples.

[0030] FIG. 7B shows exemplary performance results of an EVOH quantification model for the multiple samples with known EVOH loading levels.

[0031] FIG. 8 shows exemplary performance results of the EVOH quantification model illustrating a mean value of predicted EVOH enthalpy values for the multiple samples with the known EVOH loading levels.

[0032] FIG. 9 shows an exemplary illustration depicting separate performance results for nine different plastic samples validating the disclosed techniques in accordance with some aspects of the present disclosure.

[0033] FIG. 10 illustrates an exemplary workflow of techniques including real-time prediction of a presence and / or an amount of EVOH in an object or feedstock based on one or more hyperspectral images in accordance with some aspects of the present disclosure.DETAILED DESCRIPTION

[0034] Some embodiments of the present disclosure relate to techniques predicting a presence and / or an amount of Ethylene Vinyl Alcohol (EVOH) by processing a hyperspectral image of an object or feedstock being processed in real-time (e.g., in a recycling workflow). The object or feedstock may include, for example, EVOH-free or EVOH-based items such as food packaging films, multilayer containers, flexible pouches, thermoformed trays, co-extruded bottles, and industrial scrap materials that potentially comprise EVOH as a barrier layer. In some instances, the hyperspectral image may include an image generated based on data collected from one or more hyperspectral sensors or cameras sensitive to a set of wavelengths (or frequencies). In some other instances, the hyperspectral image is generated based on data collected fromAlly. Ref. No.: 60BN033-PCT multiple hyperspectral cameras, each of which detects signals from a set or range of wavelengths different than the set(s) or range(s) of wavelengths that are detectable by the other(s) of the multiple hyperspectral cameras.

[0035] The generation of the hyperspectral image may be performed by applying spectral fusion that involves combining multiple hyperspectral images to generate the (unified) hyperspectral image for rich spectral representation. In some examples, the multiple hyperspectral images may undergo one or more operations prior to generating the hyperspectral image. For example, the one or more operations may include image registration (e.g., spatial alignment for pixel-level correspondence), radiometric correction (e.g., normalizing pixel intensities and accounting for variations in illumination conditions or other environmental effects), and spectral calibration (e.g., aligning the spectral axis across all image bands).

[0036] Accordingly, the hyperspectral image may depict signals corresponding to nearinfrared (NIR, -700-1000 nm), shortwave infrared (SWIR, -1000-2500 nm), midwave infrared (MWIR, -3-5 pm), and / or longwave infrared (LWIR, -8-14 pm) (and may include signals corresponding to a wavelength set that span a single, multiple or all of these bands). When the data from multiple hyperspectral cameras (e.g., corresponding to different frequency bands) is spectrally fused to produce a hyperspectral image, the produced hyperspectral image may have a range and / or resolution of wavelengths that is better than the underlying data from either or any of the multiple hyperspectral cameras.

[0037] In various aspects, the disclosed techniques may determine the presence and / or quantification of EVOH, thereby addressing a range of technical problems in applications such as recycling and material handling. These technical problems may include inability of conventional material identification systems to reliably detect thin or embedded EVOH layers in multilayer packaging, which can result in inaccurate sorting, contamination of polyolefin recycling streams, or improper routing of materials for downstream processing. Additional technical problems may arise from inconsistent EVOH content across feedstocks, which complicates process control in recycling workflows such as pyrolysis, compatibilizer-assisted extrusion, or solvent-based recovery — each of which benefits from accurate, real-time composition data. Many conventional systems may lack the ability to detect or distinguish EVOH from other polymers at high throughput rates or may not quantify EVOH content with sufficient accuracy to inform automated decisions. The disclosed techniques may address theseAtty. Ref. No.: 60BN033-PCT limitations by enabling real-time, non-intrusive or non-destructive detection and quantification of EVOH using hyperspectral imaging, with operational capability at high conveyor belt speeds (e.g., up to approximately 3 meters per second (m / s)), thereby supporting high-throughput material handling while avoiding flow interruptions or requiring destructive testing.

[0038] In accordance with some aspects of the disclosed techniques, at least part of the hyperspectral image may be processed by one or more machine-learning models to predict whether the object or feedstock includes EVOH and to predict an EVOH enthalpy. The EVOH enthalpy can be used as a proxy of and / or to predict an amount of EVOH in the object by performing a differential calorimetry (DSC) analysis. The EVOH enthalpy that is predicted can correspond to a predicted area within a dip in a DSC curve within a particular temperature range. The particular temperature range may have a lower bound of (for example) 170 °C, 175 °C, 177 °C, 178 °C, 179 °C, 180 °C, 181 °C, or 182 °C and / or an upper bound of (for example) 200 °C, 195 °C, 193 °C, 190 °C, 189 °C, 188 °C, 187 °C, or 186 °C. As one example, the particular temperature range may be 180-187 °C. It will be appreciated that the DSC curve itself need not to be predicted; rather, the prediction may include a feature that corresponds to the area within the dip within the particular temperature range.

[0039] In some instances, the one or more machine-learning models include a first machinelearning model (referred to herein as an “EVOH detection model”) configured to predict whether the EVOH is present in the object or feedstock. In some instances, the EVOH detection model corresponds to a supervised machine-learning classifier that is configured to generate, based on the at least part of the hyperspectral image (or a preprocessed version thereof), a predicted probability representing a likelihood of the EVOH within the object or feedstock. The one or more machine-learning models may further include a second machine-learning model (referred to herein as an “EVOH quantification model”) that is configured to predict the EVOH enthalpy. In this case, potentially, the processing of EVOH quantification model, corresponding to a supervised regression model, may be conditionally triggered upon detecting a prediction from the EVOH detection model that the feedstock or a product includes EVOH.

[0040] In some aspects, the EVOH detection model and quantification model may be separately trained, where an entirety of the hyperspectral image (or a preprocessed version thereof) is processed by each of the EVOH detection model and the EVOH quantification model. In some instances, one or more portions of the hyperspectral image are first identified, and eachAtty. Ref. No.: 60BN033-PCT portion is then separately processed by the EVOH detection model and / or the EVOH quantification model. Each portion may be defined by predicting that the portion depicts an individual object (e.g., predominately, or exclusively with respect to other objects). The portion(s) may be defined (for example) using a bounding-box technique or segmentation technique. The portion(s) may be defined by processing the hyperspectral image or another corresponding image (e.g., an RGB image of the feedstock) using yet another machine-learning model (referred to herein as a “parsing model”).

[0041] In some aspects, the parsing model may be configured to generate one or more masks that align with the spatial dimensions of the hyperspectral image. For example, the one or more masks may include a labeled mask that assigns an object identifier or a background label to each pixel, or a set of binary masks where each binary mask corresponds to an individual object detected within the hyperspectral image.

[0042] In some other aspects, the EVOH detection model and EVOH quantification model are jointly trained as a single machine-learning model in an end-to-end manner. Accordingly, the EVOH detection model may be configured to process at least part of the hyperspectral image or (or a preprocessed version thereof). The EVOH quantification model, in this case, may generate the predicted variable based on an input comprising the prediction of the EVOH detection model and a compressed spectral representation of the at least part of the hyperspectral image. For example, the compressed spectral representation may be generated by averaging spectral values across all spatial pixels within the object (e.g., generating a one-dimensional spectral vector), dimensionality-reduced features such as principal components, spectral embeddings, or learned latent vectors (e.g., feature embeddings extracted from deeper or final layers of the EVOH detection model).

[0043] The EVOH detection model and / or EVOH quantification model may be trained on a training dataset comprising multiple samples (e.g., EVOH-based and EVOH-free). Each training sample may be associated with a corresponding enthalpy value derived from an associated set of DSC curves (such as curves associated with first heating, first cooling and second heating). To produce the training dataset, DSC analysis can be run on each of the multiple samples to determine the amount of EVOH. When running DSC, different polymers including EVOH have different phase transition temperatures. These transitions can be detected by first heating, then cooling, and finally heating the training sample again. EVOH can be identified by the dip in theAtty. Ref. No.: 60BN033-PCT curve using one or more thresholds and / or ranges disclosed herein (at -180-187 °C) in the second heat. EVOH can then be included as an additional signal passed along other spectral sensor fusion inputs (e.g., as disclosed in U.S. Provisional Patent Application Number 17 / 820,946, filed on August 19, 2022, which is hereby incorporated by reference in its entirety for all purposes).

