Compared with conventional migration method, least-squares migration has a greater advantage, but the large amount of calculation limits its popularization and application. Phase encoding which combines multi-shot data into a super-gather can effectively improve the computational efficiency, and suppress crosstalk noise. However there are a quite few coding strategies, and very few comparison and analysis among them are carried out. Based on least-squares reverse time migration method, we have conducted a detailed comparison of four kinds of commonly used encoding schemes(amplitude encoding, polarity encoding, random time delay encoding, and plane wave encoding) by numerical tests. Besides, we extend the stochastic optimization idea to multi-source least squares reverse-time migration and propose an optimized multi-source least- square reverse time migration. Our algorithm considers the random characteristic of the gradient, and the gradient at current iteration is the exponential decay weighted average of the previous iteration. Through the weighted average of the previous gradient, it reduces the random fluctuations of the gradient. Theoretical model processing confirms that convergence of the proposed optimized multi-source least-square reverse time migration is faster than conventional phase-encoding least-square reverse-time migration method. Therefore, it can save computational cost and improve computational efficiency.