Figure 1
Overview map of the northern coast of Honduras showing Guadalupe’s location. Elevation data: ASTGTM Version 3 (NASA et al. 2019).
Figure 2
Selection of pottery examples from Guadalupe. a) Tripod vessel with geometric incisions, punctates, and zoomorphic appliqué lug (PAG-15-683). b) Constricted vessel with wave patterns and zoomorphic appliqué lug (PAG-40-3). c) Rim sherd with geometric paint pattern (PAG-43-3). d) Shallow tripod vessel with zoomorphic appliqué lug and anthropomorphic appliqué supports (PAG-53-29). e) Turtle-shaped ocarina (PAG-43-1). f) Roller stamp with geometric pattern (PAG-120-1). Photo credits: a) and b) F. Fecher, c) K. Engel, d) T. Remsey, e) P. Bayer, and f) M. Lyons.
Figure 3
Examples of the five ceramic fabric types analyzed. Row 1) Fabric a: Inclusion-sparse birefringent, Row 2) Fabric c: Amphibole-rich, Row 3) Fabric d: Common, fine inclusions, Row 4) Fabric e: Microfossil-rich, and Row 5) Fabric w: Poorly sorted angular inclusions. Column a) Exterior surface; scale: in cm, Column b) fresh break; scale: in mm, Column c) close-up of thin section under cross-polarized light; scale: image width = 2.8 mm, and Column d) thin section under cross-polarized light; scale: see image.
Figure 4
Conceptual schematic of the model architecture in use. Transfer learning combines the ‘frozen’ base layers of the VGG19 model with the additional trainable layers used to ‘tune’ the model to the thin section dataset.
Table 1
Approach A data partition of training, validation, and test image counts per fabric.
| FABRIC | TRAINING | VALIDATION | TEST | TOTAL |
|---|---|---|---|---|
| a | 80 | 10 | 11 | 101 |
| c | 127 | 16 | 16 | 159 |
| d | 323 | 40 | 41 | 404 |
| e | 140 | 18 | 18 | 176 |
| w | 183 | 23 | 23 | 229 |
| total | 853 | 107 | 109 | 1,069 |
| percent | 79.79% | 10.01% | 10.20% | 100% |
Figure 5
Training and testing results for VGG19 with Approach A. a) Training and validation accuracy (bottom) and loss (top) per epoch. b) Confusion matrix of test data showing the model’s predicted fabric types (Prediction) vs. actual fabric types (Reference).
Figure 6
Training and testing results for ResNet50 with Approach A. a) Training and validation accuracy (bottom) and loss (top) per epoch. b) Confusion matrix of test data showing the model’s predicted fabric types (Prediction) vs. actual fabric types (Reference).
Table 2
Approach B data partition of training, validation, and test image counts per fabric.
| FABRIC | TRAINING | VALIDATION | TEST | TOTAL |
|---|---|---|---|---|
| a | 58 | 7 | 36 | 101 |
| c | 102 | 13 | 44 | 159 |
| d | 284 | 35 | 85 | 404 |
| e | 92 | 11 | 73 | 176 |
| w | 153 | 19 | 57 | 229 |
| total | 689 | 85 | 295 | 1,069 |
| percent | 64.45% | 7.95% | 27.60% | 100% |
Figure 7
Training and testing results for VGG19 with Approach B. a) Training and validation accuracy (bottom) and loss (top) per epoch. b) Confusion matrix of test data showing the model’s predicted fabric types (Prediction) vs. actual fabric types (Reference).
Figure 8
Training and testing results for ResNet50 with Approach B. a) Training and validation accuracy (bottom) and loss (top) per epoch. b) Confusion matrix of test data showing the model’s predicted fabric types (Prediction) vs. actual fabric types (Reference).
Table 3
Summary of results showing the accuracy of fabric predictions for training, validation, and testing images with respect to each combination of data partitioning approach and base model.
| APPROACH | BASE MODEL | TRAINING | VALIDATION | TEST |
|---|---|---|---|---|
| A | VGG19 | 100% | 100% | 99.1% |
| A | ResNet50 | 100% | 99.1% | 100% |
| B | VGG19 | 99.5% | 96.6% | 96.3% |
| B | Resnet50 | 99.5% | 97.7% | 93.6% |