Abstract: Under the global context of environmental and climate change, soil salinization has become a critical issue threatening ecological stability and sustainable development. The Yellow River Delta region exhibits prominent soil salinization problems. Compared to arid areas, this region features complex ecological heterogeneity, where traditional two-dimensional feature spaces—limited to coupling only two environmental variables—show inherent limitations in characterizing such complex environments. Furthermore, conventional research often restricts itself to a few specific indices, failing to systematically construct and evaluate a comprehensive index pool to identify the optimal parameters for representing salinization, which limits the full extraction of salinization information. In contrast, three-dimensional feature spaces can more effectively utilize multi-band remote sensing data and integrate multiple environmental variables, thereby possessing stronger monitoring capabilities. However, studies focusing on salinization monitoring in the Yellow River Delta by constructing three-dimensional feature spaces based on spectral indices, combined with feature selection and Bayesian-optimized XGBoost models, remain relatively scarce. To address this, this study employed Landsat 9 satellite imagery to build a spectral index pool, extracting a total of 35 spectral indices, including vegetation indices, salinity indices, water indices, and other relevant indices. To fully enhance modeling efficiency and parameter screening effectiveness, a Bayesian-optimized XGBoost model was utilized to evaluate and select features based on the built-in Gain metric, retaining the top two most important indices from each category. Using these selected representative indices, multiple three-dimensional feature space models were constructed through cross-category combinations. Within these three-dimensional spaces, the three coordinate axes respectively represent different index types, and any point ( x, y, z) in the feature space corresponds to the values of three indices for a specific pixel in the remote sensing image. Simultaneously, multiple two-dimensional feature spaces were built using cross-category combinations of the single most important index from each category. By comparing accuracy evaluation metrics with field-measured data, the optimal three-dimensional and two-dimensional feature space models for soil salinization inversion in the Yellow River Delta were determined, and regional salinization spatial analysis was subsequently conducted. The results demonstrate that: 1) The Bayesian-optimized XGBoost model effectively screened the most relevant indices. Salinity indices achieved the highest modeling accuracy ( R ² = 0.921, RMSE = 0.964 g/kg, RPIQ = 8.422), with the salinity index 7 SI7 showing the highest feature importance (0.341). Ultimately, eight of the most informative feature indices were selected. 2) Compared to two-dimensional feature spaces, three-dimensional feature spaces more fully exploit spectral information. The optimal three-dimensional model showed improvements of 0.059 in R ² and 1.191 in RPIQ, and a reduction of 0.069 g/kg in RMSE, confirming that the three-dimensional approach enables high-precision prediction of soil salinization. 3) Among the constructed three-dimensional feature space models, the model based on SI8-Albedo-WI achieved the highest accuracy ( R ² = 0.922, RMSE = 0.863 g/kg, RPIQ = 7.645, Kappa coefficient = 86%), whereas the ERVI-WI-Albedo model performed the worst ( R ² = 0.519, RMSE = 3.464 g/kg, RPIQ = 1.087). 4) In the Yellow River Delta region, moderately salinized areas account for the largest proportion (29.7%), primarily distributed in the central-western part of Kenli District and Lijin County; severely salinized areas constitute the smallest proportion (9.8%), mainly located in the eastern part of Kenli District. The findings of this study provide crucial references and decision-making support for the prevention and remediation of soil salinization in the Yellow River Delta.