More sustainability in industrial buildings through AI

Introduction to the project

The idea of the project is to recognize the sustainability of industrial buildings and use artificial intelligence to determine ideas on how buildings can be retrofitted sustainably by entering the address of a building and outputting a list of possible retrofit ideas. The calculations that convert the address into the information needed to evaluate sustainability are described below.

Data collection and first steps

The first step is to determine the coordinates, i.e. the longitude and latitude, of the address entered. Then coordinates are calculated that describe the area of approximately one square kilometer around the center of the entered address. These coordinates are used to search for satellite data that can provide information about the building. There are various satellites that record maps of the earth's surface (called remote sensing), but the images differ in several respects. One point that is important for this project is the spatial resolution, i.e. how many meters a pixel of the map measures, i.e. how much can be seen on the map. Another point that is interesting and important for satellite images that do not only observe in the optical range is the spectral resolution, i.e. how many wavelengths are observed.

Analysis of the satellite data

Presentation of relevant satellites and their data quality

Three satellites that were considered for this project are presented below:

Sentinel-2A data has a spatial resolution of 10 m in the optical wavelength range and 60 m in the infrared wavelength range. This means that small industrial buildings are only mapped with a few pixels. However, the advantage of the Sentinel data is that data is available in the infrared, which means that the albedo, i.e. the reflectivity of a surface, can be calculated using the recorded spectrum (see info box). This information indicates how much the roof surface, and therefore the building, heats up due to solar radiation.

Higher spatial resolution data is recorded using the Pleiades and SPOT satellites, among others. These maps must be purchased; there are several providers who sell complete maps or only sections of the individual maps. Pleiades maps have a spatial resolution of 0.5 m (NEO data of 0.3 m) and SPOT data of 1.5 m. This provides a good spatial representation of buildings and allows objects on the building roof, such as windows, to be recognized. However, these satellites do not record any or sufficient information in the infrared, which means that the albedo cannot be calculated. In this case, conclusions about the heating of the building are drawn from the brightness of the roof surface.

The satellite maps are available as tif files containing geo-information, which means that individual pixels or areas can be found using geographical coordinates.

Integration of OSM maps to make the data more precise

For an analysis, further information is required in the form of an additional map containing the building outlines. These are Open Street Map (OSM) maps that can be downloaded for free. The address coordinates are used to find the correct building on the OSM map and the building outline is read using the "buildings" files. This information is transferred to the satellite map so that the building can be cut out on the satellite map. In order to examine the immediate surroundings of a building, e.g. whether there are green areas, trees or mainly asphalted areas, a necessary number of pixels around the recognized building are cut out from the satellite map and made available for analysis.

Detailed evaluation of the building environment

Material analysis through spectral evaluation

The analysis steps that determine the brightness of the roof and its surroundings differ depending on the satellite map. Sentinel measures several wavelengths for each recorded pixel, both in the optical and infrared range, so that a spectrum can be created for each pixel. These spectra can be used to determine the material composition of the observed area. For this purpose, spectra of different materials measured in the laboratory are used and compared with the observed spectra. This makes it possible to determine material information for each pixel, i.e. also for the selected building roof and the building surroundings. One disadvantage here, however, is the low spatial resolution. As a pixel measures an area of 10×10 m or 60×60 m, it can be assumed that different materials can also be found in this area, which mix to form a spectrum. Unfortunately, the exact composition of the materials cannot be clearly determined using this method, as the spectral resolution is also not high enough, but it is possible to recognize green areas. However, the spectra are well suited to calculating the albedo of the surfaces, i.e. the reflectivity of a surface. This value can be used to determine how much the roof, and therefore the building, is heated by solar radiation.

The albedo cannot be calculated with the Pleiades and SPOT maps as there is insufficient information in the infrared. In this case, the brightness is calculated using the data in the optical wavelength range. For this purpose, a gray image is created and the brightness is read off. As the resolution for these images is very high in some cases, it is possible to cut out the roof precisely, detect any shadows or detect windows, which allows conclusions to be drawn about heating and ventilation.

Recognition and analysis of green spaces

Another approach that only relates to the optical range, i.e. is the same for all satellite data, is the recognition of green areas by color. For this purpose, the maps in the optical wavelength ranges are combined into one map and the color green is selected. This allows the proportion of green areas on and around the building to be determined. It is assumed that the green areas are vegetation, i.e. trees and grass areas. In connection with this project, photos of the building are also examined from all four wind directions. For each individual photo, the building is first cut out using object recognition. Window detection can also be performed using object detection, or using a different approach in which edges on the building are recognized (edge detection), which determine the windows. 

For the latter approach, the optical maps are converted to gray images and edge detection is performed. It is advisable to repeat this process with different gray image representations and with inverses of the gray images. The windows found on each of these gray images are then compared with each other. The proportion of windows on the house wall is determined in this way. This provides further information about ventilation possibilities and heating of the building. The calculation of the brightness of the house walls is carried out on the building sections without windows. In contrast to the roof brightness detection, the main colors of the façade are detected for the photos. The brightness is then determined on the basis of the main color, which must account for more than 50% of all colors, or the proportional average of all colors. 

Detailed evaluation of the building environment

With this calculated information, it is possible to give an assessment of how sustainable the building is and to provide information on how it can be made more sustainable.
Scroll up