<p>Accurate and rapid acquisition of soil organic carbon (SOC) content is important for field fertilization and management in tobacco fields. Spectral reflectance were extracted from unmanned aerial vehicle (UAV) remote sensing multispectral images and spectral indices were calculated. Red-edge bands were introduced to modify original indices. Modified spectral indices dominated by near-infrared and red-edge bands were formed. Three analysis approaches including gray correlation, dispersion and multicollinearity were used for variables determination. Four machine learning (ML) algorithms, including random forest (RF), support vector machine (SVM), back propagation neural networks (BPNN) and extreme gradient boosting (XGBoost), were employed to construct prediction models. Modified spectral indices have better correlation with SOC content. Models based on combination of three indices as variables have highest accuracy of all. The coefficient of determination (R²) and root mean square error (RMSE) were used to comprehensively evaluate the model accuracy. The best prediction models for tobacco fields a and b were constructed using RF algorithm (with R² of 0.913 and 0.907, with RMSE of 0.504 and 0.525, respectively). In contrast, the best prediction models for tobacco fields c and d were constructed using XGBoost algorithm (with R² of 0.923 and 0.915, with RMSE of 0.559 and 0.561, respectively). Through model validation, the optimal models hold the highest prediction accuracy with R² of 0.923. The optimal models were used to obtain the spatial distribution of SOC. Using modified spectral indices is an effective approach for predicting SOC from UAV images.</p>