Shoreline extraction from landsat-8 satellite imagery by using deep learning algorithms

Shorelines are under rapid change due to natural influences such as global warming, population increase of urbanization and environmental pollution. Temporal monitoring of coastal change has a critical importance for coastal resource management, protection of coastal areas and sustainable coastal planning. By using optical satellite images, current, accurate, temporal and reliable information about coastal areas can be obtained. Machine learning based Deep Learning algorithms can be used to extract shoreline from satellite images that enable automatic shoreline extraction. In this thesis, SegNet deep learning architecture, which is a semantic segmentation algorithm is used for shoreline extraction from LANDSAT-8 satellite imagery. SegNet architecture consists of encoder layers, decoder layers and a classification layer. The principle is that the low-resolution feature maps generated in the encoder layers match to the same resolution as the input image in the decoder layers. In the study, blue, red and near infrared bands of 34 LANDSAT-8 images have been used. 29 of them have been used to obtain training SegNet Deep Learning Network and 5 of them for testing for the results. NDWI (Normalized Difference Water Index) of LANDSAT-8 images were created to generate labels for training data set. After applying the threshold value to the NDWI images, binary images have been created and the labels have been set. 4880 data sets were prepared. Two different SegNet deep learning networks based on the VGG16 and VGG19 architectures were created in the MATLAB platform and the training process was carried out using prepared training set. Using SegNet Deep Learning Networks, Land and Water classes have been obtained in binary form from 5 test images. Binary images were converted into vector format and shorelines were obtained. Manual digitized shorelines were used as reference data for accuracy assessment. As a result of study, for each test data, the calculated average errors were 13.38 m (0.45 pixel), 16.16 m (0.54 pixel), 20.32 m (0.68 pixel), 9.99 m (0.33 pixel) and 12.63 m (0.42 pixel) for SegNet-VGG19 network respectively. The average errors for SegNet- VGG16 network were 9.78 m (0.33 pixels), 25.67 m (0.86 pixels), 17.07 m (0.57 pixels), 11.11 m (0.37 pixels) and 10.69 m (0.36 pixels) respectively. The results show that deep learning and SegNet semantic segmentation architecture can be used efficiently for shoreline extraction.