Quality of forest plantations using aerial images and computer vision techniques

Abstract

Geotechnology has provided several tools that allow the spatial and temporal variability of soils and plants to
be investigated, leading to the consolidation of Precision Agriculture. The great challenge for studies using sensors mounted
aboard Remotely Piloted Aircraft (RPA) lies in interpreting the high-dimensional data, since most sensors do not measure the
biometric parameters of a plant directly. Therefore, the aim of the present study was to develop a methodology for using digital
images (obtained by means of an airborne RGB sensor mounted aboard an RPA) in the quality control of forest plantations,
specifically Eucalyptus (Eucalyptus ssp.), planted in a commercial area. A Phantom 4 Pro multirotor RPA was used, equipped
with a 20 Megapixel RGB sensor, acquiring images with 80% and 60% longitudinal and lateral overlap, respectively. From
the generated orthomosaic, a Test Area was outlined to be used in developing the processing routine based on computer vision
techniques. In general, the proposed methodology maps the individual location of each plant in the orthomosaic, resulting in a
mesh that allows the automatic generation of report maps of various silvicultural variables, such as plant count, planting failures,
and spacing between rows and plants. In addition to high computer performance, with real-time processing, the methodology
was highly accurate in correctly identifying more than 93% of plants in an area of more than 3,000 plants.