Aerospace and Electronic Systems - August 2018 - 65

Addabbo et al.

Figure 8.

Detection times for both panel and defects.

We have evaluated the effectiveness of the classification method
with the procedure summarized below:
1. build a testing set T (gold standard) in which modules are identified and classified by an expert as "Good" or "Bad";
2. apply the classifier to each element of T to detect modules and
defects;
3. compare classifier's detection with the gold standard and evaluate performance metrics.
The metrics obtained are shown in Table 1. Time columns represent the average time, expressed in seconds, needed to build the
classifier and to perform the detection on two flight sequences. In
Figure 8, times are reported for a wider set of sequences of different
sizes. The trend confirms that panel identification weighs more than
defects detection. The reason of this difference resides in the sizes
of templates and the number of passes to perform classification.
First of all, templates of panels are larger than the template of defects and this influences cross-correlation computation. Moreover,
panel identification requires a two-pass algorithm using two templates, one with defects and one without defects, i.e., a healthy one.
The following consideration can be made:
C

the F-measure ranges from 0,18 to 0,83 for panel detection
and from 0,01 to 0,75 for defect detection;
best results for panels and defects detection are obtained for
the threshold of respectively 0.96 and 0.985; above these
values, F-measure starts to decrease.

AUGUST 2018

reducing the threshold increases the Recall but quickly lowers the Precision: this can be seen more clearly in Figure 9
that reports Precision and Recall versus the threshold for
Panels and Defects detection.

CONCLUSIONS AND FUTURE WORK
In this article, a computer vision approach applied to a PV plant
for the identification of the panels and the detection of thermal
anomalies has been proposed. The approach integrates geographic
information gathered from GNSS with results of a computer vision template matching algorithm applied to thermal images. This
combination allows to perform panel identification by assigning to
each module an identifier that remains consistent across different
flight sessions. Through the proposed matching algorithm different
templates have been exploited, to detect both panel extension and
the presence of defects. The proposed approach opens up to further
work to be focused mainly on the following points, as already discussed in [3]: a barometric altimeter to be tested in order to assess
the RPAS altimeter resolution; a UAV/RPAS ground speed of less
than 3 m/s; a gimbal orientation normal to object reference system;
the evaluation of the misalignments (few milliseconds) introduced
by wind gusts; assessment of the PV panels in the case of lack of
information about PV panels placement in a planar surface; inclination of the PV panels that reduces their cross-section as seen
from the UAV/RPAS; introduction of sensor lens distortion that
has to be assessed and mitigated during the tests, since it accounts

IEEE A&E SYSTEMS MAGAZINE

Aerospace and Electronic Systems - August 2018

Table of Contents for the Digital Edition of Aerospace and Electronic Systems - August 2018