Multi-spectral detection method for wheat don toxin based on cross-scale biochemical mechanism and physical prior deep attention network

Abstract

This application discloses a multispectral detection method for wheat DON toxin based on a deep attention network with cross-scale biochemical mechanisms and physical prior constraints, belonging to the field of non-destructive testing of agricultural products and intelligent sensing of food safety. The method includes: acquiring tissue cavitation microstructure information of wheat grains with different infection levels using scanning electron microscopy; and obtaining the spatial distribution and depletion characteristics of protein, lipid, and carbohydrate biochemical components using synchrotron radiation Fourier transform infrared microscopy. Based on this, cross-scale mechanism priors are integrated to guide the selection of characteristic bands in short-wave infrared hyperspectral imaging, constructing a multispectral dataset, and establishing a physical prior-driven deep derivative-enhanced spectral attention network to achieve a nonlinear quantitative mapping between SWIR spectra and DON concentration. This method achieves high-precision non-destructive detection and heterogeneity analysis of DON toxin at the single-grain scale.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing and food safety sorting technology for agricultural products, specifically to a multispectral detection method for wheat DON toxin based on cross-scale biochemical mechanisms and physical prior deep attention networks. Background Technology

[0002] Deoxynivalenol (DON) is a secondary metabolite produced by Fusarium fungi during the infection of grains such as wheat. Due to its thermal stability and toxicological hazards, DON contamination not only causes huge economic losses to the global wheat industry chain, but also poses a serious threat to human and animal health. Therefore, achieving rapid, non-destructive, and accurate screening of DON concentration in wheat grains is a major strategic requirement for ensuring national food security and food processing quality.

[0003] Currently, traditional chemical analysis methods such as high-performance liquid chromatography (HPLC) and enzyme-linked immunosorbent assay (ELISA), while highly accurate, suffer from inherent drawbacks such as sample destruction, time-consuming processes, the need for specialized operation, and the inability to achieve online high-throughput detection. In recent years, short-wave infrared hyperspectral imaging (SWIR-HSI) technology has been widely used in the field of non-destructive testing of agricultural product quality due to its ability to simultaneously acquire spatial and spectral information of the target analyte. However, due to the extreme complexity of the internal matrix of wheat grains and the dynamic variability of fungal infection processes, existing hyperspectral-based DON toxin detection methods still face serious theoretical and algorithmic bottlenecks in practical industrial applications. Firstly, existing deep learning models are trapped in a purely data-driven black box, neglecting physical priors and resulting in extremely poor robustness. Current convolutional neural networks (CNNs) and similar systems largely rely on pure data-driven approaches, essentially acting as an uninterpretable black box. Lacking prior knowledge of the microscopic pathogenic mechanisms of fungi, such as cell wall rupture and macromolecular hydrogen bond depolymerization, deep models are prone to fitting background noise or spurious features. The DON distribution within a single wheat grain exhibits strong spatial heterogeneity. Conventional global average pooling operations can obliterate the spectral distortion features of local extreme pathogenic regions, leading to a sharp decline in the model's generalization ability under complex testing environments.

[0004] Secondly, ignoring the kinetics of substrate depletion leads to nonlinear prediction saturation in extremely high concentration ranges. In severe infections, the starch and protein in the wheat endosperm are largely depleted, causing the spectral response to enter a nonlinear plateau. Conventional networks, using the standard mean squared error (MSE) loss function, cannot perceive the biological substrate depletion saturation limit, resulting in a prediction ceiling and severe numerical underestimation when faced with extremely high toxin concentrations.

[0005] Furthermore, the lack of deep integration between mechanistic bands and end-to-end algorithms hinders high-throughput conversion. Full-band hyperspectral imaging is extremely expensive and computationally slow. Although some studies have extracted specific bands, current feature extraction methods largely rely on external static mathematical preprocessing, such as fixed second derivatives. This prevents dynamic and adaptive spectral derivative feature enhancement within deep networks based on sample damage levels, limiting the application potential of multispectral equipment in laboratory-level high-throughput precision screening. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: This application provides a multispectral detection method for wheat DON toxin based on cross-scale biochemical mechanisms and physical prior deep attention networks, including: Acquire wheat grain samples covering different fungal infection gradients and their corresponding shortwave infrared hyperspectral imaging full-band spectral data; The wheat grain samples were characterized by microscopic morphology using scanning electron microscopy to obtain structural cavitation features reflecting tissue degradation. At the same time, the wheat grain samples were subjected to in-situ chemical spatial imaging using synchrotron radiation Fourier transform infrared microscopy to obtain biochemical depletion features of the samples. Using the structural cavitation features and the sample biochemical depletion features as prior information of the mechanism, multiple mechanism feature bands related to DON content are extracted by jointly analyzing the original average spectrum and derivative spectrum of short-wave infrared hyperspectral imaging. Based on the multispectral image data corresponding to the characteristic bands of the mechanism, a physical prior-driven deep derivative enhanced spectral attention network model is constructed, and the model is trained to obtain the DON nonlinear quantitative detection model. The multispectral image data of the unknown wheat sample to be tested under the mechanistic characteristic band is input into the DON nonlinear quantitative detection model, and the absolute concentration value of DON in the sample to be tested is output.

