Unsupervised spectral-spatial processing of drone imagery for identification of pine seedlings

Abstract

Reliable and accurate monitoring of replanted forest areas is a vital component of the plantation cycle if sustainable management practices are to be maintained following resource extraction. Unmanned aerial vehicles (UAVs) can rapidly survey newly planted forest areas using high resolution imagery. Automated image processing techniques can cost-effectively examine large quantities of data. Correctly combined, such technologies can efficiently create accurate maps of seedling locations, which offers the prospect of management practices for forestry and ecology operations that involve less fieldwork. This paper presents a technique based on unsupervised machine learning for detecting healthy young pinus caribaea (Caribbean pine) and pinus radiata (Monterey pine) seedlings. The locations of individual trees requiring replacement are also determined based on the spatial distribution of the identified healthy trees. The approach was tested on data from 30 sites covering over 700 ha, with seedlings ranging from 9 months to 3 years’ old (mean tree heights 30–2 m). Test sites contained a range of geomorphologies and levels of weed infestation. The results indicate detection precision and specificity in excess of 90% and recall approaching 99% for dense point cloud resolution of 4–6 cm. As network pre-training is unnecessary, there is no need for high resolution imagery, which allows greater operational coverage per unit time. The algorithm may also be used on multi- or hyperspectral data; and may be useful if employed in conjunction with other supervised learning techniques to reduce the need for manual labelling of training data sets.