Cotton foreign matter (FM) harms fiber quality as it may damage cotton fiber during ginning processing or cause flaws in finished textiles. Therefore, detecting and classifying foreign matter are important in the cotton production process and quality assessment. The mulching film is a unique impurity in machine-harvested seed cotton in China. Since the mulching film is commonly used to grow cotton in Xinjiang, the remaining fragments are mixed into cotton during mechanical harvesting. This study placed 12 types of common cotton foreign matter, including mulching film fragments, between two lint layers. A push-broom-based hyperspectral imaging system was used to acquire images of the mixed foreign matter and lint samples in transmittance mode at the spectral range of 400~1 000 nm. The hyperspectral transmittance images were first corrected using flat-field correction and cropped due to noise at the edges. The images at 500 nm were chosen for manual region-of-interest (ROI) selection. Mean transmittance spectra were extracted from the ROIs and normalized across all samples. Canonical discriminant analysis (CDA) and the first three canonical variables were used to group foreign matter and lint, and multivariate analysis of variance (MANOVA) was employed to evaluate the differences between each combination of two types of foreign matter using the first three canonical variables. Then, the interval Random Frog (iRF) method was used to extract 12 feature wavelengths. A support vector machine (SVM) classifier was used to classify the transmittance spectra of full and selected wavelengths respectively, and the accuracies were compared and analyzed. The results show that the average classification accuracy of all types of foreign matter and lint at the full wavelength was 84.4%. The method in this paper was feasible for classifying foreign matter in the inner layer of cotton, including plastic packaging, paper, and mulching film. After extracting the feature wavelengths, the classification accuracy of 4 types of foreign matter with similar appearance and similar chemical composition (broken stem, hull, bark, brown leaf) was lower, but all exceeded 73%. The classification accuracy of seed meat, green leaf, paper, plastic package, mulching film, and lint was over 90%. The average classification accuracy of all foreign matter and lint types was 86.2%. Compared with the classification results of the full-wavelength, the average classification accuracy of the selected wavelength was improved by 1.8%.The results of this study can provide a theoretical basis for the research on the detection of foreign matter in the inner layer of cotton and have a guiding role for the application of hyperspectral transmittance imaging technology.