|
Semantic segmentation is a suitable deep learning method to determine the area of an object without marking the object in a rectangular box and assigning it to a class. With semantic segmentation, each pixel of an image can be optimally assigned to a class. This is particularly advantageous when determining the degree of coverage of different habitats in the Wadden Sea, such as mussel beds. In this paper, two well-known segmentation methods, U-Net and Mask R-CNN, are compared. Experimental results of the deep learning evaluation are presented. Evaluation metrics such as intersect over union (IOU) and mean average precision (mAP) are compared. IOU is a particularly interesting metric in this use case, as it specifically concerns the degree of coverage, i.e. the correct recognition of an area. In addition, the specific selection criteria of the networks used and a justification for the final network used are given. Furthermore, the requirements analysis lists the specifications that distinguish this project from others. Images taken by a drone are used as training and test data. More information about the drone, the camera and the flight altitude can be found in the paper. The aim is to validate satellite data, which in the past has always been done by hand or on foot. In particular, the strengths and weaknesses of U-Net and Mask R-CNN are explored and described in this work. The dataset used, the parameters set and the computational and time effort are explained in more detail. In particular, our deep learning evaluation shows a significant milestone with an outstanding IOU of over 85%, confirming the ability of semantic segmentation to accurately define object areas within the diverse habitats of the Wadden Sea.
|
