Optimized building extraction from high-resolution satellite imagery using deep learning

Building extraction is very essential in various urban dynamics like disaster management and change detection, finding the estimated population, and so on. Building extraction from satellite data is a challenging task as the images may be subjected to different illumination or structure due to very large variations of the appearance of buildings which may correspond to the different area/terrain. Although satellite imagery is readily available from various sources, translating the imagery includes intensive effort. Many computer-vision tasks have been carried out successfully but understanding the impact of them on building extraction with remote sensing imagery is a growing need.To overcome this kind of problem, an algorithm is proposed which extends the convolutional neural network for pixel-wise classification of images. Furthermore, to resolve the problem of extraction and masking of images, Mask-RCNN (i.e., Mask Region-based Convolutional Neural Network) algorithm is used which makes this process easier and more efficient.The model is trained on a complex dataset that is significantly larger. Also, to make this algorithm more scalable, an advanced image augmentation technique is used in the pre-processing step.The results show that the algorithm achieves better performance in terms of accuracy.

Automated summary

Building extraction is crucial in urban dynamics for disaster management, change detection, and population estimation. However, extracting buildings from satellite data is challenging due to variations in illumination and structure. To overcome this, a convolutional neural network and Mask-RCNN algorithm with advanced image augmentation technique are proposed. The results show improved accuracy in building extraction.