Abstract: To achieve rapid and non-destructive detection of aflatoxin B₁ (AFB₁) content in maize, this study collected near-infrared hyperspectral data and corresponding AFB₁ chemical values from 337 groups of moldy maize samples. The Bootstrap resampling method was employed for data augmentation on the original spectral data. Ten preprocessing methods were applied to preprocess the raw spectra, and a depthwise separable convolution XCeption net (DESCXCeptionNet) model, improved based on the XCeption architecture, was constructed for quantitative analysis of AFB₁ in moldy maize. This model efficiently extracts deep spectral features through hierarchically expanded depthwise separable convolution modules and a multi-scale feature fusion mechanism, establishing an intrinsic relationship between spectral information and the measured chemical values to achieve quantitative analysis of AFB₁ content in moldy maize. Meanwhile, five comparative models were used for analysis: Partial Least Square, Support Vector Regression, Extreme Learning Machine, Back Propagation neural network, and Convolutional Neural Network. The results showed that for quantitative analysis of AFB₁ in moldy maize, the combination of Bootstrap data augmentation and second derivative (D₂) preprocessing with the DESCXCeptionNet model achieved the highest accuracy. Its root mean square error of prediction (RMSEP), coefficient of determination of prediction (R²_p), and relative percent deviation (RPD) were 2.942 2 μg/kg, 0.979 0, and 6.899 9, respectively. Compared with the PLS, SVR, ELM, BP, and CNN models, the RMSEP values decreased from 17.779 7, 15.365 9, 10.323 1, 14.432 7, and 7.939 9 μg/kg to 2.942 2 μg/kg, with prediction accuracy improved by 83.43%, 80.85%, 71.51%, 79.62%, and 62.94%, respectively. In summary, near-infrared hyperspectral technology combined with the DESCXCeptionNet model has good feasibility for quantitative analysis of AFB₁ in moldy maize, effectively enhancing the model's prediction accuracy and robustness. This study provides data support for achieving rapid, non-destructive, and accurate detection of AFB₁ content in moldy maize. To meet the demand for rapid, non-destructive detection of aflatoxin B1 (AFB₁) in maize while overcoming prominent drawbacks of conventional detection approaches including sample damage, procedural complexity, and high costs, this study proposes a novel Depth-enhanced Separable Convolution XCeption Network (DESCXCeptionNet) by integrating near-infrared hyperspectral imaging (NIR-HSI) technology on the basis of an improved XCeption neural network architecture. A controlled artificial mold incubation experiment generated 337 maize seed samples representing diverse AFB₁ contamination levels. Hyperspectral signals ranging from 1049.16 nm to 1653.68 nm were acquired for each sample, while fluorescence-based immunochromatography simultaneously quantified the actual AFB₁ content, constructing a dataset that correlates spectral signatures and AFB₁ concentrations. To address the limitations of insufficient samples and severe spectral spatial heterogeneity, Bootstrap resampling was adopted to augment the original dataset. Comparative experiments with multiplicative stochastic scaling (MS) and conditional generative adversarial network (CGAN) methods further verified its superior data augmentation performance. The augmented data was partitioned into calibration and prediction sets using the Kennard-Stone algorithm. Ten established spectral preprocessing methods, including first derivative (D₁), second derivative (D₂) preprocessing, Fourier transform (FT), moving average (MA), mean centering (MC), multiplicative scatter correction (MSC), savitzky-golay smoothing (S-G), wavelet transform (WT), standard normal variate (SNV) and smoothing spline (SS), were systematically evaluated for noise reduction and feature enhancement, assessing their comparative impact on model optimization. The DESCXCeptionNet model consisting of four core modules, namely the input flow, middle flow was further constructed to conduct depth expansion and channel expansion of spectral data for quantitatively analysis of AFB₁ in moldy maize. Leveraging hierarchically expanded depthwise separable convolution blocks and a multi-scale feature fusion mechanism, this model efficiently extracts deep spectral features and establishes an intrinsic relationship between spectral information and the measured chemical values, thereby realizing accurate regression prediction of AFB₁concentrations in moldy corn. Meanwhile, five comparative models were used for analysis: Partial Least Square (PLS), Support Vector Regression (SVR), Extreme Learning Machine (ELM), Back Propagation neural network (BP), and Convolutional Neural Network (CNN). The results indicated that for quantitative analysis of AFB₁ in moldy maize, the combination of Bootstrap data augmentation and second derivative (D₂) preprocessing maximized feature enhancement and baseline noise suppression. Consequently, the DESCXCeptionNet model achieved optimal detection performance, with the root mean square error of prediction (RMSEP), coefficient of determination of prediction (R²p), and relative percent deviation (RPD) were 2.942 2 μg/kg, 0.979 0, and 6.899 9, respectively, significantly surpassing all benchmark methods. Compared with the PLS, SVR, ELM, BP, and CNN models, the RMSEP values decreased from 17.779 7, 15.365 9, 10.323 1, 14.432 7, and 7.939 9 μg/kg to 2.942 2 μg/kg, with prediction accuracy improved by 83.43%, 80.85%, 71.51%, 79.62%, and 62.94%, respectively. In summary, near-infrared hyperspectral technology combined with the DESCXCeptionNet model has good feasibility for quantitative analysis of AFB₁ in moldy maize, effectively enhancing the model's prediction accuracy and robustness. This study provides data support for achieving rapid, non-destructive, and accurate detection of AFB₁ content in moldy maize.