Lutein is a natural antioxidant that has numerous benefits for human health. Heterotrophic Chlorella sorokiniana has the advantage of high purity and production of lutein. In contrast, the production of lutein in Chlorella sorokiniana mainly depends on two factors: biomass productivity and lutein content. However, conventional approaches such as the optical density method for measuring biomass productivity and high-performance liquid chromatography for measuring lutein content suffer from drawbacks, including complex procedures and limited timeliness. A visible near-infrared dual-mode snapshot multispectral imaging detection system was constructed to rapidly and non-destructively determine the variations in lutein production during the growth process of Chlorella sorokiniana. Based on the spectral response range, the visible camera was used to obtain the spectral information image of lutein content, and the near-infrared camera was used to obtain the spectral information image of biomass productivity to build a visible near-infrared dual mode multispectral dataset containing biomass productivity and lutein content information. To address the issue of wide spectral range and limited wavelengths in the snapshot multispectral camera used in the system, a novel approach combining sequential floating forward selection with a modified successive projections algorithm (mSPA) was proposed. A comparative study was conducted, evaluating mSPA against successive projections algorithm, genetic algorithm, and random frog algorithm for wavelength selection. Multiple linear regression and extreme learning machine models were constructed based on the selected feature wavelengths. Finally, the optimal predictive models for biomass productivity and lutein content were used to generate a visualization distribution map of lutein production in Chlorella sorokiniana. The results indicated that when using near-infrared and visible cameras for biomass productivity and lutein detection in Chlorella sorokiniana, the mSPA algorithm consistently yielded fewer feature wavelengths for both biomass productivity and lutein and achieved the highest prediction accuracy. The optimal models of biomass productivity and lutein content were established using the mSPA-selected feature wavelengths in combination with an extreme learning machine. The corresponding coefficients of determination for the prediction sets were 0.947 for biomass productivity and 0.907 for lutein, with root mean square errors of 0.698 g·L-1 and 0.077 mg·g-1 and residual prediction deviations of 3.535 and 3.338, respectively. The models demonstrated good predictive capabilities. The visualization distribution successfully achieved intuitive monitoring of lutein production variations in Chlorella sorokiniana, which is beneficial for online detection of lutein content in practical production scenarios. The mSPA algorithm, employed in the snapshot multispectral detection of biomass productivity and lutein content in Chlorella sorokiniana, effectively avoided the incorrect selection and omission of feature wavelengths by evaluating each sorted wavelength individually, thereby improving the prediction accuracy of the models. This approach provides a new wavelength selection strategy for applying snapshot multispectral imaging technology.