Classifying unmanned aerial vehicle images for urban vegetation mapping utilizing SVM

Abstract

This study focuses on the potential of sUAVs for mapping urban vegetation. The researchers compared the effectiveness of maximum likelihood and SVM algorithms for classification purposes. Additionally, they tested different window sizes to determine the optimal size for calculating textural indices. An ortho-mosaic image was used to analyze the vegetation. A total of 748 images were collected from a height of 100 meters using a low-cost UAV, resulting in a resolution of 2.56 cm per pixel. To ensure accurate results, a high overlap of 90% forward and 80% side overlap was maintained to minimize vegetation masking by tall buildings. Ground control points were collected using GPS RTK technology, and all images were processed using Agisoft PhotoScan v1.27 software with a root mean square error of 0.2 pixels. Eight textural indices, including mean, standard deviation, homogeneity, contrast, dissimilarity, entropy, correlation, and angular second moment were extracted using gray-level co-occurrence matrix (GLCM). These texture indices were calculated using six different window sizes ranging from 3×3 to 45×45. The findings of this study will contribute to the understanding of sUAV-based remote sensing for mapping urban vegetation and provide insights into the most effective classification algorithms and window sizes for calculating textural indices.