A dataset and deep learning framework for building object segmentation from high resolution Cartosat-2S satellite images

Abstract

Building object extraction plays a crucial role in the monitoring and strategic planning of urban infrastructures. With the rapid advancement of deep learning (DL), segmentation tasks are becoming more effective, even though models trained on datasets from different regions may not perform well in Indian areas due to their complex, unstructured cityscapes. Therefore, this paper introduces a novel building object segmentation dataset, namely, NITK Computer Vision and Deep Learning (NCVDL) Cartosat-2S building dataset, which covers three diverse regions of India. Moreover, this paper presents a lightweight DL model, namely, the KAN Building Detection Network (KANBDNet), for extracting buildings from satellite images, aiming to reduce the number of parameters without compromising performance using Kolmogorov–Arnold Networks, and the training is guided by a boundary-focused mIoU metric. It is mainly structured using three modules: a split-wise feature extraction module to efficiently extract features by splitting them into two and then process separately, a dense multiscale feature extraction block to extract and densely connect features at various scales, and a skip layer which effectively combines the encoder stage and high level features to enhance the features passing through the decoder. The performance of the proposed KANBDNet and state-of-the-art (SOTA) DL-based building segmentation models is evaluated using the new NCVDL Cartosat-2S dataset and the publicly available GF-7 building dataset. The proposed KANBDNet outperforms all SOTA models, achieving ( F1-score, kappa coefficient, jaccard index, dice coefficient) of (87.75%, 75.50%, 78.87%, 80.69%) and (92.10%, 84.19%, 85.83%, 86.62%) on the NCVDL Cartosat-2S and GF-7 building datasets, respectively.