Detection of karst depression in Brazil comparing different semantic and instance segmentations and global digital elevation models

This research assessed the efficacy of deep segmentation in identifying and measuring natural karst depressions in the Bambuí Group's carbonate rocks in Western Bahia, Brazil. The investigation used five Global Digital Elevation Models (DEMs) with 30-m resolution: Advanced Spaceborne Thermal Emission and Reflection Radiometer - Global Digital Elevation Model version 3 (ASTER-GDEM v3), Advanced Land Observing Satellite World 3D – 30 m version 3.2 (AW3D30 v3.2), Copernicus 30 m global DEM (GLO-30), NASA Digital Elevation Model version 1 (NASADEM v1), and Shuttle Radar Topography Mission version 3 (SRTM v3). The karst feature detection analysis compared five semantic segmentation architectures: Feature Pyramid Networks (FPN), LinkNet, UNet, UNet++, DVL3+, using an EfficientNet-B7 backbone, one instance segmentation model (Mask-RCNN), and a semantic-to-instance conversion method. This comparison considers different datasets, including one, two, or eleven variables (DEM, DEM-based sink depth, and nine terrain attributes). Besides, we performed a combinatorial analysis of the DEMs to evaluate the improvement in karst feature detection. The methodology involved the generation of geomorphometric attributes, sample labeling from Sentinel-2 and Operational Land Imager-Landsat 8 datasets, training-validation-testing steps with 128 × 128 samples, reconstruction of large images for semantic and instance segmentation, and accuracy analysis. The findings revealed that GLO-30 and AW3D30 data were the most accurate, while ASTER GDEM performed poorly in both segmentation forms. Among the semantic models, FPN showed the highest accuracy. The 11-variable models preferentially outperformed those with fewer in both types of segmentation. The approach of semantic-to-instance conversion with Geographic Information System tools favored individualizing karst depressions and obtaining efficient quantification, consisting of an easy alternative to achieve instance segmentation. Reconstruction of large remote sensing images through sliding windows considering the specifics of instance and semantic segmentation demonstrated that smaller stride enhances object coverage and reduces the risk of losing crucial information, effectively improving the predictions. The combinatorial analysis for semantic and instance segmentation indicates that models incorporating more DEMs (4- and 5-DEM models) and variables achieve generally higher accuracy. The best result in semantic segmentation was combining the five DEM datasets using 11 variables, reaching an F-score of 85.06 and IoU of 74.00. In instance segmentation, the best result for the bounding box was the model that integrates AW3D30, GLO-30, NASADEM, and SRTM with eleven variables reaching Average Precision (AP) of 45.64, AP50 of 88.40 and AP75 of 39.38, while the best result for the segmentation mask was the model that integrates the five models with eleven variables with AP of 42.61, AP50 of 86.79 and AP75 of 35.43. These results demonstrate that integrating a wide range of DEM data, even the lowest performing ones, improves model generalization and accuracy in segmentation, leveraging strengths and mitigating individual weaknesses. Future work could explore high-resolution DEMs and the integration of various deep-learning methods.