Aprendizagem profunda na segmentação semântica de rios em imagens de alta resolução espacial

Abstract

Mapping rivers is essential for diagnosis and environmental planning, and remote sensing has helped in research related to this water resource. A method that has gained attention and helped in the task of mapping rivers is the deep learning method. The most used deep learning methods in the mapping task (semantic segmentation) are based on convolutional network architectures (CNN - Convolutional Neural Networks) and, more recently, Vision Transformers. They are robust methods that have shown good performance for this type of task, however, there are still challenges to be overcome when mapping features of narrow rivers. In this sense, considering the diversity of river features and the absence of an approach that goes deeper into their characteristics such as their width, the objective of this study is to map rivers with different sizes in RGB aerial images of high spatial resolution using deep segmentation networks semantics. For this, the present work sought to evaluate the performance of deep semantic segmentation networks when identifying large rivers (width greater than 10 m) in RGB images of high spatial resolution; characterize the capabilities and limits of deep semantic segmentation networks to identify narrow rivers (width less than 10 m) in high spatial resolution RGB images; and to evaluate the performance of deep semantic segmentation networks experimented in Vision Transformer to segment rivers, comparing them with traditional neural networks experimented in convolutions. The aerial images have 1 m of resolution and were trained and tested in a computational environment, using semantic segmentation methods based on convolution and Vision Transformer. Then, the results were compared qualitatively, visually, and quantitatively, using evaluation metrics (Accuracy, F1-Score, Precision, Recall, IoU). The results appreciated that the performance of deep semantic segmentation networks varies according to the width of the rivers. The best approach to segmenting large rivers (greater than 10 m) and narrow rivers (less than 10 m) is to use both river widths to train the deep semantic segmentation networks. Another finding is that the deep networks in Vision Transformers outperformed the deep networks in convolutions for the semantic segmentation task of large and narrow rivers in RGB images of 1 m resolution. SegFormer outperformed as evaluation indicators for river segmentation with an F1-Score above 98%. Future work should continue to investigate the aspiring deep networks in Vision Transformers, being able to explore multitemporal and multimodal images, in order to improve the monitoring of water resources.