This paper proposed a solution method of spectral classification modeling for multi-scene optimization. Firstly, a modified power function equation was constructed through a simulation analysis of the Mie scattering of particles, and the direct fitting method was used to achieve accurate turbidity correction of the sample spectra. Then, the absorbance normalization method was employed to obtain the linear feature spectra of different scenes and develop a scene-based feature library. Subsequently, the partial least-squares (PLS) method was applied to build a solution model for each scene and thereby establish a chemical oxygen demand (COD) solution model library. When an unknown water sample went through the COD detection, its normalized spectrum was first matched with the linear feature spectra of the scene-based library through the Jaccard similarity theory for the identification of the scene it belonged to. Then, its COD concentration was calculated with the optimal solution parameters obtained from the solution library. The experimental results show that the proposed method holds application value in that it delivers high scene-matching accuracy and reduces the COD solution error under multi-scene conditions.