Weakly Supervised Semantic Segmentation in Aerial Imagery via Explicit Pixel-Level Constraints