Soil organic matter (SOM) is a key indicator for assessing soil fertility and ecosystem carbon cycle. In this study, measured soil data and remote sensing data were combined, multiple model integration of SOM inversion was compared, high-precision, large-area, and long-time-series inversion of surface SOM in the Hexi Corridor from 1990 to 2024 was achieved. The results showed that among the indicator screening methods, the Max-Relevance and Min-Redundancy (mRMR) method was superior to the Pearson Correlation Coefficient (PCC) and Gray Relational Analysis (GRA) methods. Among the inversion models, the accuracy of Random Forest (RF) was better than Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), BP neural Network (BPNN), and Multiple Linear Regression (MLR) models. The mRMR-RF model integration provided the optimal solution for regional SOM inversion. The multi-year average SOM content was 10.57 g/kg in the Hexi Corridor. SOM was low in the west and north, and high in the east, south, and around the oasis areas. Among the main land types, forestland had the highest SOM at 17.80 g/kg and unused land had the lowest at 9.43 g/kg. SOM generally showed an increasing trend before 2000 and a decreasing trend after 2000. The shift in SOM was concentrated between slightly low and moderate grades, while the shift between the high and low grades was relatively weak. This study provides a high-precision remote sensing inversion framework for SOM monitoring in the arid zone, and provides a scientific basis for sustainable land management and ecological restoration in the region.