[0044] The prediction(s) regarding the presence and / or quantity of EVOH in an object or feedstock may provide several technical advantages across various applications. For instance, such predictions can be utilized to enable automated routing of feedstock materials. In some implementations, the feedstock may be routed to a particular bin or conveyor belt, where the particular bin or conveyor belt is used to collect multiple feedstocks that are to be handled in a given manner. One exemplary type of handling is to route the feedstocks to begin a given type of recycling process (e.g., to be transformed into a pyrolysis oil). The recycling process may be one that is performed with predefined parameters and / or with at least one parameter that is defined based on the predicted cumulative amount of EVOH and / or of one or more other substances in the feedstocks being processed. This predictive routing capability supports more efficient and accurate separation of materials, reducing contamination in recycling streams and improving yield quality. In certain cases, the predicted EVOH concentration may exceed acceptable thresholds for recycling compatibility, in which case the feedstock may be automatically diverted to alternative disposal streams, such as landfilling or incineration. In some instances, the handling may further be influenced by one or more other predictions or measurements pertaining to the feedstock.

[0045] In some instances, the handling of the feedstock is further influenced by an output of one or more other types of processing that do not relate to EVOH. For example, part or all of an image (e.g., a hyperspectral image used to predict a presence or amount of EVOH, spectrally fused hyperspectral image, or an RGB image) may be processed to predict the presence or amount of another component (e.g., base polymer, contaminant, PE, PT, PET, etc.), which may influence a handling determination. As another example, a weight or mass of a feedstock may be used to inform a handling of the feedstock. One or more exemplary other types of processing that may be used to influence a handling of a feedstock or for another purpose disclosed herein may include a processing disclosed in U.S. Provisional Patent Application Number 17 / 820,946, filed on August 19, 2022, which is hereby incorporated by reference in its entirety for all purposes. In some instances, the handling may be influenced by data in FTIR-ATR (Fourier TransformAlly. Ref. No.: 60BN033-PCTInfrared Spectroscopy - Attenuated Total Reflectance) data. Though FTIR-ATR data may predict base polymers (as each polymer has a distinct FTIR), EVOH is not represented in FTIR-ATR data. Therefore, the EVOH predictions may be complementary to the predictions from FTIR- ATR data.

[0046] Another technical advantage of the prediction(s) regarding the presence and / or amount of EVOH in the object or feedstock may relate to applications in manufacturing and quality assurance environments. For example, the prediction(s), in some aspects of the present disclosure, may be used to assess whether a multilayer film or molded product includes an EVOH layer of the appropriate thickness or distribution. In some implementations, this assessment may be conducted in-line during production using hyperspectral imaging or similar non-destructive techniques, enabling real-time monitoring and control of the co-extrusion or lamination process. Deviations in predicted EVOH content may trigger automated adjustments to processing parameters, such as feed rates, temperature profiles, or layer ratios, thereby improving consistency in product performance.

[0047] A further example of technical advantage includes the use of EVOH prediction data for verifying compliance with product specifications or industry standards. In certain cases, a material certification system may rely on predicted EVOH content to confirm that packaging meets functional barrier requirements or remains within regulatory thresholds for recyclability. In some instances, these predictions may be stored as part of a digital product passport or batch record, enabling traceability across the supply chain. Furthermore, the ability to detect and quantify EVOH may assist in early identification of equipment malfunctions or raw material inconsistencies, reducing waste and rework. These technical capabilities collectively enhance production efficiency, support quality assurance protocols, and enable more transparent documentation of material composition.

[0048] FIG. 1 shows an exemplary architecture 100 for automated detection and quantification of Ethylene Vinyl Alcohol (EVOH) in one or more objects based on hyperspectral imaging, in accordance with some aspects of the present disclosure. EVOH is commonly used in multilayer plastic materials as a barrier layer (e.g., in food packaging), due to its excellent gas barrier properties. However, the presence of EVOH in plastic materials may interfere with various downstream processes, such as mechanical recycling, chemical recycling (e.g., pyrolysis,Atty. Ref. No.: 60BN033-PCT solvolysis, or depolymerization), and material recovery, due to its distinct chemical composition, polar nature, and frequent use as the barrier layer in multilayer structures.

[0049] EVOH is commonly embedded in objects such as food packaging films, co-extruded bottles, laminates, and blow-mold containers, where it serves to enhance oxygen impermeability. While beneficial for product preservation, its integration with non-polar polymers such as Polyethylene or Polypropylene may complicate recycling efforts. The high oxygen barrier functionality and hydrophilic properties of EVOH can disrupt thermal degradation profiles, reduce material compatibility during melt reprocessing, and lead to product contamination or yield loss in recycling streams. As such, the accurate detection, spatial mapping, and quantitative assessment of EVOH content are a concern for enabling targeted sorting, pretreatment, or material separation strategies. These capabilities are particularly valuable in closed-loop recycling systems, advanced sorting facilities, and circular economy frameworks, where precise material composition is efficiently managed to meet regulatory compliance, eco-design goals, and material reuse standards.

[0050] The exemplary architecture 100 may include multiple hyperspectral sensors or cameras 102-1, 102-2, 102-3, . . ,,102-n that are configured to capture a feedstock being processing in real-time e.g., in a recycling workflow as illustrated in FIG. 1. In some instances, the feedstock includes one or more objects 106-1, 106-2, 106-3, 106-4 placed on a surface such as a conveyor belt 104. It may be understood that, for illustrative purposes, four objects are depicted; however, the number of objects is not limited to this example and may vary depending on the specific implementation or application context. In some instances, the disclosed techniques may include accessing a hyperspectral image that is generated based on data collected from the multiple hyperspectral cameras, each of which detects signals from a set or range of wavelengths different than the set(s) or range(s) of wavelengths that are detectable by the other(s) of the multiple hyperspectral cameras.

[0051] The generation of the hyperspectral image may include transmitting the data comprising multiple hyperspectral images, via a network 108 (wired or wireless), to an EVOH assessment engine 110. The EVOH assessment engine 110 may combine the multiple hyperspectral images by applying spectral fusion 112 to generate the (unified) hyperspectral image for rich spectral representation. In some instances, the hyperspectral image may include an image generated based on data collected from one or more hyperspectral sensors or camerasAtty. Ref. No.: 60BN033-PCT sensitive to a set of wavelengths (or frequencies). Accordingly, the hyperspectral image may include signals corresponding to a single set of wavelengths spanning one, multiple, or all of the spectral bands including, near-infrared (NIR, -700-1000 nm), shortwave infrared (SWIR, -1000-2500 nm), midwave infrared (MWIR, -3-5 pm), and longwave infrared (LWIR, -8-14 pm).

[0052] In some aspects, the EVOH assessment engine 110 may additionally be provided with one or more RGB images of the feedstock from an RGB camera 109. These RGB images may be utilized for object segmentation, visual labeling, quality inspection or parsing operations via a parsing model 114. The parsing model 114 may be optionally employed to segment the hyperspectral image to identify individual objects (e.g., using a bounding box or pixel-level segmentation). When multiple objects are present in a single frame, the parsing model 114 may segment the objects or feedstock enabling per-object EVOH analysis or assessment. The EVOH assessment engine 110 may further include an EVOH detection model 116 that is configured to predict presence or absence of EVOH based on the hyperspectral image of a given object.

