Centre de documentation
scientifique

Research on automatic identification method of terraces on the Loess plateau based on deep transfer learning / Mingge Yu in Remote sensing, vol 14 n° 10 (May-2 2022)  

Public [article]  inRemote sensing > vol 14 n° 10 (May-2 2022) . - n° 2446 Titre : Research on automatic identification method of terraces on the Loess plateau based on deep transfer learning Type de document : Article/Communication Auteurs : Mingge Yu, Auteur ; Xiaoping Rui, Auteur ; Weiyi Xie, Auteur ; et al., Auteur Année de publication : 2022 Article en page(s) : n° 2446 Note générale : bibliographie Langues : Anglais (eng) Descripteur : [Vedettes matières IGN] Traitement d'image optique [Termes IGN] apprentissage profond [Termes IGN] Chine [Termes IGN] classification par réseau neuronal convolutif [Termes IGN] détection automatique [Termes IGN] échantillonnage [Termes IGN] extraction de traits caractéristiques [Termes IGN] image à haute résolution [Termes IGN] image panchromatique [Termes IGN] image Worldview [Termes IGN] modèle de simulation [Termes IGN] surface cultivée [Termes IGN] terrasse Résumé : (auteur) Rapid, accurate extraction of terraces from high-resolution images is of great significance for promoting the application of remote-sensing information in soil and water conservation planning and monitoring. To solve the problem of how deep learning requires a large number of labeled samples to achieve good accuracy, this article proposes an automatic identification method for terraces that can obtain high precision through small sample datasets. Firstly, a terrace identification source model adapted to multiple data sources is trained based on the WorldView-1 dataset. The model can be migrated to other types of images for terracing extraction as a pre-trained model. Secondly, to solve the small sample problem, a deep transfer learning method for accurate pixel-level extraction of high-resolution remote-sensing image terraces is proposed. Finally, to solve the problem of insufficient boundary information and splicing traces during prediction, a strategy of ignoring edges is proposed, and a prediction model is constructed to further improve the accuracy of terrace identification. In this paper, three regions outside the sample area are randomly selected, and the OA, F1 score, and MIoU averages reach 93.12%, 91.40%, and 89.90%, respectively. The experimental results show that this method, based on deep transfer learning, can accurately extract terraced field surfaces and segment terraced field boundaries. Numéro de notice : A2022-402 Affiliation des auteurs : non IGN Thématique : IMAGERIE Nature : Article DOI : 10.3390/rs14102446 Date de publication en ligne : 19/05/2022 En ligne : https://doi.org/10.3390/rs14102446 Format de la ressource électronique : URL article Permalink : https://documentation.ensg.eu/index.php?lvl=notice_display&id=100705 [article]