Soil salinity is a critical concern for sustainable agriculture, particularly in arid and semi-arid regions where saline water irrigation is often employed to meet crop water demands. While remote sensing has long been explored for estimating soil salinity, significant spatial autocorrelation in the data raises concerns about the reliability of model performance. In this study, a novel approach is proposed to predict soil salinity using aerial multispectral data designed to account for spatial autocorrelation in the datasets based on Bayesian Neural Network (BNN). The study also evaluated the advantages of the red-edge (REG) band and spectral indices over spectral bands in enhancing predictive skill. Aerial multispectral data and in situ soil observations were used to establish relationships between spectral features and soil salinity, estimated using electrical conductivity observations. Initially, four BNN models were developed using unfiltered multispectral features to assess the contributions of the REG band and spectral indices. Subsequently, four additional BNN models were constructed using spatially filtered covariates, processed via the Moran Eigenvector Filtering (MEF) technique, to evaluate the impact of spatial autocorrelation on model performance. The results showed that incorporating the REG band improved model efficiency by 10.77 % during training and 10.27 % during testing when included as a feature in the bands-only (BO) model. Additionally, using REG-derived spectral indices in the indices-only (IO) model led to improvements of 17.13 % during training and 19.44 % during testing. Using spectral indices in place of bands has also led to improvement of model efficiencies by up to 20.35 % during testing. Moreover, the use of decomposed covariates derived through MEF resulted in a substantial improvement in BNN model performance, with a 94.61 % increase in model efficiency and a 32.89 % reduction in estimation error. The best model for predicting salinity was the one that incorporated spatially filtered spectral indices, including those derived from the REG. The study also demonstrated the importance of static land characteristics in the prediction of soil salinity by incorporating the spatially structured component of soil salinity. This study highlights that failure to account for spatial autocorrelation can compromise both the interpretability of model features and the accuracy of predictions, leading to potentially misleading conclusions in spatially structured data.