Remote Sensing Ground Object Segmentation Algorithm Based on Edge Optimization and Attention Fusion.

Considering the characteristics of remote sensing land cover images with a wide variety of types and complex object edges as well as the limited receptive field of local convolutions in existing segmentation networks resulting in inadequate utilization of contextual information, leading to issues such as blurred object edges and low segmentation accuracy, this paper proposes a remote sensing land cover segmentation algorithm based on the UNet3+ network architecture. Firstly, during the decoding process, a similarity-aware point affiliation operator is introduced as an upsampling method. This operator aggregates multiple proposals from the feature pyramid to enhance the segmentation capability for object boundary details. Secondly, in the encoding process, a selective kernel module is introduced to optimize the downsampling approach. This module enables neurons to achieve an adaptive receptive field size, facilitating the acquisition of multi-scale information from target features and precise capture of valuable detailed semantic information. Finally, in the skip-connection phase, a dual multi-scale attention module is added to perform weighted fusion of features from different scales, enabling the model to better focus on both local details and global contextual information. Experimental results on the WHDLD and ISPRS Potsdam datasets demonstrate that the proposed algorithm achieves mean intersection over union (MIoU) improvements of 64.4% and 75.4% respectively, compared to baseline models, the improvement is about 2.6 percentage points and 3.2 percentage points respectively. This also validates the effectiveness of the proposed algorithm in addressing the issue of blurry segmentation edges. [ABSTRACT FROM AUTHOR]