This paper explores an underwater localization system aimed at addressing the challenge of accurately locating a target node in an inhomogeneous underwater environment. In this system, multiple receivers with known positions collaborate to enhance the localization process. The presence of the stratification effect causes underwater acoustic signals to propagate along a curved path, introducing non-Gaussian noise and leading to localization errors. Traditional localization algorithms, such as the weighted least squares algorithm, are designed for ideal underwater environment with low-Gaussian noise, assuming the signal is transmitted in a straight line. However, these algorithms are limited in their applicability. To overcome this limitation, we put forth a deep learning-based underwater range difference correction network (DL-URDCN). This innovative approach effectively removes the non-Gaussian noise in range difference, converting it into low-Gaussian noise. Furthermore, an iterative weighted least squares method based on Taylor series expansion (TIWLS) algorithm for the time-difference-of-arrival (TDOA) is utilized to solve a localization problem. Simulation results demonstrate that the DL-URDCN-TIWLS algorithm outperforms existing benchmarks in terms of localization performance and satisfies the Cramér Rao Lower Bound (CRLB).