[0007] In one possible implementation, a scanning electron microscope is used to perform microscopic morphological characterization of the wheat grain sample to obtain structural cavitation features reflecting tissue degradation, including: Representative grains were selected from wheat grain samples covering different fungal infection gradients. After glutaraldehyde fixation, gradient ethanol dehydration, critical point drying and vacuum gold sputtering pretreatment, the grains were placed under a scanning electron microscope for high-resolution imaging observation to obtain SEM microscopic images. The SEM images were subjected to in-depth analysis to extract morphological features reflecting the rupture of endosperm cell walls, collapse of starch matrix structure, and cavitation of internal tissues in wheat grains caused by fungal infection, and these morphological features were used as structural cavitation features.

[0008] In one possible implementation, the in-situ chemical spatial imaging of the wheat grain sample using synchrotron radiation Fourier transform infrared microscopy to obtain the sample's biochemical depletion characteristics includes: The absolute concentration of DON toxin in the wheat grain samples was quantitatively analyzed by liquid chromatography-tandem mass spectrometry. Healthy wheat grain samples and those with different toxin levels, quantified by liquid chromatography-tandem mass spectrometry, were embedded, sliced ​​to a predetermined thickness, and then laid flat on a high-transmittance substrate. On the window slats; Using synchrotron radiation Fourier transform infrared microscopy for surface scanning imaging, and relying on the high brightness light source of synchrotron radiation, we can accurately analyze the spatial distribution evolution characteristics of the core nutrient components of starch, lipids and proteins during fungal infection. Identify the breakage and enzymatic depletion states of key biochemical fingerprints, including starch COC and OH bonds, protein amide bands, and lipid CH bonds.

[0009] In one possible implementation, the structural cavitation characteristics and the sample biochemical depletion characteristics are used as prior information of the mechanism. Multiple mechanism characteristic bands related to DON content are extracted through joint analysis of the raw average spectrum and derivative spectrum of short-wave infrared hyperspectral imaging, including: The average spectra of the regions of interest for samples with different fungal infection gradients were extracted to obtain the original average spectra of each gradient sample. The first and second derivative spectra are calculated based on the original average spectra of each gradient sample. The original average spectra, first derivative spectra, and second derivative spectra are superimposed and compared to identify the spectral characteristic peaks and distortion regions that change regularly with increasing DON pollution concentration. Using the structural cavitation features and the sample biochemical depletion features as prior information of the mechanism, and based on the theory of overtone and combination of molecular spectra, candidate feature wavelength ranges associated with starch, protein and lipid depletion were initially identified in the SWIR region. Based on the sensitivity analysis results of derivative spectra changing with pollution gradient, the candidate bands were cross-validated to eliminate non-specific bands affected by moisture or background noise, and finally, several mechanism characteristic bands related to DON toxin content were determined.

[0010] In one possible implementation, the plurality of mechanistic characteristic bands include 994 nm, 1190 nm, 1208 nm, 1426 nm, 1432 nm and 1487 nm; The 994 nm band corresponds to the second harmonic of the OH bond in wheat starch; The absorption characteristics of CH bonds in proteins and lipids at wavelengths of 1190 nm and 1208 nm; The wavelengths 1426 nm, 1432 nm, and 1487 nm correspond to the first overtones of OH bonds or NH bonds in proteins in tissues infected by fungi.

[0011] In one possible implementation, the physics-prior-driven deep derivative-enhanced spectral attention network model includes an input module, the output of which is connected to the input of a deep derivative enhancement module, the output of which is connected to the input of a spatial-spectral dual attention fusion module, and the output of which is connected to a regression prediction module. The spatial-spectral dual attention fusion module includes a spatial attention submodule and a spectral attention submodule.

[0012] In one possible implementation, the loss function used when training the physics-prior-driven deep derivative-enhanced spectral attention network model is: in, For the total loss, For regression error loss, For the derivative consistency constraint loss, For adaptive Box-Cox loss, , These are the weighting coefficients. Batch size This represents the actual DON concentration. To predict concentration, This is the bias value. For the first i Box-Cox transform parameters for dynamic prediction of individual samples For parameter prediction subnetwork, used for output The network layer or module, For input features.

[0013] In one possible implementation, the multispectral image data of the unknown wheat sample to be tested under the mechanistic characteristic band is input into the DON nonlinear quantitative detection model, and the absolute DON concentration value of the sample to be tested is output, including: A multispectral imaging sensor equipped with narrowband filters for the multiple mechanism feature bands was used to perform multispectral imaging on the unknown wheat grains to be tested, and to collect multispectral image data of the unknown wheat sample under the mechanism feature bands. After performing reflectance normalization preprocessing on the acquired multispectral image data, it is input end-to-end into the DON nonlinear quantitative detection model. The first and second derivative features of the input multispectral image data are extracted through the depth derivative enhancement module to enhance the weak absorption and overlapping spectral peak information caused by micro-tissue cavitation and biochemical depletion, and output derivative-enhanced spectral features. The spatial attention submodule enhances the weight of the tissue cavitation region characterized by SEM to determine the spatial location of the contaminated area, and the spectral attention submodule weights the wavelength dimension of the derivative-enhanced spectral features to highlight key spectral information related to biochemical depletion. Subsequently, spatial-spectral joint features are formed through element-wise weighted fusion; The spatial-spectral joint features are nonlinearly mapped by the regression prediction module, and the predicted value after Box-Cox transformation is output. The predicted values ​​after Box-Cox transformation are restored to the true concentration ratio, and the absolute concentration value of DON in the sample to be tested is output.

