Water quality monitoring is essential for managing aquaculture sustainably. Changes in water quality can lead to disease outbreaks and financial losses. Traditional methods of sampling are slow and reactive. The existing machine learning methods, however, are in general very challenging for capturing the interdependent features of various parameters in different environmental data. In this paper, a novel water quality classification model with transformer is proposed by integrating learned positional encoding and multi-criteria feature selection for automatic water quality classification. Our approach demonstrates the effectiveness of learning positional encoding for feature interaction modeling in tabular environmental data through strengthening the expressiveness of feature representations. Evaluation on water quality samples shows a classification accuracy of 98.37% across three categories: Excellent, Good, and Poor. The use of learned positional encoding leads to a 12.14% improvement compared to a baseline transformer with no positional encoding. The proposed model demonstrates high accuracy and low computational cost, which is fit to be widely applied in aquaculture surveillance systems.