Satellite image segmentation and pansharpening using multi-task learning /
Andrew Emel Nessem Kelada Khalel
Satellite image segmentation and pansharpening using multi-task learning / تقسيم صورة القمر الصناعى وتحريرها باستخدام تعلم المهام المتعددة Andrew Emel Nessem Kelada Khalel ; Supervised Magda Bahaa Eldin Fayek , Amir Fouad Surial Atiya , Motaz Elsaban - Cairo : Andrew Emel Nessem Kelada Khalel , 2020 - 75 P. : charts , facimiles , maps ; 30cm
Thesis (M.Sc.) - Cairo University - Faculty of Engineering - Department of Computer Engineering
Due to technical limitations, satellite imagery sensors capture two separate images simultaneously : high resolution panchromatic image PAN and low resolution multispectral image MS. Normally, these two images are fused together to obtain images having rich spectral information as well as high spatial resolution. This fusion process is called pansharpening. Images produced from pansharpening suffer loss of information that existed in original images.Our contribution is twofold. First, we propose a new loss function for better panshaprening. Second, we introduce a multi-task framework that takes MS and PAN images as inputs and generate high resolution multispectral images together with densely labeled maps. Results show that the proposed loss function yields better optimization compared to other loss functions from the literature. Additionally, learning pansharpening and dense labeling tasks jointly is shown to exhibit better performance than each individual task. We also show that our approach outperforms the existing approaches in the literature
Deep learning Satellite imagery Segmentation
Satellite image segmentation and pansharpening using multi-task learning / تقسيم صورة القمر الصناعى وتحريرها باستخدام تعلم المهام المتعددة Andrew Emel Nessem Kelada Khalel ; Supervised Magda Bahaa Eldin Fayek , Amir Fouad Surial Atiya , Motaz Elsaban - Cairo : Andrew Emel Nessem Kelada Khalel , 2020 - 75 P. : charts , facimiles , maps ; 30cm
Thesis (M.Sc.) - Cairo University - Faculty of Engineering - Department of Computer Engineering
Due to technical limitations, satellite imagery sensors capture two separate images simultaneously : high resolution panchromatic image PAN and low resolution multispectral image MS. Normally, these two images are fused together to obtain images having rich spectral information as well as high spatial resolution. This fusion process is called pansharpening. Images produced from pansharpening suffer loss of information that existed in original images.Our contribution is twofold. First, we propose a new loss function for better panshaprening. Second, we introduce a multi-task framework that takes MS and PAN images as inputs and generate high resolution multispectral images together with densely labeled maps. Results show that the proposed loss function yields better optimization compared to other loss functions from the literature. Additionally, learning pansharpening and dense labeling tasks jointly is shown to exhibit better performance than each individual task. We also show that our approach outperforms the existing approaches in the literature
Deep learning Satellite imagery Segmentation