Inspection work

Why this matters

Inspection differs from mapping. Mapping asks “what is where?”; inspection asks “is this damaged, and how badly?” A crack a few millimetres wide needs images many times more detailed than a map, and hot spots on a solar panel are invisible to an ordinary camera. Drones reach high places without scaffolding, but a poor plan produces thousands of images that cannot answer the question.

Plan from the damage you must see

Start by asking how large is the smallest damage that must be seen, then decide how many pixels must span it. Lee and colleagues (2026) found that cracks only 3 pixels wide or less are measured with very high uncertainty, but they set no fixed number. The pixel count is therefore a requirement agreed between the team and the user, not a universal standard.

Then use the GSD formula from module 1, replacing height with the distance to the wall $D$.

SENSOR_W, PX_W = 13.2, 5472          # mm, pixels
CRACK_MM, PX_ACROSS = 2.0, 4
gsd_mm = CRACK_MM / PX_ACROSS

for name, focal in (("wide 8.8 mm", 8.8), ("zoom 50 mm", 50.0)):
    distance_m = gsd_mm * focal * PX_W / SENSOR_W / 1000
    width_m = distance_m * SENSOR_W / focal
    print(f"{name}: GSD {gsd_mm} mm needs distance <= {distance_m:.1f} m, image covers {width_m:.2f} m of wall")

wide 8.8 mm: GSD 0.5 mm needs distance <= 1.8 m, image covers 2.74 m of wall
zoom 50 mm: GSD 0.5 mm needs distance <= 10.4 m, image covers 2.74 m of wall

Facade inspection lanes usually run vertically or horizontally parallel to the wall at a constant distance, with enough overlap to mosaic the images or build a model, and with the camera perpendicular to the wall so crack sizes are not distorted. Every image must be locatable on the structure so the repair team can find the spot.

Thermal imaging

Every object above absolute zero emits infrared radiation. The Stefan–Boltzmann law says a perfect blackbody emits power per unit area $M=\sigma T^4$. Real objects emit less, scaled by their emissivity ($\epsilon$), between 0 and 1. Vollmer and Möllmann explain that for an opaque object, the part not emitted ($1-\epsilon$) is reflected. The camera therefore receives both the object’s own radiation and radiation from the surroundings reflected by it, represented by the reflected temperature ($T_{\text{refl}}$).

SIGMA = 5.670374419e-8

def to_k(c): return c + 273.15

def received(t_obj_c, eps, t_refl_c):
    return eps * SIGMA * to_k(t_obj_c) ** 4 + (1 - eps) * SIGMA * to_k(t_refl_c) ** 4

def read_temp(w, eps_set, t_refl_c):
    return ((w - (1 - eps_set) * SIGMA * to_k(t_refl_c) ** 4) / (eps_set * SIGMA)) ** 0.25 - 273.15

w = received(60.0, 0.85, 30.0)          # true object 60 °C, true emissivity 0.85
for eps_set in (0.85, 0.95, 1.00):
    print(f"camera set eps {eps_set:.2f}: reads {read_temp(w, eps_set, 30.0):.1f} °C")

camera set eps 0.85: reads 60.0 °C
camera set eps 0.95: reads 57.2 °C
camera set eps 1.00: reads 56.0 °C

NETD (noise equivalent temperature difference) is the smallest temperature difference the camera can resolve, in mK; lower is more sensitive. Thermal cameras have far fewer pixels than colour cameras, for example 640 × 512, so the GSD of the thermal camera must always be computed separately.

Solar panels and other structures

Outdoor thermography of solar panels is covered by the technical specification IEC TS 62446-3:2017, which sets inspection conditions and the classification of anomalies. The IEA PVPS report recommends imaging when irradiance exceeds 600 W/m², with no clouds and low wind, because the panels must be carrying real current for defects to heat up enough to be seen. (The 600 W/m² value is quoted from the IEA PVPS report and was not checked against the table in the IEC specification itself.)

from statistics import median

temps = [  # °C per panel
    [44.8, 45.1, 45.3, 44.9, 45.6, 45.0, 44.7, 45.2],
    [45.4, 45.0, 62.3, 45.1, 44.9, 45.5, 45.3, 44.8],
    [45.0, 44.6, 45.2, 45.8, 51.4, 45.1, 45.0, 45.3],
]
ref = median(t for row in temps for t in row)
print(f"median {ref:.2f} °C")
for r, row in enumerate(temps, 1):
    for c, t in enumerate(row, 1):
        dt = t - ref
        if dt >= 5:
            level = "report now" if dt >= 10 else "re-inspect"
            print(f"row {r} panel {c}: {t:.1f} °C, dT {dt:+.1f} K -> {level}")

median 45.10 °C
row 2 panel 3: 62.3 °C, dT +17.2 K -> report now
row 3 panel 5: 51.4 °C, dT +6.3 K -> re-inspect

The limits of drone inspection must be stated plainly to the user. The US bridge inspection standards (NBIS 2022) say drones may be used to supplement inspection by qualified personnel but cannot replace tasks that need contact, such as sounding with a hammer. ASTM D4788, for detecting bridge-deck delamination with thermography, was written for vehicle-mounted scanners, so applying the method from a drone must be shown to be suitable.

Module lab

Common mistakes

Summary

• Start from the damage size to be seen and the agreed pixel count, then compute the stand-off distance and choose the lens
• A thermal camera receives both the object’s radiation and reflected radiation; wrong emissivity or reflected temperature settings give wrong temperatures
• Find hot spots by their difference from the median, with criteria set in advance, then confirm them by another method
• Drones supplement inspection by qualified personnel; they do not replace contact inspection

Check your understanding

1. A GSD of 0.5 mm is required with a 50 mm lens and a 13.2 mm, 5472-pixel sensor. What is the maximum distance from the wall?
2. How much power per unit area does a blackbody at 300 K emit?
3. For an opaque object with emissivity 0.9, what fraction is reflected?
4. Which thermal camera is more sensitive, one with 30 mK NETD or one with 50 mK?
5. Why is the median used instead of the mean as the reference when looking for hot spots?