Automatic Segmentation of Sinkholes Using a Convolutional Neural Network

Sinkholes are the most abundant surface features in karst areas worldwide. Understanding sinkhole occurrences and characteristics is critical for studying aquifers mitigating sinkhole-related hazards. Most sinkholes appear on land as depressions or cover collapses commonly mapped from elevation data, such digital models (DEMs). Existing methods identifying DEMs often require two steps: locating separating non-sinkhole depressions. In this study, we explored deep learning to directly identify DEM data aerial imagery. A key contribution of our study an evaluation various ways integrating these types raster data. We used image segmentation model, U-Net, locate sinkholes. trained separate U-Net based four input images data: a image, slope gradient DEM-shaded relief image. Three normalization techniques (Global, Gaussian, Instance) were applied improve model performance. Model results suggest that viable method particular, provided best U-net The using with Gaussian achieved performance intersection-over-union (IoU) 45.38% unseen test set. Aerial images, however, not useful training IoUs below 3%.