[0014] In one possible implementation, the predicted value after Box-Cox transformation is restored to the true concentration ratio, and the formula for calculating the absolute concentration value of DON in the sample to be tested is as follows: in, The final output is the predicted concentration of DON toxin in the wheat sample to be tested; It is the inverse Box-Cox transform function. The DON prediction value after Box-Cox transformation is directly output by the model network. These are the optimized Box-Cox transform parameters.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: This application breaks through the algorithm black box, realizing the analysis of heterogeneity in single wheat grains and the precise quantification of extreme concentrations, providing a core algorithm barrier for multispectral high-throughput automated detection and accurate quality identification.

[0016] This application breaks through the limitations of traditional chemometric models, which only know the "what" but not the "why," by endowing multispectral detection features with microscopic structure and tissue decay sources. This upgrades the detection algorithm from a purely statistical, haphazard approach to a method based on physicochemical causality, enhancing the scientific foundation and academic barriers of the technical solution. Through a Box-Cox normalization combined with a RF nonlinear ensemble strategy, the prediction ceiling effect caused by substrate depletion in heavily infected samples is overcome. The model maintains high fitting accuracy across an extremely wide range of toxin concentrations, broadening the detection range and readily addressing the highly complex sample heterogeneity in agricultural fields.

[0017] The 6-band optimization scheme provided in this application successfully compresses the high full-spectral dimension limit to 2.3%. This not only improves the algorithm's computational efficiency exponentially, but also provides direct core algorithmic support for developing low-cost, high-throughput industrial-grade multispectral cameras equipped with customized narrowband filters, clearing away the hardware cost barrier from non-destructive testing laboratories for wheat toxins to grain processing and sorting production lines. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a multispectral detection method for wheat DON toxin based on cross-scale biochemical mechanisms and physical prior deep attention networks, provided for embodiments of this application; Figure 2 This is a diagram of the physical prior-driven deep derivative-enhanced spectral attention network model architecture provided in the embodiments of this application. Detailed Implementation

[0019] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0020] Figure 1 A flowchart illustrating a multispectral detection method for wheat DON toxin based on a deep attention network with cross-scale biochemical mechanisms and physical priors, provided for embodiments of this application, is shown below. Figure 1 This embodiment presents a multispectral detection method for wheat DON toxin based on cross-scale biochemical mechanisms and physical prior deep attention networks, comprising: S101, acquire wheat grain samples covering different fungal infection gradients and their corresponding shortwave infrared hyperspectral imaging full-band spectral data.

[0021] In this embodiment, during the wheat flowering stage, targeted inoculation was carried out in the field by artificially spraying a suspension of Fusarium graminearum spores to simulate the natural disease infection process. After the wheat was fully mature, it was harvested uniformly to obtain grain samples with a continuous toxin concentration gradient. Each harvested grain was individually analyzed for absolute DON toxin concentration using high-performance liquid chromatography-tandem mass spectrometry (HPLC-MS / MS) as a true reference value for modeling. Based on these quantitative results, the samples were strictly divided into a healthy control group and an infected / toxin-containing sample group.

[0022] S102, the wheat grain sample is characterized by microscopic morphology using a scanning electron microscope to obtain structural cavitation features reflecting tissue degradation. At the same time, the wheat grain sample is subjected to in-situ chemical spatial imaging using synchrotron radiation Fourier transform infrared microscopy to obtain the biochemical depletion characteristics of the sample.

[0023] In this embodiment, representative wheat grains were selected from wheat grain samples covering different fungal infection gradients. After glutaraldehyde fixation, gradient ethanol dehydration, critical point drying, and vacuum gold sputtering pretreatment, they were placed under a scanning electron microscope for high-resolution imaging observation to obtain SEM microscopic images. The accelerating voltage of the scanning electron microscope was set to 15 kV. Deep analysis of the SEM images was performed to extract morphological features reflecting fungal infection, including endosperm cell wall rupture, starch matrix structure collapse, and internal tissue cavitation in wheat grains. These morphological features were then used as structural cavitation characteristics.

[0024] As infection progresses, the cell wall network of the wheat endosperm gradually degrades, exposing the previously tightly bound starch granules and causing severe porosity, i.e., cavitation. This collapse of the physical structure alters the depth of light transmission within the grain and amplifies Mie scattering, providing a physical mechanism explanation for subsequent macroscopic spectral baseline shifts and spatial heterogeneity. The morphological characteristics of this cavitation lesion region are summarized as prior knowledge of physical cavitation.

[0025] Relying on the ultra-high spatial resolution of synchrotron radiation sources, the dynamic relationship between toxin accumulation and depletion of macromolecular nutrients by enzymatic hydrolysis can be analyzed in situ at the microscopic level.

[0026] Healthy wheat grain samples and those with different toxicity levels, quantified by liquid chromatography-tandem mass spectrometry, were embedded, sliced ​​to a thickness of 10 μm, and then laid flat on a high-transmittance substrate. On the window slide, surface scanning imaging was performed using a synchrotron radiation Fourier transform infrared microscope. Relying on the high brightness light source of synchrotron radiation, the spatial distribution evolution characteristics of the core nutrients of starch, lipids and proteins during fungal infection were accurately analyzed, and the breakage and enzymatic depletion states of key biochemical fingerprints of starch COC and OH bonds, protein amide bands, and lipid CH bonds were identified.

