Determination of soil diversity with artificial neural networks and k nearest neighbor algorithm

Ground penetrating radar (GPR) systems, one of the methods used for the detection of buried structures on the ground surface, are mainly used in military and civilian areas as well as in many different areas. GPRs perform scanning by detecting the reflected waves along with the transmission of the magnetic wave. Developed for the modeling of GPRs in a computer environment, gprMax is designed as open-source computer software. There are A, B and C scanning models produced by the GprMax program. A-scan image is used for the data obtained from a single scan modeling when the antennas in ground penetrating radar systems are in a fixed position, and B-scan image expressions are used for the sum of the sections collected by A-scans when they are in different positions. It is not possible to classify the category of an object correctly with a small number of classification examples for the classification of a structure or any object buried on the ground surface. Different training samples are needed to obtain a classification with a high success rate. Even if a successful classifier training takes place in accordance with the structure of the soil type (e.g dry sand), a change in the type of soil may cause the classifier to fail. As a result, the classification type to be obtained in the study was used and trained with many data types (e.g dry sand, wet sand, moist sand). In this study, by creating models with artificial neural networks (ANN) and K nearest neighbor algorithm (k-NN) with the help of gprMax software, it has been seen that classification of a single soil type will be an unsuccessful classification type, but classification can produce better results as a result of increasing data diversity. In the ANN and k-NN models used within the scope of the thesis, 3000 gprMax B-images were analyzed via Matlab and the results were used as input. Statistical analyzes were conducted in order to determine the most suitable architecture in ANN models. The data obtained in this analysis were divided into three different training and testing categories, and modeling was conducted using three different types of layers within each category and three different learning algorithms, two different activation functions, and different numbers of neurons in these layers. In K-NN models, on the other hand, three different data sets and predictions were categorized, and five different k neighborhood values were determined in each category. Statistical analysis and evaluations and performance results were compared between 270 different ANN models and 15 different k-NN architectures. Key Words: GprMax, Ground Penetrating Radar, Artificial Neural Networks, K- Nearest Neighbor Algorithm