A depth-based land-use semantic extraction method for landscape images

Integrated aerial and ground-based observations are essential for rapid land-use classification in modern land surveys, requiring efficient analysis of large volumes of landscape images. These images often contain background noise that impedes classification accuracy and exhibit spatial compression of the main scene, which causes semantic spatial distortion as depth increases. To address these challenges, this paper proposes a depth-based land-use semantic extraction (DLSE) method that effectively reduces background noise and corrects spatial distortions. The DLSE method follows three main steps: (1) depth estimation per pixel using a multi-resolution neural network to delineate the main scene range, filtering semantic noise; (2) conversion of perspective projection to orthographic projection per pixel to correct spatial distortion; and (3) improved land-use semantic extraction through MobileViT-enhanced feature extraction in SegNet. Experimental results demonstrate that DLSE achieves a 96.84 % accuracy with more detailed outputs and operates at 23 fps, positioning it as an efficient tool for automated land surveys and land-use decision-making. • Projection Transformation: Converts perspective projection to orthographic projection, reducing spatial distortion in land-use semantics. • Enhanced Segmentation: Integrates MobileViT into SegNet, optimizing feature extraction for better efficiency and accuracy in segmentation. • Accurate Depth Estimation: Uses multi-resolution deep neural networks for precise pixel depth estimation, filtering background noise. • High Performance Metrics: Achieves 96.84 % accuracy in land-use semantic extraction with a processing speed of 23 fps on mobile devices.