[0053] The EVOH detection model 116 may correspond to a machine-learning model that performs a binary classification (i.e., presence or absence). This model may be trained using hyperspectral data and corresponding DSC (differential scanning calorimetry) reference labels, where EVOH is identified by an endothermic dip in a second heating cycle (e.g., 180-187 °C range). In various instances, an EVOH quantification model 118 may predict an amount of presence of EVOH of the given object based on a prediction by the EVOH detection model 116 that EVOH is present. The EVOH quantification model 118 may correspond to a regression model that is configured to predict an enthalpy value, which correlates with the amount or loading of EVOH in the object. The enthalpy value may correspond to the area under the dip in the DSC curve.

[0054] In certain aspects, the EVOH assessment engine 110 may generate an output 124 that indicates the presence or absence of EVOH, along with a quantitative estimate of its concentration or amount within the analyzed object. Additionally, the output 124 may include a variety of visual and analytical representations tailored to the application domain. For instance, spectral maps, color overlays, and material distribution heatmaps may visually depict the spatial location indicating presence and relative abundance indicating the amount of EVOH. Additionally, the output 124 may present confusion matrices, ROC (receiver operatingAlly. Ref. No.: 60BN033-PCT characteristic) curves, precision-recall curves, and model confidence scores to evaluate detection performance. Similarly, box plots, histograms, and distribution charts may be used to summarize statistical variations across batches, regions, objects, or time points.

[0055] The output 124 may further include quantitative parameters or metrics, such as layer thickness estimates, volumetric content, or purity scores, which may be logged for quality control, material certification, regulatory reporting, or adaptive feedback in production lines. The format and depth of output 124 may be dynamically configured, supporting use in research and development, inline industrial inspection, forensic packaging analysis, and automated recycling systems, among others. These outputs may be displayed on a user interface or a user device 122 for visualization by a user 120 or a technology expert. In certain instances, the output 124 may be programmatically forwarded to downstream processes such as sorting control module.

[0056] The output of the EVOH assessment engine 110 may be leveraged to perform various actions. In some instances, the feedstock or the objects on the conveyor belt 104 may be automatically routed based on the predictions or the output 124. For example, objects with high EVOH content (e.g., above a predefined upper threshold) may be diverted to a landfill, EVOH- free or low-EVOH objects (e.g., below a predefined lower threshold) may be routed to pyrolysis, mechanical recycling, or chemical depolymerization processes. Routing may be achieved by actuating diverters or directing the conveyor belt 104 towards a particular bin 126 or destinations. In some instances, routing and handling decisions are made dynamically, based on cumulative EVOH concentration in a batch or stream of feedstock.

[0057] FIG. 2 illustrates an exemplary architecture 200 for spectral fusion 202 of multiple hyperspectral images in accordance with some aspects of the present disclosure. The exemplary architecture 200 may include multiple hyperspectral sensors or cameras 102-1, 102-2, 102-3, . . ,,102-n capturing an object e.g., a packaged box 106-3 lying on a surface 220. Each hyperspectral camera may be configured to capture a hyperspectral image (e.g., 201-1, 201-2, or 201-n) over a specific and distinct spectral band or wavelength range. Unlike conventional imaging systems that operate in visible spectrum (i.e., red, green, and blue), these hyperspectral cameras 102-1 , 102-2, 102-3, . . ., 102-n may capture rich spectral data across a wide range of the electromagnetic spectrum. In this configuration, the hyperspectral cameras are selectively tuned to different spectral regions, such as the near-infrared (NIR, -700-1000 nm), shortwave infrared (SWIR, -1000-2500 nm), midwave infrared (MWIR, -3-5 pm), and longwave infrared (LWIR,Atty. Ref. No.: 60BN033-PCT-8-14 pm). As illustrated in region 214, each spectral band may be particularly sensitive to specific physical or chemical properties of a target material. For example, for various target materials, NIR may be useful for detecting vegetation, organic substance, and moisture content, SWIR may help distinguish plastics or mineral, MWIR may detect thermal contrasts and chemical signatures, and LWIR is typically used for thermal imaging and heat mapping.

[0058] A hyperspectral camera producing the hyperspectral image e.g., 201-1, 201-2, ..., 201-n may represent a three-dimensional (3D) structure with spatial dimensions 216 (x and y representing pixels width and height) and a spectral dimension 218 (z). The spectral dimension 218 may correspond to a number of discrete wavelength channels captured within the assigned spectral range of the hyperspectral camera. For example, a hyperspectral camera e g., 102-n sensitive to SWIR may output an image 201-n of dimensions 640*480* 100, where 100 represents spectral dimension 218 representing the number of narrowband (or wavelength) channels across 1000-2500 nm range.

[0059] It may be understood that the exact spectral resolution and the number of spectral bands may vary depending on the specific configuration and design parameters of each hyperspectral camera. The modular, multi-band configuration may enable simultaneous or sequential acquisition of complementary hyperspectral data, significantly enhancing the ability to detect, classify, or monitor objects or phenomena that would otherwise be invisible or ambiguous in standard optical imaging. This capability is achieved by the use of multiple hyperspectral sensors or cameras 102-1, 102-2, 102-3, ..., 102-n, which provide high spectral fidelity, increased spatial coverage, and more robust material discrimination across a broad wavelength domain.

[0060] In some aspects, the hyperspectral images 201-1, 201-2, ..., 201-n, each covering a different portion of the spectral domain, may be transmitted via the network 108 to a component performing spectral fusion 202. In some instances, the spectral fusion 202 may include one or more operations to generate a hyperspectral image representing a unified 3D hypercube 212. For example, the one or more operations may include image registration 204 that may be performed to spatially align all incoming hyperspectral images 201-1, 201-2, ..., 201-n for pixel-level correspondence across the band-specific images. The image registration 204 may address misalignments caused by sensor position differences, optical distortions, or platform motion. Similarly, radiometric correction 206 may be applied to normalize pixel intensities and accountAtty. Ref. No.: 60BN033-PCT for variations in illumination conditions, sensor gain, exposure differences, or environmental effects. Furthermore, spectral calibration 208 may be carried out to align the spectral axis across all image bands, compensating for sensor-specific spectral response shifts and consistent wavelength mapping.

[0061] In one example, noise reduction and artifact removal techniques may be employed to improve signal quality. Once all input data comprising hyperspectral images 201-1, . . ,201-n are corrected and calibrated, data fusion 210 may be executed to combine the band-specific images into a single unified hyperspectral cube, also referred herein as 3D hypercube 212, which comprise spatially and spectrally aligned information across all wavelengths of interest. The 3D hypercube 212 may enable advanced downstream analysis such as EVOH presence detection and / or quantification. Additionally, the fused 3D hypercube 212 may enable enhanced material discrimination, improved prediction accuracy, and more robust object-level classification, particularly when multiple materials or objects are present within the field of view.

[0062] FIG. 3 illustrates an exemplary network 300 for constructing labeled training dataset 324 for machine-learning (ML)-based EVOH detection and quantification in accordance with some aspects of the present disclosure. In some aspects, the disclosed techniques may involve accessing a set of training samples 302 comprising a diverse set of instances that include both EVOH-based and EVOH-free materials. These samples may represent a wide range of plastic items and polymer-based objects, such as films, flexible and rigid containers or packaging, bottles, jugs, wraps, laminates, and other manufactured forms. The training samples 302 may further include samples composed of single-layer or multilayer polymer structures, where EVOH may be present as a distinct barrier layer or blended within a composite. To enhance robustness and generalizability, the training samples 302 may vary in terms of EVOH concentration, layer thickness, material composition, surface characteristics, manufacturing processes, and environmental exposure conditions (e.g., humidity, heat, or UV). Such diversity enables the ML models to accurately detect, localize, and quantify EVOH across a broad spectrum of real-world use cases and material configurations.

