Recent developments in computer science have made available new techniques to analyze big data and derive the patterns that naturally occur in them. According to the SAS Institute, these techniques commonly called machine learning are a branch of artificial intelligence that group all the methods where computer learn from data, observe and evaluate patterns to make decisions without or with minimal human intervention. In parallel, the development of low cost unmanned aerial systems (drones) has also eased data collection, aerial surveys. A major constraint posed by drones is the limitation of the existing techniques to analyze the large volume of image data collected via drones. Making sense of these big data will provide valuable information that can be used to address critical issues faced by communities. However, artificial intelligence, in general, and machine learning, in particular, can address the challenges posed by the analysis of drones’ big data.
The present pilot project tests how drones can be combined with machine learning tools for urban planning. Conducted by Benin Flying Labs, the main objective of this pilot project is to provide municipalities with accurate and update data about their communes. These data can, in turn, serve in urban planning and address issues such as soil erosion, land management, monitoring of roads construction, waste management and their impact on health.
From October 21 to 24, 2018 the Benin Flying Labs team collected aerial images of the town of Dassa. More than 20GB of georeferenced images collected were stitched to produce an orthomosaic map of the town. This map is condensed but contains valuable information that can provide useful insights to managers. However, examining each section of this map before gathering the information needed is inefficient. It is a time-consuming task that could take days and the manager might not deduct a clear pattern after his inspection. In this pilot project, we rely on a supervised classification (supervised machine learning) to identify common objects on the maps such as constructions, vegetations, roads, railroads, hills, and nude soils. Figure 1 presents the results of the classification model and the original orthomosaic map. Figure 2 illustrates the same result but on a larger scale.
The results, which are highly accurate, show that the machine learning model can classify or categorize the different objects on the orthomosaic generated from the images collected with the drone. Such results also confirm that machine learning techniques can be used to quickly analyze large volumes of drones data. By using these techniques, managers at the local level, can quickly gain insights from the data collected with the drones and address the issues their constituents are facing.