Position-aware graph-CNN fusion network: an integrated approach combining geospatial information and graph attention network for multi-class change detection

• Published in: IEEE transactions on geoscience and remote sensing p. 1
• Main Authors: Wang, Moyang, Li, Xiang, Tan, Kun, Mango, Joseph, Pan, Chen, Zhang, Di
• Format: Journal Article
• Language: English
• Published: IEEE 04.01.2024
• Subjects: Artificial intelligence; Convolutional neural networks; Deep learning; Geospatial analysis; Geospatial artificial intelligence; graph attention network; multi-class change detection; position information encoding; Remote sensing; Semantics; Task analysis; urban changes
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3350573

Urban change detection is crucial for informed decision-making but faces various challenges, including complex features, rapid changes, and extensive human interventions. These challenges underscore the urgent need for innovative multi-class change detection (MCD) techniques that extensively incorporate deep learning. Despite several successes achieved with the deep learning based MCD methods, still certain shortcomings persist, including the disregard for spatial principles, which significantly hinders the seamless integration of geoscience-knowledge and artificial-intelligence. In this paper, a novel deep learning model known as the Position-aware Graph-CNN Fusion Network (PGCFN) is introduced, integrating spatial position encoding to effectively detect urban changes. The model's first part encodes geospatial positions following Tobler's first law of geography. It then integrates encoded positions into a multi-class change detection model, combining a graph attention network with a convolutional neural network to enhance performance. The model was tested on 0.5-meter resolution remote sensing images, achieving an impressive minimum Mean Intersection over Union (MIoU) score of 91.20%. Additionally, the model's position-aware graph attention module exhibited a strong emphasis on geographic-proximity when evaluating connections between superpixels. Overall, these findings affirm that our model could effectively addresses urban change detection challenges and significantly enhances the integration of geoscience knowledge and artificial intelligence.