Spectral signature mapping and deep learning detection of Frankincense (Boswellia Sacra) using worldview-3 imagery.

Abstract

The identification of plant species in arid ecosystems relies on spectral signatures that capture their reflectance properties. Despite the ecological and economic importance of the Frankincense tree (Boswellia sacra), its spectral characteristics remain underexplored in remote sensing applications. This study presents a spectral characterization of Frankincense using field spectra collected with an Analytical Spectral Devices (ASD) spectroradiometer (350-2,500 nm) and validated with WorldView-3 (WV3) multispectral imagery in Oman's Dhofar region. The ASD spectra were resampled to WV3 bands and integrated into supervised classification using the Spectral Angle Mapper (SAM) algorithm. SAM significantly outperformed NDVI thresholding, which overestimated vegetation cover and failed to distinguish Frankincense from surrounding vegetation. Accuracy assessment showed strong performance, with user's accuracy of 88.0%, producer's accuracy of 99.2%, overall accuracy of 93.7%, and a Kappa coefficient of 0.87. Three deep learning models (SSD, YOLOv3, and RetinaNet) were also evaluated for automated Frankincense detection. SSD achieved the best performance, detecting 701 trees with a precision of 0.83 and recall of 0.93. The findings demonstrate the potential of integrating field spectroscopy, high-resolution imagery, and deep learning for monitoring ecologically important vegetation in arid environments.