The fibre content of sugarcane is a non-negligible factor in the breeding process of sugarcane and the production of sugar, paper and other industries. It is of great significance to detect the fibre content of living sugarcane non-destructively by using visible-near infrared spectroscopy in transmission form. One hundred and twenty-three sugarcane samples of six varieties at different growth stages were collected and divided into a calibration set (82 samples) and prediction set (41 samples) in the ratio of 2∶1 using the Duplex sample set division method. The transmission spectra of sugarcane cane stem in the original state and dewaxing state were acquired at a measurement angle of 120°, and the band from 670 to 950 nm with less noise and obvious amplitude fluctuations was chosen as the actual modeling band. The waveforms were observed to find a significant increase in transmittance after dewaxing, and a PLS (Partial least squares) regression model was established to analyze the effect of wax coverage on the predictive model ability. The sugarcane samples were modeled more effectively after dewaxing. The 9 preprocessing methods, including first derivation (FD), continuous wavelet transform (CWT), and standard normal transform (SNV), are divided into four steps: baseline correction, scattering correction, smoothing, and scale scaling. The order of the steps was permuted to produce 108 combined preprocessing methods, and the PLS modeling analysis was performed for the spectra after each combined preprocessing separately, finally, the preprocessing method FD+SG with the best comprehensive modeling effect was acquired. To screen for the wavelengths carrying the most effective information, effective variables screening algorithms such as uninformative variable elimination (UVE), genetic algorithm (GA), competitive adaptive reweighted sampling (CARS), and random frog algorithm (RF) were taken to select important wavelengths for the transmission spectra after optimal pretreatment. The important wavelengths extracted by each algorithm were analyzed for PLS modeling separately, in which the important wavelengths extracted by the UVE method were modeled the best, with the number of selected wavelengths being 40, accounting for 14.3% of the full band. The R2p is 0.73, which improves 14.1% over the full-band modeling results with the same preprocessing method, and the RMSEP is 0.88, which decreases 14.6% over the full-band modeling results. The results showed that the transmittance visible-NIR spectrum could effectively predict the fibre content of living sugarcane. This study can provide a theoretical basis for developing corresponding portable sensors, and provide technical support for sugarcane breeding and production efficiency in various industries.