Building Footprint Semantic Segmentation using Bi-Channel Bi-Spatial (B2-CS) LinkNet

High-resolution satellite imagery provides the information about the planet's surface whose automated labelling helps in various practical applications. Segmentation of building footprint from remote sensing images is a challenging task. This paper proposes a novel architecture bi-channel bi-spatial (B-2-CS) LinkNet for implementing the semantic segmentation of building footprint. The proposed network uses two parallel semantic segmentation architectures (LinkNet) of RGB and HSV channels with dual spatial and channel attention schemes. By enhancing the features specifically in various channels and positions, result of the segmentation scheme improves. The proposed model outperforms other models with different characteristics like scales, spatial resolution, and object shapes while the analysis was done in two state-of-art datasets: INRIA aerial image labelling dataset and Massachusetts buildings dataset. Accuracy of the proposed work is 97.92% for both datasets which is better than many of the existing methods.