[0063] Each training sample (such as 303) may be placed in a designated area or a surface 220 and subjected to hyperspectral scanning using one or more hyperspectral cameras 102-1, . .. , 102-n. As discussed earlier, each hyperspectral camera may be configured to capture a hyperspectral image across a distinct spectral range (e.g., NIR, SWIR, etc.). In some instances,Atty. Ref. No.: 60BN033-PCT multiple hyperspectral images of the training sample are collected via multiple hyperspectral sensors and may be forwarded for the spectral fusion 202. The spectral fusion module may apply one or more operations e.g., including spatial alignment, radiometric correction, noise removal, and spectral calibration. Subsequently, the multiple hyperspectral images may be fused, thereby generating a hyperspectral image corresponding to the unified 3D hypercube 212. In some other instances, a hyperspectral image may be captured via a single hyperspectral sensor, which may be directly passed on to the labeling module via the network 108, bypassing the spectral fusion 202, as illustrated by a dotted line 321.

[0064] In certain aspects, differential scanning calorimetry (DSC) analysis 304 may be employed for the creation of the labeled ground-truth dataset (or training dataset 324) for training models aimed at EVOH detection and quantification. DSC provides a reliable and standardized method for identifying the presence, absence, and relative content of EVOH in polymer-based samples by measuring heat flow associated with thermal transitions. This thermal analysis is particularly useful for validating spectral features derived from hyperspectral, FTZR, or NIR data, enabling training of machine-learning models using accurate and physically verified labels. To perform DSC analysis 304, a section or portion of the training sample 303 (e g., ~ 5-15 mg) may be excised preferably from a region suspected to include EVOH. This sample may then be placed into a sample pan 308a, which is positioned within a DSC chamber 306 alongside an empty reference pan 308b, both resting on a thermally conductive disc or plate 310, typically made of high-conductivity metal such as silver, aluminum, or a nickel-chromium alloy (e.g., Chromel). This disc serves as the sensor platform, enabling uniform heat distribution and accurate thermal measurements.

[0065] The pans (308a and 308b) and sensor disc 310 are typically surrounded by integrated temperature sensors or thermocouples, which continuously monitor and record thermal differential signal (AT) between the sample and the reference as the DSC chamber 306 undergoes a controlled temperature program using internal heating elements or a furnace 312 (e.g., from room temperature to 300°C). EVOH exhibits a distinct melting peak, generally in the range of 160-190°C, which enables its identification and quantitative estimation based on the area under the melting peak (enthalpy, AH). These thermal signatures can distinguish EVOH from other common polymers such as polyethylene (PE), polypropylene (PP), or polyethylene terephthalate (PET), each of which has its own characteristic melting behavior.Ally. Ref. No.: 60BN033-PCT

[0066] The portion of training sample in the sample pan 308a undergoes three sequential phases during heating and reheating 318 that includes a first heating phase, a cooling phase, and a second heating phase. During the first heating phase, the sample is gradually heated at a controlled rate (e.g., 10 °C / min) via surrounding heating elements or furnace 312 to remove prior thermal history (e.g., relaxation or pre-melting effects). Temperature readings from the temperature sensors may be continuously sampled, and the thermal differential signal (AT) may be computed between the sample and reference pans. The first heating phase typically reaches above the known melting temperature of EVOH to enable thermal reset of the polymer.

[0067] Subsequently, the sample is cooled back to a sub-ambient or baseline temperature (e.g., 25 °C to 50 °C) during the cooling phase at a controlled rate. The cooling phase may promote re-crystallization or molecular organization thus making the EVOH signature more visible in subsequent reheating. Once the sample is cooled down, it is reheated in the second heating phase over a critical temperature range (e.g., -180-187 °C). If the sample or the portion of the sample includes EVOH, an endothermic dip appears in the resulting DSC curve 314, corresponding to the melting transition of EVOH. During both the heating and reheating phases, temperature readings from integrated thermocouples or temperature sensors located beneath the sample pan 308a and reference pan 308b are continuously recorded. The DSC system monitors the temperature difference (AT) between the two pans as the temperature ramps upward at a controlled rate. When the sample undergoes a thermal event — such as melting, it either absorbs or releases heat, causing its temperature to deviate from the reference. This deviation results in a measurable AT, which is then used to calculate the corresponding heat flow using internal calibration constants and thermal conductivity principles.

[0068] The heat flow (Q) (measured in e g., mW, J / g, or W / g) is directly proportional to AT (measured in °C or K), based on Fourier’s law of heat conduction. As such, the DSC curve 314 is generated by plotting temperature (X-axis) versus heat flow (Y-axis) representing the rate of energy absorbed or released by the sample relative to the reference. Thermal transitions (e.g., melting, crystallization, or glass transition) appear as distinct peaks or shifts in the curve. A pair of machine-learning compatible outputs may be derived from the DSC curve 314. The first output may comprise a classification label generated during EVOH detection 316, indicating whether DSC curve included an EVOH-specific dip in the defined range. The second output may include a continuous regression label generated by performing EVOH quantification 320. DuringAtty. Ref. No.: 60BN033-PCT this quantification, the area under the peak, which represents the enthalpy change (H), can be integrated to estimate the amount of EVOH present in the sample. This process enables accurate thermal characterization of materials, providing labels for the training dataset 324 in EVOH detection and quantification models. Once generated or estimated, the DSC-based ground-truth labels are paired with corresponding hyperspectral image by a labeling module 322. Following this approach, each training sample is fully annotated and then added to training dataset 324, forming a dataset usable to fine-tune or train machine-learning models.

[0069] FIG. 4 illustrates an exemplary block diagram 400 of various models associated with EVOH assessment engine 110, in accordance with some aspects of the present disclosure. In certain instances, a hyperspectral image 401 (e.g., representing a 3D hypercube 212 or an image collected via a single hyperspectral camera) may include multiple objects appearing within a single frame. To enable individual EVOH assessment for each object, in one aspect of the present disclosure, the hyperspectral image 401 may be segmented via the parsing model 114 to isolate and separate the objects from one another. Unlike conventional RGB or grayscale images, the hyperspectral image 401 may provide a broad range of contiguous spectral bands (e.g., in hundreds) for each pixel associated with two spatial dimensions x, y and one spectral dimension. This enables fine-grained discrimination of materials based on their unique spectral signatures, which is particularly useful in applications such as remote sensing, surveillance, biomedical imaging, and material classification.

[0070] To perform parsing, preprocessing 402 may be performed to the hyperspectral image 401, e.g., including normalization of spectral intensities across wavelengths, dimensionality reduction with methods such as principal component analysis (PCA), independent component analysis (ICA), or t-SNE. These methods may be applied to reduce computational load — caused by the high spectral dimensionality of hyperspectral image — and to enhance the signal -to-noise ratio (SNR) by generating compact, informative representations of the data. Specifically, these techniques may transform the original spectral bands into a smaller set of uncorrelated or statistically independent components, which preserve the most significant spectral variance or underlying sources. For example, a hyperspectral image 401 originally having dimensions m x n x b (spatial x spatial x spectral) may be reduced to m x n x , where k « b represents top k components retaining most of the meaningful spectral information. This dimensionalityAlly. Ref. No.: 60BN033-PCT reduction may facilitate more efficient and robust downstream tasks such as object segmentation, classification, or anomaly detection.

[0071] Once the hyperspectral image 401 is preprocessed, the parsing model 114 may apply segmentation 404 e.g., via a machine-learning model, where both unsupervised and supervised methods may be employed. In some instances, the segmentation 404 is performed via unsupervised techniques such as k-means clustering, mean-shift, or spectral-spatial clustering that aim to group pixels based on similarity in spectral space independent of prior labels. These unsupervised techniques are particularly useful for exploratory analysis or when ground-truth data is unavailable. In some other instances, segmentation 404 may be performed via supervised learning that involves training machine-learning or deep learning models using annotated hyperspectral datasets. Algorithms such as support vector machines (SVM), random forests, or more advanced deep networks such as 3D CNNs or hybrid CNN-RNN architectures may exploit both the spectral and spatial patterns to accurately classify and segment objects. In some instances, deep learning approaches may be utilized that involve spectral attention mechanisms or transformer-based models to selectively focus on the most informative wavelengths.

