Abstract

Robust semantic segmentation of VHR remote sensing images from UAV sensors iscritical for earth observation, land use, land cover or mapping applications.Several factors such as shadows, weather disruption and camera shakes makingthis problem highly challenging, especially only using RGB images. In thispaper, we propose the use of multi-modality data including NIR, RGB and DSM toincrease robustness of segmentation in blurred or partially damaged VHR remotesensing images. By proposing a cascaded dense encoder-decoder network and theSELayer based fusion and assembling techniques, the proposed RobustDenseNetachieves steady performance when the image quality is decreasing, compared withthe state-of-the-art semantic segmentation model.

 

Quick Read (beta)