Improving of global digital elevation models with İCESat-2 and GEDI data

Digital Elevation Models (DEMs) are a crucial data source in many professional disciplines, with the with the aid of gridded elevation data. Because of this, numerous DEMs have been produced and continue to be produced, whether for pay or without charge, locally or globally. However, global DEM (KSYM) data contains various errors due to data structures or DEM production steps. To eliminate these errors, data improvement methods have been developed by using different DEMs together. In this study, it is aimed to improve KSYM data (AW3D30, ASTER KSYM and SRTM) with space-based LiDAR altimeter data (GEDI and ICESat-2). In the thesis study, firstly, the DEM production performance of GEDI and ICESat-2 data from space-based LiDAR altimeter systems, which started collecting data in late 2018, was investigated. In this instance, Puerto Rico, New Zealand, and the United States were chosen as test regions. Airborne LiDAR data was used as reference data in the test the DEM accuracy of the study areas. In the test areas, DEMs were produced by using the Kriging interpolation method and the results were compared with AW3D30, ASTER GDEM, and SRTM data. While it was observed that the accuracy of the DEMs created by using GEDI and ICESat-2 data independently was low, higher accuracy was obtained in the models created by using ICESat-2 and GEDI data together. In particular, it has been determined that there is a high correlation of 99% between the reference data and the produced DEM data. In the second stage of the thesis study, five different test areas were selected to improve the Global DEMs data (AW3D30, ASTER GDEM, and SRTM) with space-based LiDAR altimeter data, which is the main purpose of the study. These areas are Istanbul and Ankara provinces of the Türkiye, in addition to the other three areas used in DEM production. GNSS data was used as ground truth in Istanbul and Ankara data. Three different methods named Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN) and Extreme Gradient Boosting Machines (XGBoost) were used to improve the accuracy of GDEM data with GEDI and ICESat-2 data. Considering the point-based accuracies of GEDI and ICESat-2 data, it varies between 6.48 m and 11.29 m. According to these results, while GEDI data had worse results than GDEMs only in test area-3, higher vertical accuracy was obtained compared to GDEM data in remaining areas. Considering the DEM improvement potential, GDEMs have been improved in all areas. The greatest improvement in overall accuracy was achieved with the CNN method on ASTER KSYM data with 4.35 m in test area-4 according to RMSE. According to land cover classes, the best improvement was obtained in rangeland class in ASTER GDEM data with 4.72 m using the CNN method in test area-3 according to RMSE. The best improvement in the slope group classes was obtained in the 60->% slope group in ASTER GDEM data with 9.77 m using the CNN method in test area-3 according to RMSE. When the results of the study were examined in terms of considered methods, it was concluded that the most successful method was the CNN method. The best accuracy improvement success in land cover classes was achieved in the forest class. In slope groups, better results were obtained at higher slopes.