[0072] In some aspects, segmentation 404 may output pixel-wise masks, while in others, it may generate bounding boxes or object-level metadata — depending on the type of segmentation model used and the downstream application requirements. In some instances, semantic segmentation may be applied, where each pixel in the hyperspectral image 401 is classified into a predefined category. The output is typically a 2D label mask that aligns with the spatial dimensions of the hyperspectral image 401 (or preprocessed hyperspectral image), assigning an object identifier — or a background label — to each pixel. This pixel-level labeling may distinguish spatial regions corresponding to individual physical objects captured in the image or frame.

[0073] In some instances, the segmentation 404 may output a set of binary masks, where each mask corresponds to an individual object detected within the hyperspectral scene. For example, if there are 5 objects in a hyperspectral image or hypercube, the output is a set of 5 binary masks, each having dimensions m x n. These object-level masks may be accompanied by class labels and confidence scores, depending on the segmentation model architecture. This type of output is particularly valuable when downstream processing — such as EVOH assessment involving EVOH detection and quantification — is to be performed independently for eachAtty. Ref. No.: 60BN033-PCT object. In addition to binary masks, a set of spatial positions may also be generated that corresponds to a bounding box spatially enclosing a segmented object. These bounding boxes provide a compact and computationally efficient means of isolating object regions from the 3D hypercube 212, enabling targeted analysis or cropping.

[0074] Subsequently, a portion 405 of the hyperspectral image including an object of the multiple objects may be processed by the EVOH detection model 116 that is configured to determine whether the object includes EVOH. The EVOH detection model 116 may correspond to a supervised classifier trained using labeled hyperspectral images along with corresponding DSC (differential scanning calorimetry) profiles of known EVOH-based and EVOH-free objects or feedstock, as discussed in reference to FIG. 3. In some examples, the training dataset 324 employed for training the EVOH detection model 116 may include the hyperspectral image (such as 401 or 405) of a training sample acquired in conjunction with a corresponding DSC profile that confirms presence or absence of the EVOH in the training sample based on endothermic peak transitions and onset temperatures. The EVOH detection model 116 may be applied on a per object basis following the parsing model 114, such that each segmented portion 405 of the hyperspectral image corresponding to a distinct object or material is individually evaluated.

[0075] To capture both spatial and spectral characteristics of EVOH-based objects or materials from a segmented portion 405 of the hyperspectral image 401, the EVOH detection model 116 may be implemented using a 3D convolutional neural network (CNN), spectral- spatial transformer, or hybrid spectral attention network, where spatial-spectral features are jointly learned. Alternatively, the EVOH detection model 116 may operate on spectral features generated from the segmented object region or portion 405 of the hyperspectral image 401. In such cases, the spatial region corresponding to an object (e.g., as defined by a bounding box or mask) is used to generate a representative spectral signature, typically by computing the mean reflectance (or intensity) at each spectral band by aggregating values from all pixels that belong to the object, effectively collapsing the spatial dimensions and retaining only the spectral signature. The result is a one-dimensional spectral vector — typically of length equal to the number of spectral bands — that characterizes the average spectral response of the object across wavelengths. This approach is particularly useful when using models such as one-dimensional convolutional neural networks (ID CNNs) or recurrent neural networks such as long short-termAtty. Ref. No.: 60BN033-PCT memory (LSTM), which are configured to capture sequential dependencies and patterns across spectral bands, rather than spatial variations.

[0076] In some embodiments, the portion 405 of the segmented hyperspectral image may undergo normalization, dimensionality reduction (e.g., PCA, ICA), and / or contrast enhancement prior to classification (similar to the preprocessing 402 involved in parsing model 114). The output of the EVOH detection model 116 may correspond to a probabilistic prediction score representing the likelihood that the object includes EVOH. This score may be expressed as a value between 0 and 1, where higher values indicate greater certainty in the model’s prediction. Such scores may be used to threshold decisions e g., accurately identifying EVOH presence when the score exceeds a detection threshold, or flagging objects for further inspection or manual review when the score falls within an uncertain range or below the detection threshold.

[0077] In some aspects, processing performed by the EVOH quantification model 118 may be conditionally triggered upon detecting a prediction from the EVOH detection model 116 that a product or object includes EVOH. In some examples, the input to the EVOH quantification model 118 may include the portion 405 of the hyperspectral image including the individual object. Alternatively, the input to the EVOH quantification model 118 may comprise a compressed spectral representation extracted therefrom. For example, a mean spectral vector, obtained by averaging the spectral values across all spatial pixels within the object, or dimensionality-reduced features such as principal components, spectral embeddings, or learned latent vectors (e.g., generated feature embeddings extracted from the deeper or final layers of the EVOH detection model 116).

[0078] The EVOH quantification model 118 may be implemented as a supervised regression model that is trained on the training dataset 324 comprising hyperspectral images of various training samples with corresponding measured enthalpy values derived from the DSC analysis 304. Depending on the input representation, several types of supervised regression models may be applied for EVOH quantification. In some instances, when the input comprises the full segmented portion 405 of the hyperspectral image — retaining both spatial and spectral information — models capable of learning complex spatial -spectral correlations are employed. These include 3D convolutional neural networks (3D CNNs), spectral-spatial transformers, and hybrid attention-based architectures, which can directly operate on multi-dimensional data and learn to extract relevant features for regression. Alternatively, when the input is a compressedAtty. Ref. No.: 60BN033-PCT spectral representation — such as an object-level mean spectral vector, principal components, or latent embeddings from earlier detection model — simpler or lower-dimensional models may be employed. These may include traditional supervised regressors such as linear regression, partial least squares regression (PLSR), and support vector regression (SVR), as well as more flexible nonlinear models such as multi-layer perceptrons (MLPs) or ID convolutional neural networks (ID CNNs). The choice of model typically balances factors such as the complexity of the spectral patterns, the amount of training data available, computational constraints, and the required interpretability of the output.

[0079] The output of the EVOH quantification model 1 18 may include a predicted EVOH enthalpy 406, which serves as a proxy of and / or to predict an amount of EVOH in the object. The EVOH enthalpy that is predicted can correspond to a feature within a DSC curve specifically, the area under an endothermic dip corresponding to EVOH melting within a particular temperature range. The particular temperature range may have a lower bound of (for example) 170 °C, 175 °C, 177 °C, 178 °C, 179 °C, 180 °C, 181 °C, or 182 °C and / or an upper bound of (for example) 200 °C, 195 °C, 193 °C, 190 °C, 189 °C, 188 °C, 187 °C, or 186 °C. As one example, the particular temperature range may be 180-187 °C. It will be appreciated that the DSC curve itself need not be predicted; rather, the model is trained to regress a single-value feature that corresponds to the enthalpic area within the specified dip of the DSC curve. In some aspects, the predicted enthalpy may be used to infer presence of multilayer structures, determine laminate compositions, or trigger downstream classification or sorting actions based on material content.

[0080] FIG. 5A illustrates an exemplary DSC graph 500-Afor a sample including EVOH, in accordance with some aspects of the present disclosure. In some implementations, as discussed earlier, the DSC apparatus may include temperature sensors positioned beneath both the sample and reference pans 308a and 308b, respectively to monitor temperature differences. The system calculates the heat flow based on the energy required to maintain a uniform temperature profile between the sample and the reference. These heat flow signals may be continuously recorded or processed to generate various thermal curves within the DSC graph 500, as depicted in FIG. 5A. The DSC graph 500-A shows thermal profiles in which heat flow (Q) normalized per unit mass watts per gram (W / g) is plotted on vertical axis 502 and temperature (°C) on horizontal axis 504.

[0081] The thermal profiles comprise three distinct thermal curves, each representing the heat flow response of the sample over a range of temperatures (e.g., -100 °C to 300 °C). AsAtty. Ref. No.: 60BN033-PCT indicated by legend 503, a red curve 506 represents a first heating phase during which the sample is heated from a sub-ambient temperature up to above an expected melting point range. A green curve 508 represents a thermal profile of the first cooling phase, where the molten sample is cooled under controlled conditions. Similarly, a blue curve 510 represents a thermal profile of a second heating phase, which reveals thermal transitions that are not obscured by prior thermal history.

