Evaluation of Deep Learning Transfer Techniques for Mangrove Segmentation with Images of the Sentinel-2A

Fine-tuning techniques allow the use of weights from pre-trained networks in other models across different contexts, potentially improving training performance as it generally requires fewer computational resources and less data. Finetuning has become more widespread in the natural domain (RGB) with the availability of pre-trained model weights from the ImageNet database. However, pre-trained models in the same domain are not readily available for the remote sensing domain, such as in mangrove identification. Both nationally and in the state of Paraná, there are few studies employing deep learning for mangrove segmentation. Developing models using deep learning transfer can help establish automated monitoring systems. Thus, this study evaluated fine-tuning techniques for mangrove segmentation in Paraná using the U-Net model with pre-trained encoders in the same domain, remote sensing, and the natural domain. The dataset for training the U-Net was generated using bands from the Sentinel-2A satellite and annotations from the MapBiomas project maps. The fine-tuned networks discussed in this study accurately identified mangroves in Paraná, all achieving accuracies above 95.1% and F-scores greater than 92.6%.