Abstract: ‍ ‍The reconfigurable intelligent surface （RIS） is considered a pivotal technology for future wireless communication and target perception due to its characteristics of enhancing spectrum and energy efficiency while being cost-effective. RIS integrates a multitude of low-cost passive reflecting elements on a planar surface， and by connecting them to an intelligent controller， it allows for the control of the phase and amplitude of signals incident on these reconfigurable elements， thereby reconfiguring the propagation of incident signals. Due to its ability to control the phase and magnitude of reflected signals， RIS can be utilized to enhance the performance of wireless communication and target sensing. The accurate estimation of the direction of arrival is a pivotal element in achieving target sensing objectives. RIS can substantially enhance the precision of DOA estimation through various strategies. One approach involves focusing radar beams in specific directions to amplify target signals while concurrently reducing interference. Additionally， RIS can be employed to adjust reflection paths， mitigating the adverse effects of multipath propagation. Traditional DOA estimation methods rely on phase difference information between antennas in scenarios with multiple receiving antennas. However， these methods have limitations. Conventional beamforming methods and similar beamforming techniques are associated with low resolution. Subspace-based methods like Multiple Signal Classification and maximum likelihood-based approaches like Maximum Likelihood parameter estimation methods， suffer from high computational complexity and susceptibility to environmental influences. With the continuous development of fields such as wireless communication and radar， researchers have gradually explored and combined deep learning to overcome the limitations of traditional DOA estimation methods. Nevertheless， practical implementation of RIS faces challenges such as mutual coupling effects arising from the proximity of elements being less than half a wavelength， as well as reflection mismatches （errors in reflection phase and amplitude） resulting from imperfect control of the reflection process， which significantly impact the performance of RIS. In this paper， we address the two-dimensional DOA estimation problem in RIS-based systems by establishing an RIS system model that takes into account mutual coupling effects and reflection mismatches. Based on this model， we propose a novel method for two-dimensional angle estimation in DOA estimation. This method initially reconstructs the signals received by RIS using a deep neural network to mitigate the effects of mutual coupling and reflection mismatches. Subsequently， it employs a nonlinear least squares technique for high-precision DOA estimation. Simulation results are presented to validate the algorithm’s estimation performance， it has been demonstrated that the DNN reconstruction step effectively mitigates the impact of mutual coupling and reflection mismatches， leading to a significant enhancement in the algorithm’s DOA estimation performance. Applying the Nonlinear Least Squares algorithm to estimate the reconstructed signals yields even more precise angle estimations. Therefore， in RIS systems affected by mutual coupling and reflection mismatches， the proposed method outperforms traditional approaches such as Orthogonal Matching Pursuit and Fast Fourier Transformation in terms of estimation performance.