Automatic roof area inspection and geomarketing

Initial situation

Planning, coordination, and control are three interrelated, essential components in companies' decision-making processes to undertake and prioritize market activities. Remote sensing methods can derive information, or geodata, about large areas using airborne sensor data. This geodata helps to answer specific questions of a company and makes the underlying information more quickly tangible through deriving relevant parameters or visualization in thematic maps.

Verschiedene Ergebnisbilder von beschädigten Dächern
© Fraunhofer IOSB
Detection of damaged roof parts.
Erstellung einer thematischen Karte, in der Gebäude in Abhängigkeit ihres Beschädigungsgrades markiert sind.
© Fraunhofer IOSB
Determination of a thematic map picturing buildings in dependency of their degree of damage.
Markierung von Dachfenstern mittels CNNs. Durch die hohe Bodenauflösung (10 cm) ist eine direkte Klassifikation möglich.
© Fraunhofer IOSB
Highlighting roof windows using CNNs. Due to the high ground resolution (10 cm) a direct classification is possible.

However, such information is not only of particular interest for the evaluation of market potential. Up-to-date site plans also constitute essential support for planning emergency forces in disaster control and related relief measures. Besides, analyses of aerial images linked to thematic maps can facilitate time-critical and sometimes dangerous tasks for insurance adjusters and accelerate claims for damages.

Needs assessment

The structure of thematic maps is designed to color-code regions of particular interest depending on the issue. Relevant parameters are first derived from airborne sensor data and then visualized on a map for a better geographical overview.

For manufacturers and distributors of roof windows, it is interesting, for example, to determine the roof structures as well as the number of roof windows of residential buildings in an administrative district.

On the other hand, after significant damage events, such as those caused by a storm event, relief crews and insurance companies are interested in quickly identifying areas of heavy versus light damage so they can prioritize where quick assistance is needed. A first impression of the actual extent of damage after storm disasters is provided by the roof's condition. But after natural disasters, access to insured properties is often difficult. Besides, a damage assessor would have to inspect the roof, particularly from above, which is a time-consuming and dangerous affair.  Nowadays, access to drone images or other high-resolution data is widely easy, even for amateur users. Such images have the potential to spare examiners driveways to remoted, hardly accessible areas. Even though creating a thematic map cannot fully replace the damage assessor's job, it provides indispensable support in ensuring that help gets quickly to the places where it is needed most.


Machine learning methods are suitable for solving tasks from both application areas. Our methods are capable of deriving parameters for roof windows as well as for damaged roofs. Ideally, the resolution of the underlying image material also allows the localization of windows and damage. We refer to this procedure as direct classification. In this, the image material is first cropped to the regions of interest. This can be the outline of the entire roof or a single roof surface.


Bestimmung der Anzahl von Dachfenstern auf Dachflächen mittels Bag-of-Words. Durch die geringe Bodenauflösung (20 cm) liefert nur eine indirekte Klassifikation solide Ergebnisse.
© Fraunhofer IOSB
Determination of number of roof windows on roof planes using the Bag-Of-Words approach. Due to the low ground resolution (20 cm) only an indirect classification provides solid results.
Bunt eingezeichnete erkannte Dachfenster
© Fraunhofer IOSB
Thematic window occurance map.

In the further course, suitable segments can be selected. On the one hand, this is done to accelerate the calculations and, on the other hand, to include neighborhood features. Different features from color, texture, morphological profiles and many more are learned as geometric and radiometric features. Thus data sets are evaluated with classifiers of supervised learning like Random Forest or Deep Learning approaches. If additional 3D data is available, it can be fused with the acquired images to derive additional helpful features. A particular challenge in this context is the often complex structure and superstructure of the roofs. Damage must not be confused with regular roof structures such as solar installations, chimneys, ventilation systems, satellite dishes, etc.

An alternative approach refers to an indirect classification. For example, only the existence of a certain number of windows or damage is inferred without localizing them directly. This method is particularly suitable for underlying low-resolution image material. This approach also starts by determining the region of interest. Still, it delivers a decision for the entire region whether it belongs to the class with 0, 1, 2, 3, etc., windows or to the class damaged or undamaged. Depending on the availability and quality of training data, this can be done using a "Bag-of-Words" approach or Deep learning approaches, for example, those based on image recognition.


Department Scene Analysis

You want to learn more about our products in the area of scene analysis? Then visit the page of our department.