VEN
©) du
ant
lem
and
the
tion
on
ed).
pass
om
it's
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
important to take care of. Scene geometry has to be corrected.
Usual methodology based on simple polynomial approach
cannot model such geometry especially in a mountain region as
the study one is. Orthoprojection has to be considered to make
MIVIS data suitable for the subsequent data integration.
Again both RFM and MLP NN approaches has been tested to
correct available image, whose pixel size is about 3 meters. In
Table 3 are some results concerning reached accuracy.
N? N? AE An RMSE | RMSE
Method GCPs | CHKs | mean mean CHK GCP
CHK CHK | (pixel) | (pixel)
RFM 72 10 0.00 0.00 6.13 5.09
MLP NN 72 10 0.00 -1.07 4.00 2.56
Table 3 — Accuracy tests results obtained with the RFM and
MLP NN self-developed orthoprojection routines on the
airborne sensor MIVIS image.
Results show a better performance of the MLP NN approach
and underline the need of a high number of GCPs especially in
a mountain region as the one considered is. Best performance
of MLP we guess it can be due to the better generalisation
capability of this techniques. It is in fact known that RFM are
mathematical model suitable for pushbroom type images, and
they can presents some limits for other image geometry.
Figure 9 shows a qualitative verification of correction
performance by overlaying the 1:10000 CTR map.
3 d My Anais
afi RE Du o
Figure 9 - 1:10000 Vector map overlaid onto the orthoprojected
image.
2.2.3 Significant bands selection
The second step was to select the most useful bands for the
detection of archaeological features and anomalies. These can
be identified by the texture, soil moisture and vegetation cover
differences that are produced by buried structures.
The peculiarity of the investigated objects led us to select bands
renouncing to the principal components analysis; we proceeded
with a visual interpretation taking care of the bibliographic
references. A total of 10 of the 102 available bands of the
MIVIS sensor were chosen:
1071
e 4 in the visible range (b2= 0.4600 pm, b7=0.5600 um, b11=
0.6400 pm, b20= 0.8200 pum) for the contextual location;
e 2 in the near infrared range (b23= 1.2750 um, b28= 1.5250
um), for vegetation cover anomalies;
e 1 in the medium infrared range (b52= 2.1790 um) for soil
moisture;
e 3 in the thermal infrared range (b93= 8.3859 pm, b97=
10.0200 pum, b101= 11.9450 pm) for termal variation on the
ground.
No calibration have been made as no calibration file was
available for the test image. That has not be considered
fundamental because relative, and not absolute, differences
between objects had to be investigated.
2.2.4 Test Areas Selection and Image Masking
Queries and spatial analysis performed on the data collected in
the Marchesato di Saluzzo GIS permitted to choose 2 test areas
responding to the appropriate archaeological needs:
The Sant'Iario monastery
This monastery is near the town of Revello, close to the Po
valley mouth. The documentary sources refer to three villages
in the second half of the XII Century. Today there is no sign of
these settlements which in the documents were known as
Sant Ilario, Viverio e Paralupo. The XII Century documents
also refear to a road called “via publica” which was situated
near the monastery.
The San Massimo church
The site of this church was between the Revello and Envie
towns. Thanks to a document of the half of the XIII century, we
know there was a very important road called “via monnea
superius" which was near San Massimo church. The name of
this road seems to suggest a paved road. This road was part of a
longer, probably Roman route, which joined the town of
Saluzzo with the town of Bricherasio.
We built and applied an opportune mask to bound these two
areas. A mask is a binary image that consists of values of 0 and
1. When a mask is used in a processing function, the areas which
have values of 1 are processed and the masked 0 areas are not
included in the calculations. This procedure has permitted us to
limit the investigation and the radiometric bothers.
2.2.5 Image Classification and Validation
Nine regions of interest (ROI) were selected inside the sample
areas: buildings, industrial buildings, water, streets, soil
moisture, orchards, vegetated fields, non-vegetated fields,
shadow zones.
A spectral angle mapper (SAM) classification with an angle
threshold of 0.10 (radians) was applied. For the validation of
the classification of the eight classes the correspondent
confusion matrix (here not reported) was calculated. It shows a
correct classification of ROIs, although in the next future a
certification on the ground could be necessary.
Rule images that were generated during the classification were
considered. In rule images the dark pixels mean a similar
spectral signature to the selected class, while the gray scale
pixels mean a different one. Rule images are very helpful in