Wangheng Xiao

ABSTRACT Optical remote sensing often struggles to distinguish landslides from spectrally similar backgrounds, such as bare rocks. We address this limitation by modifying the YOLOv8-seg network to accept a Digital Elevation Model (DEM) as a fourth input channel. We process the DEM data using a patch-wise Min-Max normalization and modify the network’s initial convolutional layer to ingest this 4-channel tensor, utilizing a hybrid weight initialization to retain pre-trained RGB features.