[0027] In this embodiment, the full spectrum acquisition range is set to 4000–800. The spatial resolution is limited to 10 μm × 10 μm, and the spectral resolution is set to 4. Each pixel undergoes 64 interferogram accumulation scans to maximize the signal-to-noise ratio.

[0028] To accurately assess the physicochemical integrity of the endosperm starch matrix and internal hydrogen bond network in wheat, this method abandons the conventional semi-quantitative rough judgment that relies solely on initial chemical image color mapping. Instead, it rigorously extracts and compares the actual infrared spectral curves of each pixel. Quantitative verification using the absolute absorbance levels of the actual spectral curves establishes that healthy grains (CK) at 3400... At wavenumbers (primarily characterizing the dense hydrogen bond network between / within starch polysaccharide molecules and OH stretching vibrations), it exhibits a significant and extremely stable high absorption intensity. With Fusarium graminearum infection and the accumulation of DON toxin, the fungal hydrolytic enzyme system deeply disrupts this hydrogen bond network, leading to a significant decrease in absorption intensity at wavenumbers of 3400 in infected samples. The actual absorption intensity at that location drops sharply. Based on this conclusive biochemical and physical correspondence, this embodiment will use 3400 The absolute attenuation of the true spectral curve of the band serves as the core reference baseline for anchoring the degree of damage to microscopic biochemical quality and for subsequent spatial normalization of the full-spectrum data.

[0029] Chemical fingerprint analysis of the depletion of core macromolecular nutrients: extracting the integral area of ​​characteristic absorption peaks of specific functional groups to perform microscopic imaging of carbohydrate, protein and lipid metabolic degradation.

[0030] Spatial degradation of carbohydrates / starch (1100-900) The study focused on monitoring the characteristic absorption fingerprints of COC glycosidic bonds and CO stretching vibrations in starch polysaccharides. It confirmed that fungal amylase leads to the massive consumption of both amylose and amylopectin, resulting in a sharp shrinkage of the integrated area in this band of infected samples, microscopically revealing the disintegration of the starch granule skeleton. This pattern is summarized as prior knowledge of biochemical decay kinetics.

[0031] Protein network depolymerization (1700-1500) ): Focus on monitoring amide I band (approximately 1650) (C=O stretching vibration) and amide II band (approximately 1540) (NH bending and CN stretching vibrations). The results show that the gluten network is cut off, the secondary structure is depolymerized, and the characteristic absorption peaks show a gradient decay.

[0032] Lipid consumption (2900-2800) ): Monitors aliphatic methylene (-CH2) and methyl (-CH3) vibrations. As toxin concentration increases, triglycerides are degraded, and lipid distribution becomes less dense.

[0033] S103 uses structural cavitation characteristics and sample biochemical depletion characteristics as prior information of the mechanism. By jointly analyzing the original average spectrum and derivative spectrum of short-wave infrared hyperspectral imaging, multiple mechanism characteristic bands related to DON content are extracted.

[0034] In this embodiment, a pushbroom shortwave infrared hyperspectral imaging system was used to acquire macroscopic three-dimensional data cubes of samples, with a spectral range of 900-1700 nm. The average spectra of regions of interest for samples with different fungal infection gradients were extracted to obtain the original average spectra for each gradient sample. First and second derivative spectra were calculated based on the original average spectra of each gradient sample. The original average spectra, first derivative spectra, and second derivative spectra were superimposed and compared to identify spectral characteristic peaks and distortion regions that showed regular changes with increasing DON contamination concentration. Using structural cavitation characteristics and sample biochemical depletion characteristics as prior knowledge, and based on the harmonic and combination theory of molecular spectroscopy, candidate characteristic wavelength ranges associated with starch, protein, and lipid depletion were initially identified in the SWIR region. Based on the sensitivity analysis results of derivative spectra changing with contamination gradients, candidate bands were cross-validated, and non-specific bands affected by moisture or background noise were eliminated. Finally, six mechanistic characteristic bands related to DON toxin content were determined, including 994 nm, 1190 nm, 1208 nm, and 1426 nm. The wavelengths 994 nm, 1432 nm, and 1487 nm are used to characterize the absorption characteristics of CH bonds in proteins and lipids. The 994 nm band corresponds to the second-order overtone absorption peak of OH bonds in wheat starch, directly reflecting the degree of damage to endosperm starch granules caused by fungal metabolism. The 1190 nm and 1208 nm bands correspond to the absorption characteristics of CH bonds in proteins and lipids, used to characterize the loss of core reserve substances within the grain. The 1426 nm, 1432 nm, and 1487 nm bands correspond to the first-order overtones of OH bonds or NH bonds in proteins in tissues infected by fungi.

[0035] S104. Construct a physical prior-driven deep derivative-enhanced spectral attention network model based on the multispectral image data corresponding to the mechanism characteristic band, and train the model to generate a DON nonlinear quantitative detection model.

