Abstract
Studies of karst landforms, such as sinkholes, have advanced in recent years due to technological advances and the rise of automated detection methods. Such methods offer notable time-saving strengths, but their limitations in complex terrains are frequently overlooked in modern literature. To compare these automated methods to traditional manual methods, we selected four karst massifs in the Cantabrian Mountains (Spain) and two in the Tasman Mountains (New Zealand) as study areas. High-resolution LiDAR-derived Digital Elevation Models were used to identify and map sinkholes. Automated detection was performed using a Geographic Information System and included: i) a depression-filling method; ii) a sky-view factor method; and iii) different deep learning models trained with manually mapped sinkhole data and a leave-one-region-out approach. The manual method used photointerpretation to draw the curvature of a sinkhole and field validation confirmed this method correctly identified 97.8% of sinkholes. Results in sinkhole detection and delineation varied substantially between methods. Deep learning models—HRNet in particular—outperformed other methods, achieving a mean F1 score of 0.74 and an Intersection over Union of 0.59, capturing 63.9% of sinkholes with minimal morphometric distortion. Furthermore, performance was unequally distributed, declining sharply in areas with steep or forested terrain. These discrepancies highlight the influence of local characteristics on the performance of automated methods. Their widespread adoption and their perception as a gold standard have led to a decrease in the use of traditional methods, occasionally at the risk of compromising accuracy and scientific rigor.
Rights Information
DOI
https://doi.org/10.5038/9781967518012.1007
Figure 1
ExamplesFigure.jpg (5817 kB)
Figure 2
SlopeLithoLand.jpg (24032 kB)
Figure 3
Fig1_Accuracy_Final.png (244 kB)
Figure 4
Fig2_Identification_Final.png (1705 kB)
Figure 5
Fig3_Morphometric_Final.png (4251 kB)
Figure 6
Comparing Manual and Automated Methods in Sinkhole Delineation: A Cantabrian Mountains (Spain) and Tasman Mountains (New Zealand) Case Study
Studies of karst landforms, such as sinkholes, have advanced in recent years due to technological advances and the rise of automated detection methods. Such methods offer notable time-saving strengths, but their limitations in complex terrains are frequently overlooked in modern literature. To compare these automated methods to traditional manual methods, we selected four karst massifs in the Cantabrian Mountains (Spain) and two in the Tasman Mountains (New Zealand) as study areas. High-resolution LiDAR-derived Digital Elevation Models were used to identify and map sinkholes. Automated detection was performed using a Geographic Information System and included: i) a depression-filling method; ii) a sky-view factor method; and iii) different deep learning models trained with manually mapped sinkhole data and a leave-one-region-out approach. The manual method used photointerpretation to draw the curvature of a sinkhole and field validation confirmed this method correctly identified 97.8% of sinkholes. Results in sinkhole detection and delineation varied substantially between methods. Deep learning models—HRNet in particular—outperformed other methods, achieving a mean F1 score of 0.74 and an Intersection over Union of 0.59, capturing 63.9% of sinkholes with minimal morphometric distortion. Furthermore, performance was unequally distributed, declining sharply in areas with steep or forested terrain. These discrepancies highlight the influence of local characteristics on the performance of automated methods. Their widespread adoption and their perception as a gold standard have led to a decrease in the use of traditional methods, occasionally at the risk of compromising accuracy and scientific rigor.