Abstract: Accurate and rapid determination of polyester and spandex fiber content in complex waste textile matrices is a critical prerequisite for automated textile sorting and high-value resource recycling. However, existing analytical methods often suffer from limited detection accuracy, strong dependence on sample pretreatment, model overfitting, and insufficient generalization capability when applied to heterogeneous textile blends. To overcome these limitations, this study proposes a rapid, non-destructive, and quantitative detection framework that integrates near-infrared hyperspectral imaging with optimized machine learning and deep learning algorithms for efficient waste textile identification and sorting. A push-broom hyperspectral imaging system operating in the near-infrared spectral range of 1000–1700 nm was employed to acquire high-dimensional spectral data from waste textile samples with various polyester–spandex blending ratios. To improve spectral quality and suppress interference caused by noise, baseline drift, and light scattering, the raw spectral data were preprocessed using white and dark reference correction, Savitzky-Golay (S-G) smoothing, and standard normal variate (SNV) transformation. For polyester fiber content prediction, principal component analysis (PCA) was used to extract representative spectral features and reduce data dimensionality. For spandex fiber detection, Pearson and Spearman correlation analyses were conducted to identify wavelength bands that are highly sensitive to spandex content. Feature distribution characteristics were further explored using t-distributed stochastic neighbor embedding (t-SNE) visualization. Two complementary modeling strategies were developed to address the distinct spectral characteristics of polyester and spandex fibers. First, a multilayer perceptron model optimized by the Snake Optimization algorithm (MLP-SO) was proposed for polyester content prediction. The SO algorithm enabled global optimization of the MLP hyperparameters, effectively reducing the risk of getting trapped in local optima and enhancing model robustness. Second, for spandex content prediction, a random forest (RF) model was constructed using kernel principal component analysis (KPCA)-derived nonlinear features, enabling effective modeling of complex and weak spectral responses associated with low spandex concentrations. Experimental results demonstrated that the proposed models achieved superior predictive performance compared with the conventional approaches. The MLP-SO model attained a coefficient of determination (R²) of 0.96 and a root mean square error (RMSE) of 3.55 on an independent test set, significantly outperforming traditional models such as XGBoost, generalized additive models (GAMs), and partial least squares regression (PLS-R). For spandex fiber content prediction, the RF model achieved an R² of 0.90 and an RMSE of 1.48, markedly surpassing the performance of the PLS-R model. Correlation analysis revealed that the 1000–1100 nm and 1400–1500 nm wavelength regions were highly sensitive to spandex content. Meanwhile, t-SNE visualization confirmed the clear separability of spectral features corresponding to different spandex concentration levels. Overall, this study validates the feasibility of combining hyperspectral imaging with optimized learning algorithms for the rapid, non-destructive, and precise quantification of polyester and spandex fibers in heterogeneous waste textiles. The proposed framework effectively overcomes the limitations of traditional linear and conventional machine learning models when handling high-dimensional spectral data, providing a robust and scalable technical solution to support automated textile sorting and sustainable resource recovery in the textile recycling industry.