[0036] To address the spatial heterogeneity at the single-grain scale of wheat and the substrate depletion saturation phenomenon at extremely high concentrations, this embodiment abandons traditional static preprocessing and conventional machine learning such as random forests, and innovatively constructs an end-to-end physical prior-driven deep derivative-enhanced spectral attention network model. This model achieves adaptive decision-making by deeply embedding hard constraints of physical and biochemical mechanisms obtained from offline source tracing into the network architecture and loss function. See also Figure 2In this embodiment, the physics-prior-driven deep derivative-enhanced spectral attention network model includes an input module. The output of the input module is connected to the input of the deep derivative enhancement module. The output of the deep derivative enhancement module is connected to the input of the spatial-spectral dual attention fusion module. The output of the spatial-spectral dual attention fusion module is connected to the regression prediction module. The spatial-spectral dual attention fusion module includes a spatial attention submodule and a spectral attention submodule. The deep derivative-enhanced spectral module (DDES) has a differentiable deep derivative layer. It uses multi-scale convolution kernels to perform first-order, second-order, and fractional-order dynamic feature extraction on six mechanistic bands in the wavelength dimension to adaptively amplify weak overlapping peaks and spectral distortion features under contamination conditions.

[0037] The specific construction process is as follows: acquire multispectral image data of wheat grain samples prepared in the aforementioned steps in six characteristic bands, and represent them as a data cube. At the same time, load the corresponding reference value of DON absolute concentration determined by liquid chromatography-tandem mass spectrometry.

[0038] A dynamic derivative layer is established at the front end of the neural network. Traditional static derivative preprocessing cannot adaptively adjust to the distortion peaks at different infection stages. This dynamic layer uses convolution kernels of different scales (e.g., 1×3, 1×5) to perform differentiable convolutions in the spectral dimension. Unlike conventional integer-order (1st and 2nd order) derivatives, this module focuses on training the network to learn fractional-order derivative features between bands. The physical logic is that after wheat is infected by Fusarium graminearum, the physical collapse of the endosperm matrix causes extremely slight changes in the slope and peak overlap distortion of its spectral absorption peaks. Integer-order derivatives are prone to losing this weak information or amplifying noise. Through fractional-order derivative convolution, the algorithm can accurately capture the extremely small absorbance changes caused by protein depolymerization or amylose glycosidic bond breakage in the amide band, adaptively amplify the nonlinear distortion features of these overlapping peaks, and generate derivative-enhanced spectral feature maps (FDDES), providing a high signal-to-noise ratio basis for subsequent extraction of highly toxic biochemical fingerprints.

[0039] This embodiment employs a dual-attention mechanism embedding both physical prior space and spectral data to address the spatial heterogeneity of toxins caused by uneven fungal infection in individual wheat grains. Instead of the coarse strategy of whole-grain average pooling, it adopts a dual-attention mechanism with physical prior constraints: using wheat damage pathological image data resolved by SEM, image analysis is performed on high, medium, and low concentration samples to extract morphological features of cavitation and strong Mie scattering in the tissue. These features are summarized as a cavitation probability map or a scattering anomaly mask Mprior, serving as a prior hard constraint for network training.

[0040] A spectral attention network based on a squeeze-excitation structure is constructed. The importance weights of the aforementioned six feature bands are dynamically calculated based on input features, and the bands that best characterize toxin or macromolecular depletion are selected. A spatial attention branch constrained by Mprior is designed. After the network generates a preliminary spatial weight map, an additional physical constraint loss function is introduced to penalize the network for focusing on non-lesion regions. This prior guides the network to adaptively focus and assigns extremely high spatial awareness weights to tissue cavitation (i.e., lesion areas or areas with abnormal scattering) in the seed image. This thoroughly compensates for the spatial heterogeneity interference of toxin distribution in individual seeds from an algorithmic architecture perspective, avoiding signal dilution caused by averaging across all seeds. The weights generated by spectral and spatial attention are multiplied back into the original feature map to generate Fattn. Through this mechanism, the model can ignore background signal interference in healthy areas, effectively compensate for the local distribution of toxins within a single seed, and completely solve the interference of spatial heterogeneity on toxin quantification.

[0041] To address the technical bottleneck of complete enzymatic hydrolysis of the underlying starch / protein substrate during high toxin accumulation, which leads to the failure of conventional linear spectral response models and the creation of a prediction ceiling, this embodiment embeds ABC-Loss in the backpropagation layer. Instead of the conventional MLE method of offline optimization of Box-Cox parameters, an auxiliary branch network (ABC-Net) is established within the PP-DDESAN network. This branch calculates and dynamically outputs the optimal transformation parameters for the current batch of samples online based on the dynamic attenuation characteristics of the input spectral data. ,in, For parameter prediction subnetwork, used for output The network layer or module, For input features.

[0042] The formula for calculating the ABC-Loss loss function is as follows: in, Batch size, which is the number of wheat grain samples used in each loss calculation. This represents the actual DON concentration. To predict concentration, This is the bias value. For the first i Box-Cox transform parameters for dynamic prediction of individual samples.

[0043] The ABC-Loss algorithm replaces the conventional MSE algorithm for calculating prediction error. Its underlying mathematical logic is as follows: when the network input is identified as a spectral sample with severely depleted substrate (extremely high toxin concentration), the gradient descent path of the ABC-Loss algorithm automatically stretches the model's prediction error penalty gradient for that high concentration range. Through this mathematical isomorphism, the model can deeply perceive and accurately reproduce the nonlinear chemical kinetic relationship between toxin growth and substrate depletion, completely breaking through the prediction failure ceiling of linear models in heavily contaminated areas.

