Global warming, increasing population, environmental pollution and urbanization can constantly affect coastal areas. Therefore, sustainable monitoring of coastal zones is vital to detect changes which can occur due to natural and anthropogenic effects. Thus, sustainable shoreline monitoring is essential for coastal resource management, environmental protection and planning. Satellite images provide accurate, reliable, temporal and up-to-date information for this purpose. State-of-the-art deep learning (DL) and transfer learning approaches brought new opportunities for shoreline extraction. In this study, a transfer learning based water-body segmentation framework with U-Net architecture from SENTINEL-2 imagery has been proposed. The pre-trained weights have been obtained from another study which is a network trained with LANDSAT-8 imageries. The training of used U-Net architecture was carried out using SENTINEL-2 imagery which consists of blue, red and NIR bands with 8 and 7 full frames for training and testing, respectively. Images have been cropped as 512x512 pixels and 115 and 235 patches have been created for the training and testing dataset, respectively. Average accuracy, recall, precision, specivity and F-score of the model values has been calculated as 0.9917, 0.9927, 0.9908, 0.9907 and 0.9917, respectively. The results show that it is possible to obtain shoreline with high accuracy with limited data using transfer learning.
Benzer Makaleler | Yazar | # |
---|
Makale | Yazar | # |
---|