Purpose <p>The northeast black soil region is an important area for grain production, but soil salinization and degradation seriously threaten the sustainability of agriculture. Returning decomposed straw to the field can effectively improve the fertility of saline‑alkali soil, but there are still deficiencies in the dynamic monitoring of soil quality after straw return. In this study, a hyperspectral machine learning framework was developed to dynamically monitor the soil organic matter (SOM) content in the modified soil.</p> Materials and methods <p>Firstly, 900 samples were collected from the improved saline‑alkali field planted with soybean, and the hyperspectral data of 203 bands and the organic matter content measured in the laboratory were collected simultaneously. Then a combined preprocessing strategy of multiplicative scatter correction (MSC) and standard normal variate (SNV) transformations was employed to suppress spectral interference from salinity and alkalinity. A dynamic prediction model was constructed by integrating principal component analysis (PCA) with sparrow search algorithm (SSA)‑optimized support vector regression (SVR).</p> Results and discussion <p>Results showed that the PCA-SVR model achieved the best performance, with a determination coefficient (R²) of 0.8381 and a mean squared error (MSE) of 0.1619 on the test set, representing a 19.3% improvement in R²over the partial least squares regression (PLSR) models. SOM content followed a “steady growth—temporary decline—significant accumulation” pattern during soybean development, reaching a maximum of 35.40 g/kg at maturity. A brief decline occurred during grain filling, likely due to intensified microbial decomposition.</p> Conclusions <p>This study provides high-resolution spectral data and a machine learning framework for evaluating saline-alkali soil remediation across multiple growth stages. These results offer practical guidance for improving the sustainability of saline-alkali farmland.</p>