Centre de documentation
scientifique

ALERT: adversarial learning with expert regularization using Tikhonov operator for missing band reconstruction / Litu Rout in IEEE Transactions on geoscience and remote sensing, vol 58 n° 6 (June 2020)  

Public [article]  inIEEE Transactions on geoscience and remote sensing > vol 58 n° 6 (June 2020) Titre : ALERT: adversarial learning with expert regularization using Tikhonov operator for missing band reconstruction Type de document : Article/Communication Auteurs : Litu Rout, Auteur Année de publication : 2020 Note générale : bibliographie Langues : Anglais (eng) Descripteur : [Vedettes matières IGN] Traitement d'image optique [Termes IGN] apprentissage automatique [Termes IGN] apprentissage dirigé [Termes IGN] bande spectrale [Termes IGN] cohérence géométrique [Termes IGN] correction d'image [Termes IGN] dégradation d'image [Termes IGN] image Worldview [Termes IGN] pollution acoustique [Termes IGN] qualité d'image [Termes IGN] régularisation de Tychonoff Résumé : (auteur) The Earth observation using remote sensing is one of the most important technologies to assimilate key attributes about the Earth’s surface. To achieve tangible consequence, the internal building blocks of such a complex system must operate flawlessly. However, due to a dynamically changing environment, degradation in sensor electronics, and extreme weather condition remotely sensed images often miss essential information. As the sensors operate over several years in space the likelihood of sensor degradation persists. This results in commonly observed issues, such as stripe noise, missing partial data, and missing band. Various ground-based solutions have been developed to address these technological bottlenecks individually. In this article, we devise a method, which we call ALERT, to tackle missing band reconstruction. The proposed method reconstructs the missing band with the sole supervision of spectral and spatial priors. We compare the proposed framework with state-of-the-art methods and show compelling improvement both qualitatively and quantitatively. We provide both theoretical and empirical evidence of better performance by regularized adversarial learning as compared to complete supervision. Furthermore, we propose a new residual-dense-block (RDB) module to preserve geometric fidelity and assist in efficient gradient flow. We show that ALERT captures essential features such that the spatial and spectral characteristics of the reconstructed band remains preserved. To critically analyze the generalization we test the performance on two different satellite data sets: Resourcesat-2A and WorldView-2. As per our extensive experimentation, the proposed method achieves 20.72%, 13.81%, 1.05%, 15.91%, and 2.94% improvement in the root mean square error (RMSE), SAM, SSIM, PSNR, and SRE, respectively, over the state-of-the-art model. Numéro de notice : A2020-285 Affiliation des auteurs : non IGN Thématique : IMAGERIE Nature : Article nature-HAL : ArtAvecCL-RevueIntern DOI : 10.1109/TGRS.2019.2963818 Date de publication en ligne : 16/01/2020 En ligne : https://doi.org/10.1109/TGRS.2019.2963818 Format de la ressource électronique : url article Permalink : https://documentation.ensg.eu/index.php?lvl=notice_display&id=95108 [article]