NOVEL SINGLE TREE DETECTION BY TRANSFORMERS USING UAV-BASED MULTISPECTRAL IMAGERY

Abstract. Single tree detection has been a major research topic concerning automatic forest inventory using remote sensing data. Recently, deep learning-based approaches in forestry have gained attention because of the prospect improved accuracy. In this study, we present novel method based on transformer (DETR), which applies combination with pre-trained convolutional neural network to detect individual trees high-resolution multispectral imagery. The test site (Kranzberg Forest Roof Experiment - KROOF) is located Bavaria, north Munich, and characterised by mixed consists large groups European beeches (Fagus sylvatica) surrounded Norway spruces (Picea abies). image data were acquired MicaSense RedEdge-MX Dual camera mounted UAV. Two flight mission conducted at an altitude around 85 m speed 5 m/sec, resulting ground resolution about cm. 125 surveyed tacheometric means field for testing, 1390 labelled visual interpretation imagery training validation. DETR shows promising results outperforms standard, well-known object YOLOv4 deciduous plots. More detailed, F1-scores evaluated coniferous plot 83%, 86% 71%. corresponding figures are 87% coniferous, 65% 67% deciduous. terms accuracy, inferior 6% plot, however superior 28% 5% respectively. Compared YOLOv4, found that sometimes failed small trees. Moreover, both methods tend over-detect single areas. sum, transformer-based great potential improve detection.