Abstract: Efficiently obtaining soil salinity data is of great significance for the management and improvement of coastal saline-alkali soils, with the key lying in enhancing the accuracy of the results. This study takes the cultivated soil in the Yellow River Delta region as its research object, utilizing Landsat series remote sensing imagery data. It introduces the shortwave infrared band to expand the traditional band indices, and based on both the traditional band indices and the improved expanded band indices, employs MLR (multiple linear regression), RF (random forest), and SVM (support vector machine) to construct soil salinity inversion models. The accuracy of each model is compared, and the optimal model is selected. The optimal model is then used to inversely estimate and analyze the spatial-temporal changes in salinity content in the study area from 2014 to 2024.The results indicate that compared to traditional band indices, the use of extended band indices can improve inversion accuracy and yield more precise inversion results. When modeling using the aforementioned machine learning methods, the improved model accuracy shows an increase in R² (coefficient of determination) by 0.13～0.14, a decrease in RMSE (root mean square error) by 0.12～0.28, and a decrease in MAE (mean absolute error) by 0.08～0.13; When comparing the three modeling methods, SVM achieved the highest inversion accuracy, followed by RF, while MLR performed relatively poorly. The optimal model in this study was the support vector machine model based on the extended band index, with R², RMSE, and MAE values of 0.74, 1.1, and 0.7, respectively, indicating a reliable model; The soil salinization status in the study area over the past decade has shown significant fluctuations, alternating between mitigation and exacerbation. The salinization status was particularly severe in 2018 and 2021. Therefore, soil salinization management in the region still requires high priority and continuous attention

Key words: soil salinity inversion, band index, machine learning, Yellow River Delta, spatio-temporal analysis