Text this: Semantic Segmentation of Aerial Imagery Using U-Net with Self-Attention and Separable Convolutions