<p>During the propagation in the subsurface, surface multiples undergo multiple reflections and refractions, which enable them to reach areas that are difficult to access with primary reflections. When the same shot records are used, surface multiple migration can provide more extensive subsurface illumination and wider coverage. However, it also generates a large amount of cross-talk noise. Compared with conventional Reverse Time Migration (RTM), Least-Squares Reverse Time Migration (LSRTM) provides higher-resolution imaging results and more balanced amplitudes. With the same number of iterations, the first-order multiple LSRTM can produce imaging sections with higher signal-to-noise ratios. Multiple LSRTM typically relies on gradient-based algorithms and requires tens to hundreds of iterations. Traditional multiple LSRTM does not optimize the gradient, resulting in lower convergence efficiency and imaging accuracy. Moreover, additional wavefield extrapolation calculations are required in each iteration to determine the step size. This paper introduces the QHAdam optimization algorithm from deep learning into the least-squares reverse time migration of first-order surface multiples separated by a modified Surface-Related Multiple Elimination (SRME) method, which directly obtains optimized model updates at the minimum computational cost without the need for step-size calculations. Model experiments show that the first-order surface multiple LSRTM based on the modified SRME and QHAdam has higher convergence efficiency and imaging accuracy. Additionally, due to the elimination of the step-size calculation step, its computational efficiency is also improved.</p>