Precise building semantic segmentation in remote sensing images via MR-DeepLabv3+ network.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing image segmentation faces challenges with incomplete contours, blurred boundaries, and misclassification of small buildings.
• Existing models often struggle to capture multi-scale features and maintain robustness against noise in complex urban landscapes.

Purpose of the Study:

• To propose and evaluate a novel deep learning network, MR-DeepLabv3+, for enhanced building segmentation in remote sensing imagery.
• To improve accuracy and robustness in segmenting diverse building types, including small and slender structures, across various datasets.

Main Methods:

• Integration of MixConv with dataset-adapted multi-scale convolutional kernels (3×3, 5×5, 7×7) for superior multi-scale feature extraction.
• Implementation of a segmentation-optimized R-Drop Loss function with decoder-level channel-wise masking and dynamic KL divergence for noise robustness.
• Experimental validation on three distinct building datasets: Self-building, WHU, and Massachusetts, comparing against baseline DeepLabv3+ and transformer models.

Main Results:

• MR-DeepLabv3+ achieved high performance metrics: Accuracy (Acc), Mean Intersection over Union (MIoU), and Frequency Weighted Intersection over Union (FWIoU) across all tested datasets.
• Demonstrated superior performance over baseline DeepLabv3+ and four recent transformer models on building segmentation tasks.
• The proposed network exhibits a favorable balance between model compactness and inference efficiency, suitable for resource-constrained environments.

Conclusions:

• MR-DeepLabv3+ effectively enhances building segmentation accuracy in remote sensing images, significantly mitigating issues with small building detection and contour definition.
• The method offers practical value for applications like UAV-based mapping, providing reliable and efficient building segmentation solutions.
• The network's robustness and efficiency make it a strong candidate for real-world remote sensing image analysis.