[0082] The exemplary DSC graph 500-A further includes a large endothermic peak (not annotated) that may be attributed to a different polymer component such as polyethylene (PE), as the location (e.g., temperature value at x is not consistent with EVOH presence). However, a small endothermic peak 512 in the blue (second heat) thermal profile centered at approximately 186.98 °C, reflects characteristic behavior of the EVOH. The onset temperature of the thermal event where the melting starts occurs at 179.95 °C, and the associated normalized transition enthalpy, as determined by the area under the peak, is 1.5497 J / g. This specific thermal response is used as diagnostic signature for the presence of EVOH in multilayer or composite plastic materials. A relatively low enthalpy value and the location of the thermal event are consistent with EVOH present in a small fraction within the sample.

[0083] FIG. 5B illustrates a magnified view 500-B of the exemplary DSC graph 500-A from the FIG. 5 A highlighting a thermal event corresponding to the EVOH melting. For visual clarity, the FIG. 5B shows normalized heat flow as a function of temperature over a narrower temperature range of 140 °C to 240 °C. The smaller endothermic peak 512 is again noted and clearly shown with a corresponding enthalpy of approximately 1.5497 J / g and an associated onset temperature indicated by a dotted line 514 of value 179.95 °C and a peak temperature of 186.98 °C. The baseline 516 used for calculating area under the dip is shown intersecting the curve on either side of the blue curve 510 corresponding thermal profile of the second heat. The endothermic peak 512 from the second heating phase may be used to validate presence of EVOH based on the peak location and shape. Additionally, this endothermic peak 512 may quantify a relative concentration of EVOH via the integrated enthalpy. These thermal profiles obtained during DSC analysis 304 may be integrated to generate ground-truth labels for the accurate EVOH detection and quantification.

[0084] FIG. 6A illustrates a first exemplary composite diagram 600-A depicting hyperspectral and visual data acquired from multiple cameras for a plastic object or feedstockAtty. Ref. No.: 60BN033-PCT comprising EVOH. The first composite diagram 600-A includes a spectral plot 602 derived from a mid-wave infrared (MWIR) sensor FX50, corresponding to the object acquired. The spectral plot 602 exhibits characteristic absorption and reflectance features across a range of wavelengths in the mid-wave infrared (MWIR) spectrum that are indicative of EVOH presence. The first composite diagram 600-A further comprises images 604, 606, and 608 from three different sensors (or cameras) including: FX17 (SWIR), FX50 (MWIR), and JAI, respectively. Each of these cameras captures spectral or visual data from the same spatial region of the plastic object or feedstock but across different ranges of wavelengths. The FX17 (SWIR) operates in the shortwave infrared (SWIR) spectrum and highlights certain absorption features relevant to polymer classification. The FX50 (MWIR), which captures MWIR data, provides distinct contrast and spectral differentiation of EVOH within the multilayer plastic structure and corresponds to the spectral profile 602. The JAI sensor may capture visible-range imager, providing visual context for identifying the physical layout and morphology of the plastic object of feedstock.

[0085] FIG. 6B illustrates a second exemplary composite diagram 600-B depicting hyperspectral and visual data acquired from multiple cameras for another plastic object or feedstock that is EVOH-free. Similar to FIG. 6A, the second composite diagram 600-B includes another graphical representation of spectral plot 616 corresponding to the EVOH-free plastic object acquired using the sensor FX50 (MWIR). The spectral plot 616 exhibits characteristic absorption and reflectance features across a range of wavelengths in the mid-wave infrared (MWIR) spectrum that are indicative of EVOH absence. The second composite diagram 600-B further comprises images 610, 612, and 614 from three different sensors (or cameras) including: FX17 (SWIR), FX50 (MWIR), and JAI, respectively. A comparative analysis of MWIR spectral plots 602 and 616 exhibits distinctive reflectance behavior in the first few channels (approximately range: 0-20). The comparison shows a sharp decline in the MWIR spectral plot 602 indicating EVOH presence, while the MWIR spectral plot 616 remains relatively flat for the first few channels with minor fluctuations indicating absence of EVOH.

[0086] FIG. 7A shows an exemplary confusion matrix 700-A for an EVOH detection model 116 trained on the training dataset 324 including non-black plastic samples. A non-black plastic is any plastic material that does not include carbon black pigment, allowing it to reflect or transmit light for easier detection by optical or hyperspectral sensors. Examples include clear or colored PET (Polyethylene Terephthalate) bottles, white HDPE (High-Density Polyethylene)Atty. Ref. No.: 60BN033-PCT containers, and colored polypropylene (PP) items. The confusion matrix 700-A corresponds to performance on a validation set scanned at belt speed of 3 m / s, demonstrating real-time in-line detection capability in accordance with the disclosed techniques. The EVOH detection model 116 was trained using supervised machine-learning techniques to classify whether an object of feedstock includes EVOH based on its one or more hyperspectral images. The x-axis and y-axis of the confusion matrix 700-A represent actual labels and predicted labels, respectively with counts provided for each class decision.

[0087] In total, 5480 samples were evaluated, with EVOH-positive samples comprising 1403 instances, and EVOH negative samples comprising 4077 instances. With given validation set, the EVOH detection model 116 achieved performance results 702, as illustrated in FIG. 7A, comprising a classification accuracy of 99.7%, with a precision of 99.4%, and a recall of 99.5%. These performance results 702 validate the ability of the disclosed EVOH assessment techniques, including detection and quantification, to reliably distinguish EVOH-based plastics from EVOH-free objects, materials, or feedstock.

[0088] FIG. 7B shows an exemplary graph 700-B showing results of the EVOH quantification model for samples with known EVOH loading levels in accordance with some embodiments of the present disclosure. The exemplary graph 700-B displays information about EVOH in terms of its thermal behavior and concentration within various samples. In the illustrated graph 700-B, a vertical axis represents EVOH enthalpy 704, which is the amount of heat absorbed or released by the EVOH during a phase transition — measured during the DSC analysis 304 in Joules per gram (J / g). A horizontal axis shows EVOH loading 706, where “loading” refers to the concentration or amount of EVOH in the sample or multilayer structure. The loading is represented in percentage (%), indicating the weight or volume proportion of EVOH relative to the total material. Higher loadings typically indicate more EVOH present, which may affect the thermal and barrier properties of the material.

[0089] The illustrated graph 700-B comprises three box plots 708, 710, and 712 corresponding to three known EVOH loadings of 2.5%, 5%, and 10%, respectively. Each box plot corresponds to an interquartile range (IQR) representing distribution of predicted enthalpy values for each loading category, based on multiple samples per EVOH loading that are processed using the disclosed EVOH assessment engine 110. The red crosses associated with each box plot correspond to actual enthalpy values as identified during the DSC analysis 304.Atty. Ref. No.: 60BN033-PCTThe predicted enthalpy values were generated by the trained machine-learning regressor model, EVOH quantification model 118, which outputs the predicted EVOH enthalpy 406 based on one or more hyperspectral images. The circles associated with each box plot, e.g., 709 may represent the outliers. The experimental results in the illustrated example achieved an R-squared score of 0.7392, demonstrating strong correlation between the predicted and actual values.

[0090] From the exemplary graph 700-B, it may also be observed that the three box plots 708, 710, and 712 are non-overlapping, which also indicates that the disclosed techniques are capable of distinguishing between different EVOH loading levels with high confidence scores. As shown in FIG. 7B, this capability may enable selective acceptance or rejection of EVOH- based samples in real-time based on one or more enthalpy thresholds. For instance, accepting samples or items with low EVOH loading and rejecting those with high EVOH content, thereby maintaining acceptable tolerance levels for downstream recycling or processing operations.

