<p>When dealing with spatial data, not every variable is equally important. Identifying key variables can help to reduce data collection costs and computational burden. In practice, data frequently exhibit heterogeneity and heavy-tailed distributions. As an analysis method, expectile regression can effectively reveal distribution characteristics at different expectile levels. While some variable selection methods for linear spatial expectile regression model have been proposed, there are often complex nonparametric relationships between explanatory and response variables in practice. Therefore, this paper applies deep learning technology to the variable selection problem of nonparametric spatial expectile regression, achieving feature sparsity by incorporating a skip layer into the neural network architecture to investigate the distribution correlation between the explanatory variable and the response variable when there is a nonparametric relationship between them. Through numerical simulation and analysis of real dataset, we demonstrate that this method can have good prediction performance while selecting important variables.</p>