Use of machine learning methods in automatic building extraction from UAV images

In this thesis, high-resolution data obtained through photogrammetric methods using Unmanned Aerial Vehicles (UAVs) were utilized to analyze the classification performance of machine learning methods for building detection. RGB bands were employed as the primary data source, supplemented with Digital Elevation Model (DEM) and Digital Surface Model (DSM) data to evaluate the effects of different data combinations. Building detection was performed using Support Vector Machines (SVM), Decision Trees (DT), and Artificial Neural Networks (ANN). The performance of ANN was optimized with heuristic algorithms, including Artificial Bee Colony (ABC), Particle Swarm Optimization (PSO), and Differential Evolution (DE). While RGB bands alone provided limited accuracy, the integration of DEM and DSM significantly improved the results, with the RGB + DEM + DSM combination achieving the highest accuracy. ANN models trained with heuristic algorithms achieved higher accuracy, recall, and specificity compared to traditional methods, with the ABC algorithm delivering the best performance across all data combinations.