Wastewater treatment plants (WWTPs) play a crucial role in maintaining ecological balance and public health and are essential for advancing social sustainable development goals. However, the diverse architectural styles, scales, and environmental contexts of WWTPs—shaped by climate, topography, and regional economic conditions—pose significant challenges for generalizing segmentation algorithms. To address this, integrating knowledge from different regions can create more robust knowledge representations.
This paper introduces a Skip Connection Driven Two-stream Property Fusion Variational AutoEncoder (STV) for cross-region WWTP semantic segmentation. The goal is to enhance the generalization capability of STV by integrating generative probabilistic features, intrinsic regional properties, and multi-scale characteristics. Specifically, STV leverages an attention-driven variational encoder to capture generative probabilistic features, enabling better adaptation to cross-domain variations and improving segmentation robustness and accuracy. This attention mechanism aids in capturing local details while mitigating the impact of weak semantic information.
Additionally, the model employs a two-stream parallel decoder to address distributions from different perspectives. This decoder incorporates inherent regional properties to ensure spatial consistency in segmentation results. It further extracts unsupervised regional information and multi-scale features, which are fused using an entropy-based mechanism. To align domain distributions effectively, a novel adversarial strategy is utilized.
Experiments conducted on multiple tasks using OpenStreetMap (OSM) data and Microsoft Bing Maps Very High Resolution (VHR) satellite images demonstrate the superior performance of STV compared to several state-of-the-art methods. Both qualitative and quantitative results validate the effectiveness of STV, highlighting its capability to generalize across domains and expand its range of applications.
In the future, the method will be extended to a wider range of scenarios for multiple geospatial object segmentation and will enhance the interpretability of detailed segmentation by considering geological mechanisms.
Reference: https://www.sciencedirect.com/science/article/abs/pii/S1566253525000338
Related posts: https://heigit.org/de/geoai4water-3-thailand-2/
Figure : Segmentation results of different methods on three tasks. From left to right: target domain images, ground truths, results of U-Net, AdaptSegNet (DeepLab), Adapt- SegNet (U-Net), ScaleAware (DeepLab), ScaleAware (U-Net), MemoryAdaptNet, TSPDL, TSPDS and our STVAE in three target domain regions, i.e., (a)-(b) GE, (c)-(d) SP and (e)-(f) TK.