In today's world, satellite images are being utilized for the identification of built-up area, urban planning, disaster management, insurance & tax assessment in an area, and many other social-economic activities. The extraction of the accurate building footprints in densely populated urban areas from medium resolution satellite images is still a challenging task which requires the development of the new methods to solve such problem. In this paper, a novel Dilated-ResUnet deep learning architecture for building extraction from Sentinel-2 satellite images has been proposed. The proposed model has been tested on three novel building datasets that are prepared for three densely populated cities of India (viz. Delhi, Hyderabad and Bengaluru) using Sentinel-2 satellite images and Planet OSM. First FCC (false colour composite) dataset prepared by merging NIR, Red, Green bands, second FCC dataset prepared by merging NIR, Red, Green and Blue bands and third is TCC (true colour composite) dataset by merging red, green and blue bands. The proposed architecture is applied to both the FCC datasets and TCC dataset separately; it has been identified that the proposed model has obtained better building extraction results using FCC (NIR, Red, Green) dataset. The input satellite image enhancement and extensive experimentations to identify the optimal deep learning hyper-parameters using FCC spatial dataset have also been carried out to further improve the performance of the proposed model. The results of the experimentations reveal that the proposed model has out-performed the state of the art models available in literature by achieving the F1-score of 0.4718 and Mean IoU of 0.582 for building extraction from Sentinel-2 satellite images. The outcome of the research work can be utilized for urban planning and management, generate more ground truths for Sentinel-2 satellite images which further can be useful for other societal applications.
Read full abstract