Sucrose content is an important indicator to measure the quality of sugarcane. It is of great significance to study the non-destructive detection method of sucrose content in living sugarcane based on the principle of spectroscopy. Sugarcane has a cylindrical shape with a hard skin and waxy surface. Different spectra detection angles and surface conditions will affect the modeling results to a certain extent. In addition, dimension reduction of characteristic wavelengths extraction is another factor that affects the model’s accuracy. In this study, the effect of different spectral measurement styles on the accuracy of the sucrose content prediction model was evaluated, an improved method for characteristic wavelengths extraction was proposed, and a sucrose prediction model was eventually constructed. A spectra acquisition platform was designed to obtain the transmittance spectra of sugarcane stalks. When acquiring the transmittance spectra, there were three different acquisition angles (120°, 150° and 180°) between the incident light and the measurement probe and two surface conditions (wax-not-removed and wax-removed).Six data sets of 123 samples were obtained in total. Firstly, PLS modeling result was used to evaluate the effect of different spectral pretreatment methods, including S-G smoothing, standard normal variation (SNV), multiplicative scatter correction (MSC), first derivation (FD), etc., and the result showed that SNV had the best comprehensive performance and was selected for further study.Then the effect of different measurement styles on the modeling of sucrose content was evaluated.The result found that: (1) regarding the effect of wax coverage,the spectral transmittance after wax-removed was high, the spectral difference among different collection sites of a single sample was lower, and the correlation with sucrose was much higher; (2) regarding the effect of the spectra acquisition angle, the transmittance decreased as the angle increased in a certain range; (3) the best modeling result was obtained (R_p=0.790 6, RMSEP=0.898 6) with the measurement style of wax-removed and 120° measurement angle. Finally, the interval partial least squares method (i-PLS), genetic algorithm (GA), ant colony algorithm (ACO) and an improved ant colony algorithm (VRC-ACO) based on full wavelengths PLS modeling variable regression coefficient proposed in this study were used to extract the characteristic wavelengths. The result showed that the number of characteristic wavelengths selected by the VRC-ACO algorithm, which had only ten wavelengths,was the least, yet the prediction accuracy was the best (R_p=0.861 6, 9.0% higher than the full-band model; RMSEP=0.746 6, 20.0% lower than the full-band model). This research provides theoretical support for the non-destructive detection of sugarcane and the development of corresponding sensors.