A neural network architecture that processes irregularly-sampled spatial data by masking missing values during convolution operations.