3.1. Visual Inspections of LDMs: Results and Discussion
Figure 7 shows the results from PCA of Hill Shading, LRM, SVF, ASVF, Positive Openness, and Slope.
At a quick comparative glance, a microtopography attributable to shallow remains of a village including fortified structures emerges. Furthermore, the LDMs show different results in terms of feature visibility from both the quantitative (number of identifiable features) and qualitative (feature discrimination) point of view.
Some of these features are easily recognizable from all the LDMs, as in the case of the quadrangular landform at south east of the hilly plateau (see c in
Figure 7d,e,g,h) which is reasonably attributable to a tower-castle as expected on the basis of the information recorded by the historical topographic maps reported in
Figure 2a–c. Moreover, all the LDMs clearly show microrelief bordering the edges of the plateau (see w1 in
Figure 7d,e,g,h) and some linear features North east of the hill (see l in
Figure 7d,g,h). The first are probably city walls, whereas the second may refer to possible urban streets.
Finally, all the other features, representing the greatest amount of microrelief, refer to potential shallow structures of buildings (see b in
Figure 7e,g), which are less visible and more difficult to recognize and interpret compared to the bigger features of the castle and the potential city walls.
For these subtle features, a significant enhancement in terms of feature visibility has been obtained in particular by SVF, Openness, and ASFV (see
Figure 7e–g, respectively)
However, all the LDMs are complementary for their diverse capability to emphasize the micro-topographical features. For example, the slope map shows, in an effective way, the quadrangular landform, bringing out the fortified structure (see c in
Figure 7h). Also, the linear features l (showed in
Figure 7h) are very well imaged by the slope map, thus suggesting that they would be urban streets. Both the LRM and PCA_HD (see
Figure 7c,d), even if quite limited in imaging small features, display the quadrangular feature
w2 at the east of the potential city well, thus suggesting the presence of extramural walls (see also w
2 in
Figure 7e,g,h).
As a whole, the best results are obtained from SVF, ASVF, and Openness especially for the smaller features (see b in
Figure 7e,g and b
i in the map of
Figure 8) of shallow buildings and urban blocks. Moreover, compared to LRM and Slope, they provide additional details of earthworks and landforms attributed to the tower castle and walls.
The integration of the features observed from all the LDMs (showed in
Figure 7c–h) allowed the identification and mapping of all the potential microrelief of archaeological interest (see map in
Figure 8). Four types of features have been identified. They are: (i) earthworks of possible city walls (w
1.i) and an extramural wall (w
2), (ii) landform and microrelief of the castle, (iii) micro-topographical relief probably related to buried buildings and streets of the new discovered village (indicated with b
i in
Figure 7), and finally, (iv) other features of potential anthropogenic origin (denoted with o
i)
A qualitative assessment of the feature visibility of LDMs (see
Figure 8) has been performed by visual inspection. The reconnaissance of the features has been facilitated by the field survey (see
Figure 9) conducted to verify the cultural interest of landforms. The forest of oaks with dense undergrowth made the survey very difficult and, in many cases, the areas to be surveyed were inaccessible. Furthermore, most of the potential microrelief identifiable from the LDMs is not easily visible on the ground. However, the in situ observation allowed us to map the presence of materials, such as brick roof tiles, fragments of bricks, and worked stones, which helped us in the interpretation of the remotely-sensed archaeological features. The following two criteria were adopted for selecting the areas for the in-situ inspections: (a) the presence of microtopographical variations visible from LiDAR-based maps and profiles; and (b) the accessibility. As a whole, 15 areas (showed in
Figure 4) representative of the diverse types of archaeological features, such as microrelief attributable to shallow buildings and landforms related to the castle and city walls, were inspected and measured by GPS.
The field survey was conducted in order to identify the following archaeological indicators: (i) walls, wall foundations, and stone material of collapsed masonry structures; (ii) building surface materials, such as worked stones, bricks, and roof tiles; (iii) visible and surveyable microrelief (including earth works) along with surface building materials.
The existence of one of the three above said indicators has been considered enough for the ‘archaeological validation’ of the features identified from LDMs.
Figure 9 shows four areas surveyed, named 1, 3, 5, and 8. They are related respectively to city walls (1), the castle (3), the tower (5), and an area including a road and a urban block (8). In (1), the archaeological indicators were earthworks and sparse building materials, including worked stones, bricks, and tiles. In (3), sparse building materials and remains of collapsed walls were observed. In (5), remains of walls of height ranging from 40 to 120 cm were measured. The masonry is composed of slightly carved stone blocks, complemented with small ashlars and fragments of bricks laid in place on more or less regular horizontal rows. Finally, in (8), sparse building materials including bricks and tiles were observed.
As previously said, the assessment of archaeological feature visibility from the diverse LDMs has been conducted by six of the coauthors: two experts in remote sensing archaeology (N.M. and R.L.), one archaeologist (A.P.), one historian, an expert in conservation (M.B.), and two geologists (F.T.G. and M.S.).
The assessment has been performed comparing features x
i (w1.1, w1.2...b1, b2, b3..., o1, o2, o3 see
Table 2) mapped (xi
FM) in
Figure 8 with the features visible from the single derived model (xi
FDM).
Furthermore, the normalized visibility index shown in Formula (3) has been computed for the diverse LDMs.
where µx
i is normalized visibility index for the given LDMs reported in
Table 2 (µi
PCA, µi
LRM, µi
SVF etc.); Lxi
FDM is the length of the feature x
i visible from a given LDM; Lxi
FM is the length of x
i as mapped in
Figure 8.
