STVAE: skip connection driven two-stream property fusion variational autoencoder for cross-region wastewater treatment plant semantic segmentation

• Verfasst von: Li, Yuze [VerfasserIn]
• Verfasst von: Zhang, Yan [VerfasserIn]
• Verfasst von: Randhawa, Sukanya [VerfasserIn]
• Verfasst von: Yang, Chunling [VerfasserIn]
• Verfasst von: Zipf, Alexander [VerfasserIn]
• Titel: STVAE: skip connection driven two-stream property fusion variational autoencoder for cross-region wastewater treatment plant semantic segmentation
• Verf.angabe: Yuze Li, Yan Zhang, Sukanya Randhawa, Chunling Yang, Alexander Zipf
• E-Jahr: 2025
• Jahr: June 2025
• Umfang: 21 S.
• Illustrationen: Illustrationen
• Fussnoten: Im Titel ist AE tiefgestellt ; Online verfügbar: 22. Januar 2025 ; Gesehen am 12.06.2025
• Titel Quelle: Enthalten in: Information fusion
• Ort Quelle: Amsterdam [u.a.] : Elsevier Science, 2000
• Jahr Quelle: 2025
• Band/Heft Quelle: 118(2025) vom: Juni, Artikel-ID 102960, Seite 1-21
• ISSN Quelle: 1566-2535
• DOI: doi:10.1016/j.inffus.2025.102960
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: Cross-region semantic segmentation
• Sach-SW: Domain adaptation
• Sach-SW: Property fusion
• Sach-SW: Remote sensing
• Sach-SW: Wastewater treatment plant
• K10plus-PPN: 1928103723

Abstract

Wastewater treatment plant (WWTP) plays a crucial role in achieving social sustainable development goals. Precise information on WWTPs obtained through advanced semantic segmentation technologies benefits multiple applications, including urban planning, environmental protection and public health. However, the diverse architectural styles, scales and surroundings of WWTPs across regions, influenced by climate, topography and regional economic conditions, bring challenges in generalizing segmentation algorithms. Thus, fusing the knowledge learned from different regions can form a more powerful knowledge representation. In this paper, we propose a Skip connection driven Two-stream property fusion Variational AutoEncoder (STVAE) for cross-region WWTP semantic segmentation. Our motivation is to increase the generalization capability of STVAE by capturing and fusing generative probabilistic features, inherent region properties and multi-scale properties. Specifically, STVAE captures the generative probabilistic features by constructing an attention-driven variational encoder. These features make STVAE more adaptable to the cross-domain changes, improving the segmentation robustness and performance. This attention-driven structure contributes to learning local details and limiting the effect of weak semantic information. Furthermore, a two-stream parallel decoder is considered to adapt distributions from different perspectives. The inherent region properties are introduced in this decoder to highlight the spatial consistency of the results. The unsupervised inherent region information and multi-scale features, which are extracted by this decoder, are fused through an entropy-wise mechanism. Additionally, a unique adversarial strategy is utilized to align the distributions of different domains. Based on OpenStreetMap (OSM) data and Microsoft Bing Maps Very High Resolution (VHR) satellite images, multiple experiments conducted on three tasks illustrate the effectiveness of STVAE compared with several state-of-the-art techniques qualitatively and quantitatively. STVAE effectively expands its application scope.