Aims <p>This study aimed to explore the effects of increasing image texture features and removing soil background on the alfalfa salt stress diagnosis accuracy.</p> Methods <p>This study extracted spectral reflectance to construct 15 vegetation indexes, and used gray level co-occurrence matrix to calculate eight image texture features. The Canny edge detection algorithm was used to remove the soil background, and set T1 (vegetation index non-removed soil background), T2 (vegetation index + image texture features non-removed soil background), T3 (vegetation index removed soil background), T4 (vegetation index + image texture features removed soil background), as independent variables to construct salt stress diagnosis model based on the support vector regression algorithm, and determined the best salt stress diagnosis model.</p> Results <p>Compared with the T1, the modeling and validation accuracies of salt stress diagnosis model constructed based on the T2 increased by 13.39% and 13.36%, respectively, and those of salt stress diagnosis model constructed based on the T3 increased by 6.30% and 5.33%. The salt stress diagnosis accuracy constructed based on T4 was the highest, with the modeling set R², RMSE, and RPD of 0.675, 0.2143, and 1.7735, respectively, and the validation set R², RMSE, and RPD of 0.652, 0.2349, and 15,749, respectively. The modeling and validation accuracies of the salt stress diagnosis model constructed based on crop salt stress index (CSSI) reached more than 0.564 and 0.549, respectively, which can be used as a new indicator for diagnosing salt stress.</p> Conclusions <p>Both increasing image texture features and removing soil background can significantly improve the accuracy of alfalfa salt stress diagnosis.</p>