[0091] FIG. 8 shows exemplary performance results 800 of the EVOH quantification model 118 illustrating a mean value of predicted EVOH enthalpy values for the multiple samples with the known EVOH loading levels. The performance results 800 show three plots 802, 804, and 806, corresponding to known EVOH loadings of 2.5%, 5%, and 10%, respectively that are drawn on the horizontal axis, while the vertical axis represents EVOH enthalpy 704. Each plot includes blue dots labeled as 802-2, 804-2, and 806-2, representing the mean predicted enthalpy value across all test samples associated with a corresponding EVOH loading 706. Additionally, each plot includes red crosses labeled as 802-1, 804-1, and 806-1 representing a mean of actual enthalpy values measured during the DSC analysis 304 for the test samples. It may be observed from these plots 802, 804, and 806 that a vertical line connecting the means of actual and predicted values visually indicates a confidence interval range, for example, ± 95%. The confidence interval reflects the variability of the prediction of the EVOH quantification model 118 for each loading level and provides a statistical measure of certainty in estimation.

[0092] FIG. 9 shows an exemplary illustration 900 depicting separate performance results for nine different plastic samples validating the disclosed EVOH assessment techniques, in accordance with some aspects of the present disclosure. In the illustrated example 900, a blind test was conducted on these nine plastic samples which had not been previously analyzed or included during training. The DSC analysis 304 was conducted on these nine samples to determine empirical values of enthalpy corresponding to the EVOH quantity present in theseAtty. Ref. No.: 60BN033-PCT samples. The EVOH presence and EVOH enthalpy predictions are then compared against the DSC results. The exemplary illustration 900 shows nine different plots 902, 904, 906, 908, 910, 912, 914, 916, and 918, corresponding to a known EVOH loading 706 that is drawn on the horizontal axis, while the vertical axis represents enthalpy 704 measured in Joules per gram (J / g). Each plot corresponds to a separate plastic sample, where red dot is the actual enthalpy as identified during the DSC analysis 304. Blue pluses are the predicted enthalpy by the disclosed EVOH quantification model 118. Error bars show ±0.5 J / g of actual enthalpy.

[0093] FIG. 10 illustrates an exemplary workflow 1000 of techniques including real-time prediction of a presence and / or an amount of EVOH in an object or feedstock based on one or more hyperspectral images in accordance with some aspects of the present disclosure. The blocks in workflow 1000 are illustrated in a specific order, while the order may be modified, for example, some blocks may be performed before others, and some blocks may be performed simultaneously. The blocks can be performed by hardware, software, or a combination thereof.

[0094] At block 1002, the disclosed techniques may include accessing the one or more hyperspectral images 201-1, .. ., 201-n that may be acquired via one or more hyperspectral sensors (or cameras) 102-1, . . ., 102-n configured to capture the object or feedstock being processed in real-time (e.g., in a recycling workflow). The object or feedstock may include, for example, EVOH-free and EVOH-based items that potentially comprise EVOH as a barrier layer. Each of the one or more hyperspectral images is associated with a three-dimensional space comprising one spectral dimension and two spatial dimensions. In some instances, each of the one or more hyperspectral sensors is sensitive to a set of wavelengths (or frequencies). For example, the one or more hyperspectral sensors may be configured to capture the one or more hyperspectral images that include signals corresponding to a single set of wavelengths spanning one, multiple, or all of the following spectral bands: near-infrared (NIR, -700-1000 nm), shortwave infrared (SWIR, -1000-2500 nm), midwave infrared (MW1R, -3-5 pm), and / or longwave infrared (LWIR, -8-14 pm).

[0095] At least part of the object or feedstock, at block 1004, may be processed via an EVOH detection model 1 16 that is configured to predict whether the EVOH is present in the object or feedstock based on the one or more hyperspectral images. At block 1006, in response to predicting that the at least part of the object or feedstock includes the EVOH, a variable corresponding to an amount of EVOH may be predicted in the at least part of the object orAlly. Ref. No.: 60BN033-PCT feedstock using an EVOH quantification model 118. In some instances, the EVOH detection model 116 corresponds to a supervised classifier configured to generate, based on at least part of the one or more hyperspectral images, a predicted probability representing a likelihood of presence of the EVOH within the object or feedstock, while the EVOH quantification model 118 corresponds to a supervised machine-learning regressor.

[0096] In some instances, the EVOH detection model 116 and the EVOH quantification model 118 were trained using a training dataset 324 comprising multiple samples (e.g., EVOH- based and EVOH-free), each associated with a corresponding enthalpy value derived from an associated set of differential scanning calorimetry (DSC) curves. In some instances, the feedstock corresponds to one or more objects that are automatically moved as part of a material handling or inspection process. At block 1008, the prediction(s) as to the presence of and / or the amount of EVOH in the object or feedstock may be used to automatically route the feedstock. For example, the feedstock may be routed to a particular bin 126 or conveyor belt 104, where the particular bin 126 or conveyor belt 104 is used to collect multiple feedstocks that are to be handled in a given manner.

[0097] Some embodiments of the present disclosure include a system including one or more data processors. In some embodiments, the system includes a non-transitory computer readable storage medium containing instructions which, when executed on the one or more data processors, cause the one or more data processors to perform part or all of one or more methods and / or part or all of one or more processes disclosed herein. Some embodiments of the present disclosure include a computer-program product tangibly embodied in a non-transitory machine- readable storage medium, including instructions configured to cause one or more data processors to perform part or all of one or more methods and / or part or all of one or more processes disclosed herein.

[0098] The terms and expressions which have been employed are used as terms of description and not of limitation, and there is no intention in the use of such terms and expressions of excluding any equivalents of the features shown and described or portions thereof, but it is recognized that various modifications are possible within the scope of the invention claimed. Thus, it should be understood that although the present invention as claimed has been specifically disclosed by embodiments and optional features, modification and variation of the concepts herein disclosed may be resorted to by those skilled in the art, and that suchAtty. Ref. No.: 60BN033-PCT modifications and variations are considered to be within the scope of this invention as defined by the appended claims.

[0099] The present description provides preferred exemplary embodiments only, and is not intended to limit the scope, applicability, or configuration of the disclosure. Rather, the present description of the preferred exemplary embodiments will provide those skilled in the art with an enabling description for implementing various embodiments. It is understood that various changes may be made in the function and arrangement of elements without departing from the spirit and scope as set forth in the appended claims.

[0100] Specific details are given in the present description to provide a thorough understanding of the embodiments. However, it will be understood that the embodiments may be practiced without these specific details. For example, circuits, systems, networks, processes, and other components may be shown as components in block diagram form in order not to obscure the embodiments in unnecessary detail. In other instances, well-known circuits, processes, algorithms, structures, and techniques may be shown without unnecessary detail in order to avoid obscuring the embodiments.

Claims

Atty. Ref. No.: 60BN033-PCTCLAIMSWhat is claimed is:

1. A method comprising: accessing, via one or more hyperspectral sensors, one or more hyperspectral images of a feedstock being processed in real-time in a recycling workflow, wherein each of the one or more hyperspectral images is associated with a three-dimensional space comprising a spectral dimension and two spatial dimensions, and wherein each of the one or more hyperspectral sensors is sensitive to a set or range of wavelengths; processing that at least part of the feedstock includes Ethylene Vinyl Alcohol (EVOH) via an EVOH detection model that is configured to predict whether the EVOH is present in the feedstock based on the one or more hyperspectral images; in response to predicting that the at least part of the feedstock includes the EVOH, predicting a variable corresponding to an amount of the EVOH in the at least part of the feedstock using an EVOH quantification model, wherein the EVOH detection model and the EVOH quantification model were trained using a dataset comprising EVOH-based and EVOH- free samples each associated with a corresponding enthalpy value derived from an associated set of differential scanning calorimetry (DSC) curves; and routing the feedstock based on the predicted variable.

2. The method of claim 1, wherein the feedstock includes one or more objects that are automatically moved as part of a recycling handling process.

