Figure 1
The research area (black outline) on an elevation model of the Netherlands (source of the elevation model: Nationaal Georegister, 2021; coordinates in Amersfoort/RD New, EPSG: 28992; amended from Lambers, Verschoof-van der Vaart & Bourgeois 2019).
Figure 2
Excerpts of LiDAR data, visualised with Simple Local Relief Model (Hesse 2010), showing: (a) barrows; (b) Celtic fields; and (c) charcoal kilns (source of the elevation model: Nationaal Georegister 2021).
Figure 3
Flow diagram of the dataset creation Python script.
Table 1
The datasets used in this research (the columns Negative and Ratio show the amount and proportion of negative examples respectively).
| DATASET | IMAGES | NEGATIVE | RATIO |
|---|---|---|---|
| training | 3602 | 691 | 19.2% |
| test | 931 | 190 | 20.4% |
Table 2
Architecture of the vanilla versus modified YOLOv4 model, with the differences between them in bold.
| PARAMETERS | VANILLA | MODIFIED |
|---|---|---|
| backbone CNN | Darknet53 | Darknet53 |
| input resolution | 416 × 416 | 512 × 512 |
| activation function | Mish | Wish |
| data augmentation | — | Cutmix, Mosaic |
| regularisation | (DropBlock) | DropBlock |
| loss function | CIoU | GIoU |
| non-maximum suppression | greedyNMS | DIoUNMS |
Figure 4
Flow chart showing the testing of the developed workflows: The cropped subtiles of the test dataset are inputted into the Darknet53 CNN, which uses the trained weights and a configuration file to perform inference on the images. The detection results are stored in a .txt file, which is subsequently transformed into three CSV databases.
Figure 5
Excerpts of LiDAR data, visualised with Simple Local Relief Model (Hesse 2010), showing successful detections of: (a) barrows (red); (b) barrows (red) and Celtic fields (blue); (c) charcoal kilns (green); (d) barrows (red), Celtic fields (blue), and charcoal kilns (green); source of the elevation model: Nationaal Georegister 2021).
Table 3
The results, averaged on all classes, of the testing of the vanilla and modified YOLOv4 models on the test dataset.
| MODEL | AV. IOU | RECALL | PRECISION | F1 | MAP@50 |
|---|---|---|---|---|---|
| vanilla YOLOv4 | 0.45 | 0.84 | 0.58 | 0.69 | 0.75 |
| modified YOLOv4 | 0.57 | 0.93 | 0.64 | 0.76 | 0.86 |
Table 4
Confusion Matrix of the results of the modified YOLOv4 model on the test dataset. A perfect model should obtain an identity matrix.
| PREDICTIONS | ||||
|---|---|---|---|---|
| BARROW | CELTIC FIELD | CHARCOAL KILN | ||
| Truth | barrow | 0.88 | 0.10 | 0.02 |
| Celtic Field | 0.03 | 0.97 | 0.0 | |
| charcoal kiln | 0.05 | 0.00 | 0.95 | |
Figure 6
Excerpt of LiDAR data, visualised with Simple Local Relief Model (Hesse 2010), showing the problem of objects of confusion causing False Positives: a roundabout is classified as a barrow due to the similarity of the two in LiDAR data (source of the elevation model: Nationaal Georegister 2021).
Table 5
Performance (R: Recall, P: Precision, F1: F1-scores) per archaeological class. Higher is better.
| METHOD | CELTIC FIELDS | BARROWS | CHARCOAL KILNS | ||||||
|---|---|---|---|---|---|---|---|---|---|
| R | P | F1 | R | P | F1 | R | P | F1 | |
| YOLOv4 V | 0.75 | 0.75 | 0.75 | 0.92 | 0.65 | 0.76 | 0.44 | 0.90 | 0.59 |
| YOLOv4 M | 0.79 | 0.86 | 0.82 | 0.99 | 0.57 | 0.72 | 0.78 | 0.92 | 0.84 |
| WODAN1.0 | 0.53 | 0.90 | 0.15 | 0.43 | 0.21 | 0.28 | – | – | – |
| WODAN2.0 | 0.45 | 0.57 | 0.50 | 0.40 | 0.52 | 0.46 | 0.35 | 0.12 | 0.18 |
| Trier | 0.84 | 0.70 | 0.76 | – | – | – | 0.96 | 0.68 | 0.80 |
| Bonhage | – | – | – | – | – | – | 0.83 | 0.87 | 0.85 |