The recurrent crystallization and dissolution of salts within murals lead to significant internal structural damage, ultimately causing paint loss and compromising mural integrity. This study explores the effectiveness of five training dataset partitioning methods in improving the accuracy of models designed to non-invasively predict salt content in murals using spectral data collected across 350 nm–2500 nm. Firstly, spectra were acquired from laboratory-simulated murals using a spectroradiometer, followed by smoothing and denoising via the Savitzky-Golay (S-G) algorithm. To further enhance salt-related spectral features and eliminate baseline drift, both first-order and second-order differentiation techniques were applied. Secondly the performance of five partitioning strategies—Random Selection (RS), Kennard-Stone (KS), Sample Set Partitioning Based on Joint X-Y Distance (SPXY), Kernel Distance-Based Sample Set Partitioning Based on Joint X-Y Distance (KSPXY), and Sample Set Partitioning Based on Joint X-Y-E Distances (SPXYE)—was evaluated. A salt content inversion model was then developed using Random Forest (RF) and Partial Least Squares Regression (PLSR). Results showed that PLSR, combined with KSPXY partitioning and first-order derivative enhancement, achieved the best predictive performance. Validation of the model with a test dataset yielded the RMSE of 0.068 and the R² of 0.954, indicating high accuracy. Our findings underscore the pivotal role of sample partitioning method selection in enhancing model accuracy and predictive outcomes. This study provided an effective technique for the inversion of mural salt content non-invasively, which would facilitate the preservation of these invaluable cultural artifacts.