3. The method of claim 1, wherein: the EVOH detection model corresponds to a supervised machine-learning classifier that is configured to generate, based on at least part of the one or more hyperspectral images, a predicted probability representing a likelihood of presence of the EVOH within the feedstock; and the EVOH quantification model corresponds to a supervised machine-learning regressor.

4. The method of claim 1, wherein the EVOH quantification model and the EVOH detection model are jointly trained in an end-to-end manner, and wherein the EVOHAtty. Ref. No.: 60BN033-PCT quantification model is configured to generate the predicted variable based on the prediction of the EVOH detection model and a compressed spectral representation of the at least part of the one or more hyperspectral images.

5. The method of claim 1, wherein the EVOH quantification model and the EVOH detection model are trained separately, and wherein the EVOH quantification model is configured to generate the predicted variable based on the prediction of the EVOH detection model and a compressed spectral representation of the at least part of the one or more hyperspectral images.

6. The method of claim 1, wherein the predicted variable includes a predicted enthalpy indicative of a feature of a DSC curve of the set of DSC curves comprising an area under a dip or depression within a temperature range having a lower bound of 170°C to 182°C and an upper bound of 186°C to 200°C.

7. The method of claim 1, wherein: the one or more hyperspectral images include multiple hyperspectral images captured via multiple hyperspectral sensors, each sensitive to the set or range of wavelengths different than sets or ranges of wavelengths that are detectable by other of the multiple hyperspectral sensors; and the method further comprises: fusing the multiple hyperspectral images along the spectral dimension to generate a unified hyperspectral image representing spatial and spectral information along multiple sets or ranges of wavelengths associated with the multiple hyperspectral sensors, wherein the processing using the EVOH detection model or the prediction of the variable is performed using the unified hyperspectral image.

8. The method of claim 1, wherein the one or more hyperspectral images include multiple objects or feedstocks captured via the one or more hyperspectral sensors, and wherein the method further comprises:Ally. Ref. No.: 60BN033-PCT segmenting the one or more hyperspectral images via a parsing model that is configured to identify a portion of one or more portions within the one or more hyperspectral images by predicting that the portion depicts an individual object of the multiple objects or feedstocks, wherein the prediction comprises generating one or more masks that align with the two spatial dimensions of the one or more hyperspectral images.

9. A system comprising: one or more data processors; and a non-transitory computer readable storage medium containing instruction which, when executed on the one or more data processors, cause the one or more data processors to perform a set of operations including: accessing, via one or more hyperspectral sensors, one or more hyperspectral images of a feedstock being processed in real-time in a recycling workflow, wherein each of the one or more hyperspectral images is associated with a three-dimensional space comprising a spectral dimension and two spatial dimensions, and wherein each of the one or more hyperspectral sensors is sensitive to a set or range of wavelengths; processing that at least part of the feedstock includes Ethylene Vinyl Alcohol (EVOH) via an EVOH detection model that is configured to predict whether the EVOH is present in the feedstock based on the one or more hyperspectral images; in response to predicting that the at least part of the feedstock includes the EVOH, predicting a variable corresponding to an amount of the EVOH in the at least part of the feedstock using an EVOH quantification model, wherein the EVOH detection model and the EVOH quantification model were trained using a dataset comprising EVOH-based and EVOH-free samples each associated with a corresponding enthalpy value derived from an associated set of differential scanning calorimetry (DSC) curves; and routing the feedstock based on the predicted variable.

10. The system of claim 9, wherein the feedstock includes one or more objects that are automatically moved as part of a recycling handling process.

11. The system of claim 9, wherein:Atty. Ref. No.: 60BN033-PCT the EVOH detection model corresponds to a supervised machine-learning classifier that is configured to generate, based on at least part of the one or more hyperspectral images, a predicted probability representing a likelihood of presence of the EVOH within the feedstock; and the EVOH quantification model corresponds to a supervised machine-learning regressor.

12. The system of claim 9, wherein the EVOH quantification model and the EVOH detection model are jointly trained in an end-to-end manner, and wherein the EVOH quantification model is configured to generate the predicted variable based on the prediction of the EVOH detection model and a compressed spectral representation of the at least part of the one or more hyperspectral images.

13. The system of claim 9, wherein the EVOH quantification model and the EVOH detection model are trained separately, and wherein the EVOH quantification model is configured to generate the predicted variable based on the prediction of the EVOH detection model and a compressed spectral representation of the at least part of the one or more hyperspectral images.

14. The system of claim 9, wherein the predicted variable includes a predicted enthalpy indicative of a feature of a DSC curve of the set of DSC curves comprising an area under a dip or depression within a temperature range having a lower bound of 170°C to 182°C and an upper bound of 186°C to 200°C.

15. The system of claim 9, wherein: the one or more hyperspectral images include multiple hyperspectral images captured via multiple hyperspectral sensors, each sensitive to the set or range of wavelengths different than sets or ranges of wavelengths that are detectable by other of the multiple hyperspectral sensors; and the set of operations further comprises: fusing the multiple hyperspectral images along the spectral dimension to generate a unified hyperspectral image representing spatial and spectral information along multipleAlly. Ref. No.: 60BN033-PCT sets or ranges of wavelengths associated with the multiple hyperspectral sensors, wherein the processing using the EVOH detection model or the prediction of the variable is performed using the unified hyperspectral image.

16. A computer-program product tangibly embodied in a non-transitory machine- readable storage medium, including instructions configured to cause one or more data processors to perform to perform a set of operations comprising: accessing, via one or more hyperspectral sensors, one or more hyperspectral images of a feedstock being processed in real-time in a recycling workflow, wherein each of the one or more hyperspectral images is associated with a three-dimensional space comprising a spectral dimension and two spatial dimensions, wherein the feedstock includes one or more objects that are automatically moved as part of a recycling handling process, and wherein each of the one or more hyperspectral sensors is sensitive to a set or range of wavelengths; processing that at least part of the feedstock includes Ethylene Vinyl Alcohol (EVOH) via an EVOH detection model that is configured to predict whether the EVOH is present in the feedstock based on the one or more hyperspectral images; in response to predicting that the at least part of the feedstock includes the EVOH, predicting a variable corresponding to an amount of the EVOH in the at least part of the feedstock using an EVOH quantification model, wherein the EVOH detection model and the EVOH quantification model were trained using a dataset comprising EVOH-based and EVOH- free samples each associated with a corresponding enthalpy value derived from an associated set of differential scanning calorimetry (DSC) curves; and routing the feedstock based on the predicted variable.

17. The computer-program product of claim 16, wherein: the EVOH detection model corresponds to a supervised machine-learning classifier that is configured to generate, based on at least part of the one or more hyperspectral images, a predicted probability representing a likelihood of presence of the EVOH within the feedstock; and the EVOH quantification model corresponds to a supervised machine-learning regressor.Atty. Ref. No.: 60BN033-PCT18. The computer-program product of claim 16, wherein the EVOH quantification model and the EVOH detection model are trained separately, and wherein the EVOH quantification model is configured to generate the predicted variable based on the prediction of the EVOH detection model and a compressed spectral representation of the at least part of the one or more hyperspectral images.

19. The computer-program product of claim 16, wherein the predicted variable includes a predicted enthalpy indicative of a feature of a DSC curve of the set of DSC curves comprising an area under a dip or depression within a temperature range having a lower bound of 170°C to 182°C and an upper bound of 186°C to 200°C.

20. The computer-program product of claim 16, wherein: the one or more hyperspectral images include multiple hyperspectral images captured via multiple hyperspectral sensors, each sensitive to the set or range of wavelengths different than sets or ranges of wavelengths that are detectable by other of the multiple hyperspectral sensors; and the set of operations further comprises: fusing the multiple hyperspectral images along the spectral dimension to generate a unified hyperspectral image representing spatial and spectral information along multiple sets or ranges of wavelengths associated with the multiple hyperspectral sensors, wherein the processing using the EVOH detection model or the prediction of the variable is performed using the unified hyperspectral image.

Images