Attribute reduction is a commonly used technique in machine learning and data mining. The traditional granular neighborhood rough set is sensitive to noise data and fails to distinguish the representativeness differences of samples within granules, which easily leads to the interference of noise samples on the overall representation of granules. To address this limitation, a kind of distance weighting method between samples and granular-ball centers is proposed to define a weighting function. This weighting function is set as a weight to measure the similarity between samples and granular balls, thereby defining a granular-ball fuzzy neighborhood. Subsequently, a membership function is defined based on the granular-ball fuzzy neighborhood to construct more effective fuzzy approximation operators. A variable-precision granular-ball rough set with improved fuzzy neighborhood is proposed, which enhances the robustness of the model to noisy data and its classification performance. Experimental tests conducted on 11 UCI (University of California, Irvine)benchmark datasets show that the proposed model can effectively reduce attributes while maintaining high classification accuracy. Finally, the proposed algorithm is applied to the practical problem of water quality evaluation and classification. By removing redundant features and selecting core indicators, the accuracy of classification is effectively improved.