[0044] The regression error loss is used to constrain the deviation between the predicted DON concentration and the true value. The derivative consistency constraint loss is used to constrain the consistency between the input spectral derivative features and the model's internal derivative features. Through joint optimization of the total loss function, the model can enhance its robustness to the nonlinear saturation response caused by substrate depletion while maintaining the accuracy of spectral regression. Finally, the optimal PP-DDESAN quantitative model is established.

[0045] To verify the superiority of the knowledge-driven strategy, the experiment conducted a rigorous parallel quantitative comparison between the full-spectrum benchmark model based on 256 bands and the multispectral model based on 6 characteristic bands.

[0046] Experimental data show that this application utilizes only 2.3% of the original optical data (only 6 mechanistic bands), achieving a validation set determination coefficient of 0.87, with excellent relative analysis error performance. The overall prediction performance is highly consistent with the large-scale full-band model. This technical achievement demonstrates that the bands selected for cross-scale biochemical mechanism tracing not only eliminate redundant information but also possess high information purity and representativeness.

[0047] S105: Acquire multispectral image data of the unknown sample under the mechanistic characteristic band, input the multispectral image data into the nonlinear quantitative detection model, and output the true absolute concentration value of DON of the sample under test.

[0048] In this embodiment, in a laboratory-level high-throughput automated detection platform or high-precision identification terminal, a dedicated multispectral imaging sensor equipped with N specific mechanism wavelength narrowband filters is used to perform instantaneous multispectral imaging of the unknown wheat grains to be tested, and to collect multispectral image data of the sample under multiple mechanism characteristic bands.

[0049] The acquired multispectral image data first undergoes reflectance normalization correction and is then standardized to a data format that meets the input requirements of PP-DDESAN, such as normalization to a 128×128 pixel range, with N specific mechanistic bands. The standardized multispectral data of unknown samples is then directly input into the trained end-to-end nonlinear quantitative detection model. Without any manual intervention, the model automatically performs complex calculations such as depth derivative feature enhancement, spatial-spectral heterogeneity focusing, and nonlinear ceiling elimination. The deep derivative enhancement module extracts first- and second-order derivative features from the input multispectral image data to enhance the weak absorption and overlapping spectral peak information caused by microscopic tissue cavitation and biochemical depletion, outputting derivative-enhanced spectral features. The spatial attention submodule weights the tissue cavitation regions characterized by SEM to determine the spatial location of the contaminated areas. The spectral attention submodule then weights the derivative-enhanced spectral features by wavelength dimension to highlight key spectral information related to biochemical depletion. Subsequently, element-wise weighted fusion is used to form a spatial-spectral joint feature. A regression prediction module performs a nonlinear mapping on the spatial-spectral joint feature, outputting a predicted value after Box-Cox transformation. The predicted value after Box-Cox transformation is then restored to the true concentration ratio, outputting the true absolute DON concentration value of the sample. The calculation formula is as follows: in, The final output is the predicted concentration of DON toxin in the wheat sample to be tested; It is the inverse Box-Cox transform function. The DON prediction value after Box-Cox transformation is directly output by the model network. These are the optimized Box-Cox transform parameters.

[0050] The entire calculation process is controlled at the millisecond level, and external actuators such as airflow sorting valves are used to achieve high-throughput and complete removal of substandard grains, thus completing automated, high-throughput, and precise quality screening.

[0051] To comprehensively verify the effectiveness, scientific validity, and technological advantages of the proposed method, this embodiment uses a validation set sample to quantitatively compare the PP-DDESAN method with existing conventional techniques such as the PLSR linear model based on full-wavelength data and models based on conventional input preprocessing and conventional deep network algorithms. The experiments focus on verifying performance by addressing three major technical challenges: overcoming high-concentration saturation, resolving spatial heterogeneity, and maintaining accuracy under extreme data compression.

[0052] Regarding the effectiveness verification in overcoming nonlinear prediction saturation, this section focuses on verifying the effectiveness of the built-in ABC-Loss loss function in eliminating the prediction plateau effect caused by the depletion of biological substrates at the mathematical level. In the prediction scatter plot of the conventional PLSR model, it was observed that when the actual DON concentration was low, such as less than 2000 μg / kg, the predicted values ​​showed a good linear correlation with the actual values. However, as Fusarium graminearum infection deepened, when the actual DON concentration entered the extremely high accumulation zone, such as extreme infection samples greater than 8000 μg / kg, the prediction scatter plots significantly deviated from the diagonal and showed a horizontal flattening trend, i.e., the nonlinear prediction saturation ceiling. This objectively confirms that excessive fungal infection leads to the depletion of starch and protein substrates in wheat through enzymatic hydrolysis, resulting in a sharp reduction in the nonlinear correlation between the macroscopic reflectance spectral response and the actual DON concentration, rendering the conventional linear spectral response model ineffective.

[0053] This application benefits from the pioneering ABC-Loss loss function that embeds the model gradient, eliminating the need for pre-processing the network with Box-Cox targeted normalization. Based on the derivative characteristics of the input spectrum, ABC-Loss adaptively adjusts the transformation parameters during backpropagation, reconstructing the substrate depletion dynamics from the mathematical foundation of the loss function. It automatically and significantly stretches the prediction error penalty gradient for extremely high-toxicity depleted samples along the gradient descent path. Results show that the PP-DDESAN model closely follows the 1:1 target line across the entire concentration range of 0–15000 μg / kg, completely breaking through the prediction failure ceiling of linear models facing depletion limits and achieving accurate quantitative inversion across the entire range at extreme concentrations.