Table 2 lists the features x
i grouped for 4 classes related to city walls (wi), the castle (ci), small microrelief attributed to potential buried buildings or streets of the urban fabric (bi), and other features of potential archaeological interest (oi). For each class and each derived model, the weighted average of normalized visibility index µ
LDM has been computed according to Formula (4).
where the parameters are the same as in Formula (3). This index has been also computed for the features validated in situ: in this case the index is named µ
LDM′.
As whole,
Table 2 lists 33 features, also drawn in the map of
Figure 8, attributable to city walls (wi), the castle (ci), the urban blocks and buildings (bi), and other features (oi) of the medieval village. In
Table 2, column GDV indicates the LDMs-based features x
i inspected in situ and the GDR column reports the results of the in situ survey (presence of walls, surface building materials etc.).
In particular, with respect to city walls, n.5 features were recognized from the LDMs for a total length of 469 m, among which n.3 features were observed and validated on the ground for a total length of 413 m. They were characterized by microrelief, building surface materials (w1.1, w1.2), and remains of walls (w2). The other features (wi.3 and w1.4) were not inspected due to the dense vegetation.
For all the LDMs, the normalized visibility index (µLDM′) exhibits values ranging from 0.85 to 0.98. In particular, higher values (0.95 to 0.98) were obtained for LRM, SVF, ASVF, and Openness, and slightly lower values for PCA_HD and Slope (0.85 and 0.91, respectively).
Similar values have been obtained also considering the features inspected on the ground (see µ
LDM and compare with µ
LDM′ in
Table 2).
In the area of the castle, n.5 features visible from the LDMs have been all inspected. The inspection revealed some walls, microrelief, and a great amount of building materials, such as ashlars, bricks, tiles, and pottery (see column GDV and GDR of
Table 2). Moreover, the comparative assessment put in evidence higher values of µ
LDM for SVF, ASVF, and Openness (µ
LD = 1) with respect to PCA_HD, LRM, and Slope (equal to 0.54, 0.72 and 0.74, respectively).
For small features (bi), the µ
LDM values obtained from SVF, ASVF, and Openness were much higher than those PCA_HD, LRM and Slope: 0.97 to 1 versus 0.40 to 0.66, respectively (see also
Figure 9). This difference is significantly reduced considering only the features inspected (0.98 to 1 for SVF, ASVF, and Openness, versus 0.73 to 0.89 for PCA_HD, LRM and Slope). In this case, the vegetation cover density played a fundamental role; in fact, where it was possible to inspect the microtopographical features, the vegetation cover was less dense and, consequently, was characterized by a higher ground return density with a better visibility of microtopographical changes. A similar behavior has been found for the class of other archaeological features (see histograms in
Figure 10).
As a whole, the different performance of the LDMs is due to the fact that the diffuse illumination-based models (SVF, ASVF, and Openness) enhance the edges better than the other methods (PCA of Hill Shading, Slope, LRM). This is particularly evident for small features. The histograms in
Figure 10 show that, considering the entire set of LDMs, the best results in terms of feature visibility have been obtained for the city walls and the castle because they were characterized by greater dimensions, more durable construction materials, and were abandoned later compared to buildings of the village.
3.2. Automatic Feature Extraction: Results and Discussion
Automatic feature extraction (AFE) was performed using ISODATA and segmentation, described in
Section 2.5.2 (see also flowchart in
Figure 5), applied with and without Local Geary index, to assess its impact on the data processing chain. The use of Local Geary index has been successfully used for extracting archaeological looted areas in Syria and in Peru [
44]. Results from AFE performed for all the LDMs, with and without of the use of Geary index, generally provided only bigger features related to castle and city walls. In order to compare the performance obtained from the diverse LDMs, the outputs from the automatic feature extraction have been compared with the map in
Figure 8 and field surveys.
For sake of brevity, we only show and discuss the results of AFE obtained from SVF, Positive Openness, and the Local Relief Model.
Figure 11 depicts the partial and final results of AFE.
The AFE method without Geary presents a very good result from Positive Openness; whereas LRM extracts a large part of the ‘big’ features but with a high number of commission errors. Finally, for SVF, AFE extracts only some structures of the castle and sloping surface of the hill.
The use of Geary before ISODATA allowed significant improvement in the extraction of archaeological features, with a result comparable with that obtained from Openness without Geary. On the other side, with respect to Openness, the use of Geary does not provide any improvement but tends to add commission errors. Similarly, Geary, when applied to LRM, reduces the number of pixels extracted, thus losing the spatial continuity of archaeological features.
The difference in the results with and without Geary for the diverse LDMs can be explained on the basis of their different enhancement behaviors. In particular, positive openness is devised to identify and thin edges that tend to be degraded after Geary because it detects areas of dissimilarity and areas characterized by a high variability compared to the values of its neighboring pixels, thus: (i) ‘loosing pixels’ along the archaeological features of walls and castle already highlighted by Openness, and (ii) adding commission errors, particularly North and Northeast of the plateau.
While SVF mainly delineates concave features and maintains the visibility of the slope, it offers more margins of improvement in the enhancement of local microtopography; for this reason, the use of Geary, detecting dissimilarity and high variability, enabled us to sharpen microtopographical changes. AFE takes advantage of this, providing suitable extraction of archaeological features.
Figure 12 resumes the results of AFE obtained from SVF and Openness, with and without Geary, respectively, which have been compared by overlaying both of them with the map in
Figure 8. The comparison (see
Table 3) shows a rate of success (RS) ranging from 81 to 97% for the castle and city walls, with no significant difference between Openness (RS equal to 86% and 93% for the castle and city walls, respectively) and SVF (RS equal to 83% and 97%, for the castle and city walls, respectively).
The main difference in the performance from AFE applied to Openness and SVF is evident for the smaller features, such as buildings and other archaeological features: the values of RS obtained for Openness are 14% and 34%, respectively, compared to 4% and 20%, respectively, for SVF.