Land monitoring system: Comparison of traditional machine learning and U-Net convolutional neural networks approaches applied to semantic segmentation of drone imagery

This paper uses traditional machine learning techniques, including supervised learning (SVM, Random Forest, KNN) as well as the U-net algorithm via transfer learning, for the segmentation of aerial images taken by drones. Drones are evolving rapidly and are increasingly being used for a variety of professional missions, going beyond their initial use for audiovisual or leisure purposes. Drones are playing an active role in the digitization process, and are now being used for highly advanced technical digital content. Their versatility in day-to-day use means that they can replace costly helicopter and airplane operations, as well as the purchase of satellite images, by offering superior image quality over a well-defined area. In this image segmentation process, images of flooding in Dakar collected by a drone over a well-defined area were used to construct georeferenced orthophotos. Then, the second step consists in creating labels for our images and augmenting the data to broaden our training base through data augmentation, in order to assess the advantages and limitations of the different methods to guide future choices between the model derived from transfer learning with Resnet50, the SVM algorithm, Random Forest and KNN. A comparison of these different methods shows that the Resnet50 algorithm performs best, with an average IoU = 0.8232403. These results will be used to track changes in the classes defined, enabling us to efficiently monitor land use, flooding, identify logged areas and so on.