Fig 1.
UNet architecture [1].
FMs = feature maps, w = width, h = height. Arrows indicate operations while blocks indicate data: input predictor variables, intermediate feature maps, or output class logits.
Fig 2.
Residual connection [2] within a double-convolution block using an identity connection (a) and projection connection (b).
Fig 3.
Dilated convolution (DC) module implemented in geodl [27].
Fig 4.
Squeeze and excitation (SE) module [58].
Fig 5.
AG module mechanism [61].
Fig 6.
UNet implementation provided by geodl [27].
Fig 7.
Training, test, and validation areas for (a) agricultural terraces (terraceDL) dataset [29] in Iowa, USA; (b) surface coal mining valley fill faces (vfillDL) dataset [30] in southern West Virginia, eastern Kentucky, and southwestern Virginia, USA; and (c) historic mine benches (minebenchDL) dataset [31] in northern West Virginia, USA.
(d) through (i) show example terrain surfaces and associated geomorphic features as examples.
Fig 8.
Training loss for terraceDL, mineBenchDL, and vfillDL datasets using all training samples and different model configurations across 25 training epochs.
Table 1.
Overview of the available training, validation, and test chips for each dataset used.
Table 2.
Summary of UNet-based models with descriptions of each configuration.
Table 3.
Model complexity and computational cost comparison. 1 GFLOP = 1 billion FLOPs; 1 GMAC = 1 billion MACs.
Table 4.
Assessment metrics used in study. TP = true positive; TN = true negative; FP = false positive; FN = false negative.
Fig 9.
Training loss for terraceDL, mineBenchDL, and vfillDL datasets using varying training sample sizes and different model configurations across 25 training epochs.
Fig 10.
Validation F1-score for terraceDL, mineBenchDL, and vfillDL datasets using varying training sample sizes and different model configurations across 25 training epochs.
Fig 11.
F1-score calculated from the withheld test data using different architectural configurations and training set sizes.
Red points indicate the Base UNet model.
Table 5.
Testing set assessment metrics for prediction of agricultural terraces using different architectural configurations and varying sample sizes.
Table 6.
Testing set assessment metrics for prediction of historic mine benches using different architectural configurations and varying sample sizes.
Table 7.
Testing set assessment metrics for prediction of surface coal mine valley fill faces using different architectural configurations and varying sample sizes.
Fig 12.
Comparison of agricultural terraces detection results using different architectural configurations.
Fig 13.
Comparison of historic mine benches detection results using different architectural configurations.
Fig 14.
Comparison of valley fill faces detection results using different architectural configurations.
Fig 15.
Bootstrap analysis of overall accuracy, F1-score, precision, and recall across model configurations for mineBenchDL, terraceDL and vfillDL datasets.
Error bars represent variability from 10 bootstrap test samples, highlighting performance stability differences across configurations.