Segmenting ancient cemeteries under forests using synthesized LiDAR-derived data and deep convolutional neural network

Abstract This study applied a semantic segmentation model based on convolutional neural networks and LiDAR-derived data to segment an ancient cemetery in a forested area. We proposed to synthesize multiple LiDAR-derived data into three-channel and five-channel data and perform data augmentation. Moreover, the channel attention (CA) mechanism was used to improve the Unet and TransUNet models. The results indicated that it has higher precision using five-channel raster data synthesized with elevation (DEM), slope, hillshade, roughness, and curvature than one or three derived data synthesized raster data in the test dataset. For the Unet model, the intersection over union (IoU), precision, and recall reached 0.885, 0.921, and 0.924, respectively, for the TransUNet model, the IoU, precision, and recall reached 0.901, 0.921, and 0.944, respectively, successfully segmenting the unknown region cemetery. In addition, the migration of the model also indicated that the model trained by synthesizing data has better portability.