[0054] Regarding the effectiveness verification of toxin spatial heterogeneity compensation in single-grain processing, this section focuses on verifying the effectiveness of the physically-guided spatial-spectral dual-attention mechanism in addressing the heterogeneity of toxin distribution within a single grain. In contrast, purely data-driven deep learning models lack physical guidance in their network design and rely heavily on conventional global average pooling to obtain feature representations of the entire grain. Validation results show that when facing heterogeneous samples with infection confined to localized lesions within a single grain, such as the embryo or unilateral areas, the purely data-driven model significantly underperforms due to the dilution of the strong toxin signal in the lesion area.

[0055] This application's network model guides the allocation of spatial attention weights by embedding a physical prior hard constraint—that tissue cavitation lesions caused by SEM analysis lead to a surge in Mie scattering. The network autonomously performs pathological dynamic focusing through forward propagation, dynamically identifying and assigning extremely high spatial weights to the core regions of lesions exhibiting scattering abnormalities in the seed image. Results show that the model accurately focuses on toxin accumulation areas, effectively suppresses the interference of background signals from healthy areas on the model's quantification, thoroughly compensates for spatial heterogeneity, and achieves high-precision quantification for focal samples.

[0056] In terms of performance verification of extreme data compression and detection robustness, only N characteristic bands established based on SR-FTIR mechanism are retained, and the PP-DDESAN model maintains the performance of model accuracy and robustness.

[0057] In terms of technical performance comparison, the original hyperspectral data contains 256 full bands. Comparative models often introduce standard normal variable transformation and static second derivatives as external preprocessing to reduce heterogeneity. However, the PP-DDESAN model deployed in this application requires no external static preprocessing; its built-in differentiable derivative layer and Dual-Attention branch directly enhance dynamic, fractional derivative features. With only N mechanistic feature bands as end-to-end input, its validation set determination coefficient reached 0.87, and its relative analysis error reached 2.82.

[0058] Regarding the verification of physical layer mechanisms, the N mechanism characteristic bands established by this application based on cross-scale tracing are not redundant information, but rather a complete set of high-concentration biochemical and physical fingerprints covering wheat fungal diseases from physical cavitation to the disintegration of hydrogen bonds in core nutrient macromolecules.

[0059] In terms of laboratory-level high-throughput sorting and quality assessment, based on the aforementioned superior data compression ratio, prediction robustness, and technical advantages in overcoming prediction distortion and spatial heterogeneity, the PP-DDESAN method described in this application can be seamlessly transformed from a laboratory hyperspectral research platform into the core algorithm of high-throughput automated detection and sorting equipment. By abandoning the expensive and computationally intensive pushbroom hyperspectral camera, and directly employing a multispectral imaging sensor array equipped with N types of center wavelength narrowband filters, combined with a high-speed industrial microprocessor, it can meet the needs of millisecond-level and whole-grain high-throughput screening and precise automated removal of wheat mycotoxin contamination on grain storage and processing lines. This breaks through the cost bottleneck and algorithmic blind spots in the transition from single-grain identification to industrial high-throughput sorting of mycotoxins, and has the potential for industrial-grade application transformation and the value of establishing academic barriers.

[0060] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0061] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A multispectral detection method for wheat DON toxin based on cross-scale biochemical mechanisms and physical prior deep attention networks, characterized in that, include: Acquire wheat grain samples covering different fungal infection gradients and their corresponding shortwave infrared hyperspectral imaging full-band spectral data; The wheat grain samples were characterized by microscopic morphology using scanning electron microscopy to obtain structural cavitation features reflecting tissue degradation. At the same time, the wheat grain samples were subjected to in-situ chemical spatial imaging using synchrotron radiation Fourier transform infrared microscopy to obtain biochemical depletion features of the samples. Using the structural cavitation features and the sample biochemical depletion features as prior information of the mechanism, multiple mechanism feature bands related to DON content are extracted by jointly analyzing the original average spectrum and derivative spectrum of short-wave infrared hyperspectral imaging. Based on the multispectral image data corresponding to the characteristic bands of the mechanism, a physical prior-driven deep derivative enhanced spectral attention network model is constructed, and the model is trained to obtain the DON nonlinear quantitative detection model. The multispectral image data of the unknown wheat sample to be tested under the mechanistic characteristic band is input into the DON nonlinear quantitative detection model, and the absolute concentration value of DON in the sample to be tested is output.

2. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 1, characterized in that, The wheat grain samples were characterized using scanning electron microscopy to obtain structural cavitation features reflecting tissue degradation, including: Representative grains were selected from wheat grain samples covering different fungal infection gradients. After glutaraldehyde fixation, gradient ethanol dehydration, critical point drying and vacuum gold sputtering pretreatment, the grains were placed under a scanning electron microscope for high-resolution imaging observation to obtain SEM microscopic images. The SEM images were subjected to in-depth analysis to extract morphological features reflecting the rupture of endosperm cell walls, collapse of starch matrix structure, and cavitation of internal tissues in wheat grains caused by fungal infection, and these morphological features were used as structural cavitation features.

3. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 1, characterized in that, The method of using synchrotron radiation Fourier transform infrared microscopy to perform in-situ chemical spatial imaging of the wheat grain sample to obtain the biochemical depletion characteristics of the sample includes: The absolute concentration of DON toxin in the wheat grain samples was quantitatively analyzed by liquid chromatography-tandem mass spectrometry. Healthy wheat grain samples and those with different toxin levels, quantified by liquid chromatography-tandem mass spectrometry, were embedded, sliced ​​to a predetermined thickness, and then laid flat on a high-transmittance substrate. On the window slats; Using synchrotron radiation Fourier transform infrared microscopy for surface scanning imaging, and relying on the high brightness light source of synchrotron radiation, we can accurately analyze the spatial distribution evolution characteristics of the core nutrient components of starch, lipids and proteins during fungal infection. Identify the breakage and enzymatic depletion states of key biochemical fingerprints, including starch COC and OH bonds, protein amide bands, and lipid CH bonds.

4. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 1, characterized in that, Using the structural cavitation characteristics and the sample biochemical depletion characteristics as prior information of the mechanism, multiple mechanism characteristic bands related to DON content are extracted by jointly analyzing the raw average spectrum and derivative spectrum of short-wave infrared hyperspectral imaging, including: The average spectra of the regions of interest for samples with different fungal infection gradients were extracted to obtain the original average spectra of each gradient sample. The first and second derivative spectra are calculated based on the original average spectra of each gradient sample. The original average spectra, first derivative spectra, and second derivative spectra are superimposed and compared to identify the spectral characteristic peaks and distortion regions that change regularly with increasing DON pollution concentration. Using the structural cavitation features and the sample biochemical depletion features as prior information of the mechanism, and based on the theory of overtone and combination of molecular spectra, candidate feature wavelength ranges associated with starch, protein and lipid depletion were initially identified in the SWIR region. Based on the sensitivity analysis results of derivative spectra changing with pollution gradient, the candidate bands were cross-validated to eliminate non-specific bands affected by moisture or background noise, and finally, several mechanism characteristic bands related to DON toxin content were determined.

5. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 4, characterized in that, The multiple mechanistic characteristic bands include 994 nm, 1190 nm, 1208 nm, 1426 nm, 1432 nm and 1487 nm; The 994 nm band corresponds to the second harmonic of the OH bond in wheat starch; The absorption characteristics of CH bonds in proteins and lipids at wavelengths of 1190 nm and 1208 nm; The wavelengths 1426 nm, 1432 nm, and 1487 nm correspond to the first overtones of OH bonds or NH bonds in proteins in tissues infected by fungi.

6. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 1, characterized in that, The physics-prior-driven deep derivative-enhanced spectral attention network model includes an input module, the output of which is connected to the input of a deep derivative enhancement module, the output of which is connected to the input of a spatial-spectral dual attention fusion module, and the output of which is connected to a regression prediction module. The spatial-spectral dual attention fusion module includes a spatial attention submodule and a spectral attention submodule.

7. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 1, characterized in that, The loss function used when training the physics-prior-driven deep derivative-enhanced spectral attention network model is: in, For the total loss, For regression error loss, For the derivative consistency constraint loss, For adaptive Box-Cox loss, , These are the weighting coefficients. For batch size, This represents the actual DON concentration. To predict concentration, This is the bias value. For the first i Box-Cox transform parameters for dynamic prediction of individual samples For parameter prediction subnetwork, used for output The network layer or module, For input features.

8. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 1, characterized in that, The multispectral image data of the unknown wheat sample to be tested under the mechanistic characteristic band is input into the DON nonlinear quantitative detection model, and the absolute DON concentration value of the sample to be tested is output, including: A multispectral imaging sensor equipped with narrowband filters for the multiple mechanism feature bands was used to perform multispectral imaging on the unknown wheat grains to be tested, and to collect multispectral image data of the unknown wheat sample under the mechanism feature bands. After performing reflectance normalization preprocessing on the acquired multispectral image data, it is input end-to-end into the DON nonlinear quantitative detection model. The first and second derivative features of the input multispectral image data are extracted through the depth derivative enhancement module to enhance the weak absorption and overlapping spectral peak information caused by micro-tissue cavitation and biochemical depletion, and output derivative-enhanced spectral features. The spatial attention submodule enhances the weight of the tissue cavitation region characterized by SEM to determine the spatial location of the contaminated area, and the spectral attention submodule weights the wavelength dimension of the derivative-enhanced spectral features to highlight key spectral information related to biochemical depletion. Subsequently, spatial-spectral joint features are formed through element-wise weighted fusion; The spatial-spectral joint features are nonlinearly mapped by the regression prediction module, and the predicted value after Box-Cox transformation is output. The predicted values ​​after Box-Cox transformation are restored to the true concentration ratio, and the absolute concentration value of DON in the sample to be tested is output.

9. The wheat DON toxin multispectral detection method based on cross-scale biochemical mechanisms and physical prior deep attention networks according to claim 8, characterized in that, The formula for calculating the absolute concentration of DON in the sample after restoring the predicted value to the true concentration ratio after Box-Cox transformation is as follows: in, The final output is the predicted concentration of DON toxin in the wheat sample to be tested; It is the inverse Box-Cox transform function. The DON prediction value after Box-Cox transformation is directly output by the model network. These are the optimized Box-Cox transform parameters.

Images