Answers are grounded in this paper and cite their sources.
Computational study demonstrates superior segmentation accuracy in remote sensing imagery, highlighting enhanced boundary delineation through probability-guided edge enhancement.
Semantic segmentation in remote sensing images (RSI) assigns unique semantic labels to each pixel and plays a crucial role in real-world applications such as environmental change monitoring, precision agriculture, and economic assessment. Although convolutional neural networks (CNN) and Transformer-based models for semantic segmentation of RSI have achieved remarkable success, existing approaches still struggle to accurately detect weak edges and occluded objects due to the complexity and fuzziness of edges in RSI. To overcome this obstacle, we propose a novel probability-guided edge enhancement network (PEEN) for semantic segmentation of RSI, which is the first attempt to leverage the probability function to guide the segmentation model in performing edge prediction for RSI. Specifically, in the feature extraction stage of PEEN, we present a convolutional self-attention mechanism to enhance the global feature representation of the encoder-decoder network. In the edge enhancement stage of PEEN, we innovatively build an iterative probability-guided edge prediction module to refine edge prediction mathematically and iteratively. With the cooperation of the mentioned two stages, the proposed model yields precise segmentation of the objects and edge portions in RSI. Experiment results and analysis demonstrate that the PEEN model outperforms the existing popular CNN-based and Transformer-based models in semantic segmentation with 85.54% and 88.35% of Mean Intersection Over Union (mIOU) on the Vaihingen and Potsdam test datasets. Our code is available at https://github.com/Zyk517/PEEN.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: