The acquisition of farmland soil salt information based on UAV remote sensing technology provides a rapid, accurate and reliable theoretical basis for salinization management. The soil salt content of 0～20cm from the sampling point was collected on the test ground of Shahao canal irrigation field in Hetao Irrigation District, Inner Mongolia, and the images were collected by M600 hexarotor UAV platform equipped with Micro-MCA multispectral camera. Otsu algorithm was used to classify the multi-spectral images (soil background and vegetation canopy). Based on the classification results, the spectral index and image texture features before and after removing the soil background were extracted respectively. The soil salt content monitoring model was constructed by support vector machine (SVM) and extreme learning machine (ELM). The four modeling strategies were as follows: spectral index of the soil background was not removed (strategy 1); spectral index of the soil background was removed (strategy 2); spectral index of the soil background was not removed + image texture features (strategy 3); spectral index of the soil background was removed + image texture features (strategy 4). The optimal variable combination was selected by comparing the model accuracy of the four modeling strategies. The results showed that the inversion accuracy of soil salt content calculated by strategy 3 and strategy 4 was higher than that of strategy 1 and strategy 2, and their validation sets R2v were 0.614, 0.640, 0.657 and 0.681, respectively. Therefore, it was of great significance to use image texture feature and vegetation index to improve the inversion accuracy of soil salt content. By comparing strategies 3 and 4, the image texture feature + vegetation index was affected by soil background. The accuracy of the strategy 4 was lower than that of the strategy 3, whose R2v was 0.614 and 0.657, respectively. The optimal model for each variable processing was ELM model, and the modeling sets R2c were 0.625, 0.644, 0.618, 0.683, and the standard root mean square errors were 0.152, 0.134, 0.206 and 0.155, respectively. Compared with the SVM model, the ELM model improved the inversion accuracy of soil salt content.