Prediction of elevation points using three different heuristic regression techniques

Abstract

The aim of this study is to estimate the elevation points used in the creation of the digital elevation model, which is the most important data of the projects and required in the engineering project, using horizontal and vertical location informations and three different heuristic regression techniques. As the study area, an area with mid-level elevations, located in the Marmara region, and covering a part of the intersection of Edirne, Kırklareli and Tekirdağ provinces was chosen. In the study, the estimations were investigated for three different sized areas, and these areas are square areas with the dimensions of 1x1 km, 10x10 km and 100x100 km, respectively. A total of 3500 elevation points were used in the study, and this number is constant in all areas, and 60% of these points were used in the testing phase and 40% in the training phase. The models used in the study are M5 model tree (M5-tree), multivariate adaptive regression curves (MARS) and Least Square Support Vector Regression (LSSVR). The results of the models were evaluated according to three different comparison criteria. These, coefficient of determination (R2), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) were used. When the modeling results are examined; M5-Tree regression method gave the best results (1), LSSVR method was better than MARS methods (2), The most successful input data was found in datasets using X and Y coordinates information, and the worst results were found in datasets using X coordinates (3). As the study area increased, the model performance did not improve (4). The least error was obtained in the modeling of 1x1 km area, and the highest R² was obtained from the modeling of 10x10 km area (5). It was concluded that the M5-tree method is a very successful method in height modeling.