Automatic shoreline extraction from LANDSAT 8 imageries with artificial neural networks and random forest methods

Coastal areas are always very important for human being because of their habitat. In addition to it, coastal areas provide vacation and relaxation opportunities because of their microclimatic features. Activities such as tourism, industry, aquaculture and urbanization are main threads for coastal areas. Therefore, monitoring of coastal areas one of the vital issues for preserving and sustainable management of environmental heritage. Shorelines are widely used in different areas such as natural environment management, disaster management, coastal erosion studies, solid matter transport and modeling of coastal morphodynamics. For this reason, many studies have been carried out using various methods in order to extract the shoreline from satellite images in particular. Due to their advantages, Artificial Neural Network (ANN) based methods became very popular for many scientists. In this thesis, shoreline extraction from satellite image was performed by using SOM, ANN and Random Forest (RF) methods. In this study, Terkos region has been chosen for testing of the proposed methods within the scope of "TUBITAK Project (Project No: 115Y718) titled" Integration of Unmanned Aerial Vehicles for Sustainable Coastal Zone Monitoring Model – Three-Dimensional Automatic Coastline Extraction and Analysis: Istanbul-Terkos Example ". Landsat 8 imageries have been used to implement proposed algorithms. 5 Landsat 8 images have been used for training which were from different part of Black Sea region of Turkey and taken in the year of 2017. 6 Landsat 8 images were used for testing of proposed methods First data set is from Terkos-Istanbul in the years of 2013, 2015, 2017. Second one is from province of Antalya which was taken in the year of 2017, third one is from province of İzmir which was taken in the year of 2017 and last data set is from Lake Ercek which was taken in the year of 2017. NIR-RED-BLUE bands of all test data have been used for this thesis. In the first phase of the thesis, using SOM method, satellite image of Terkos region is clustered as two clusters (land – sea) and shoreline has been extracted from the clustered image. To train the ANN method, 5 different satellite images belonging to the Black Sea Region were used. From these images, sample data were collected to form 2 classes (land – sea). The same training areas have been used for both ANN and RF methods. The main aim of this thesis is to collect automatically using results of SOM method and to train ANN model and to construct tree structures for the RF method. In the second phase of the thesis, the training data-set have been classified to obtain land and sea classes. Following to this step, training of artificial network and generating of the tree structure was carried out using randomly selected data from the results of SOM. After training, SOM, ANN and RF methods have been implemented to the test images. The artificial neural networks used in the first and second steps were consisted of single hidden layer and 2000 iterations by combination of Levenber-Marquardt (TRAINLM) and Scale Conjugate Gradient (TRAINSCG) training functions with Hyperbolic Tangent Sigmoid (TANSIG) and Logistic Sigmoid (LOGSIG) transfer functions. In the final step, new networks were designed using the TRAINSCG - TANSIG combination with different numbers of hidden layers (5, 10, 15) and iteration numbers (500, 1000, 1500, 2000). The manually selected training data and results of SOM method has been used separately as training data set for all combinations. First of all, to test the defined combinations of the proposed methods were applied on Istanbul-Terkos, 2017 Landsat 8 image. The shoreline obtained by the SOM Method in the first step was compared with the manual digitization and the average error was calculated as 0,49 pixels. Among the artificial neural networks trained by user - collected training data, the TRAINSCG - TANSIG combination gave the best result with an average error of 0,61 pixels. In the results of the RF method using the same training data, which consists of 50 trees, average shoreline extraction error was 0,45 pixels. Based on the clustering results of the SOM Method at the second step, shoreline extraction error was calculated as 0,22 pixels by using TRAINSCG - TANSIG combination. The same training data were used for RF method and created 250 trees. As a result, average error was calculated as 0,79 pixels. In the third step, the best resultant network from the user-collected training data provided 10 hidden layers and 1000 iterative artificial neural networks was used and average shoreline extraction error was calculated as 0,36 pixels. The use of SOM results and by selecting of randomly training data, the ANN configuration with 5 hidden layers and 1000 iteration, the shoreline extraction error was calculated as 0.20 pixels. After this step, by considering of best combinations of proposed methods, using additional 6 images, methods were tested again. For each image accuracy assessment was realized by comparison of manual digitizing results with obtained results. The shoreline was extracted with an average error of 0,47 pixels from the LANDSAT 8 image of the Terkos region in the year of 2013. In the same way the shoreline extracted with an average error of 0,36 pixels from the LANDSAT 8 image of the Terkos region in the year of 2015. For Istanbul-Terkos, 2017 Landsat 8 image, average extraction error was calculated as 0.22 pixels. The shorelines were extracted with an average error of 0.57 pixels from the LANDSAT 8 image of Ercek Lake in the year of 2017, an average error of 0.31 pixels from the LANDSAT 8 image of the Izmir region in the year of 2017 and an average error of 1.31 pixels from the LANDSAT 8 image of the Antalya region in the year of 2017.