In complex observation environments, the number of available GNSS satellites is significantly reduced, and the observed signals are susceptible to multidimensional outliers, resulting in challenges in outlier detection and a marked deterioration in positioning performance. To address this issue, this paper proposes a robust filtering method for quasi-accurate detection based on adaptive classification using Random Forests. Initially, we construct a Random Forest classifier utilizing multidimensional data characteristics derived from dynamic observational data collected in urban settings. This classifier facilitates the adaptive categorization of observational data into two principal classes: quasi-accurate observations and non-quasi-accurate observations. Following this classification, a novel robust positioning method for GNSS quasi-accurate detection is developed. This method leverages the quasi-accurate scores obtained during the classification process to determine the random model for quasi-accurate solutions, substituting the standardized true errors for conventional least-squares residuals in the subsequent IGG3 robustness test. Taking pseudorange differential positioning as an example, we evaluate the effectiveness of the proposed method across two positioning scenarios characterized by different magnitudes of observational errors. Experimental results demonstrate that, in comparison to traditional IGG3 and quasi-accurate detection methods, the proposed approach significantly enhances positioning performance, with accuracy improvements ranging from 16% to 51% across various testing scenarios. The random forest classification model proposed in this paper effectively improves the identification accuracy of quasiaccurate observations, and therefore can effectively improve the positioning performance in complex observation environments.