A comparative study of loss functions for road segmentation in remotely sensed road datasets

Road extraction from remote sensing imagery is a fundamental task in the field of image semantic segmentation. For this goal, numerous supervised deep learning techniques have been created, along with the employment of various loss functions that play a crucial role in determining the performances of supervised learning models. However, there is a lack of comprehensive analysis of the performance differences between the loss functions for road segmentation in remote sensing imagery. Therefore, this study conducts a comparative study of 12 well-known loss functions used widely in the field of image segmentation by training and evaluating the representative D-LinkNet network for road segmentation tasks with two publicly available remote sensing road datasets consisting of very high-resolution aerial and satellite images. The results show that different loss functions could lead to very different outcomes using the D-LinkNet, with varying focuses such as on overall model performances, precision, or recall. By dividing the loss functions into the distribution-based, region-based, and compound ones, we found that the region-based loss function type led to generally better model performances than the distribution-based one in terms of F1, IoU, and the road segmentation maps, with the compound loss function type being comparable to the region-based one. This paper eventually tries to offer suggestions for choosing the loss function that best suits the purposes of road segmentation-related studies.