Keywords: plant phenotyping, UAS, image analytics, gridding,
open-source software, Python.
INTRODUCTION
Plant phenotyping is a vital practice in plant breeding and
biotechnology to discover a new variety or characterize desired genetic
traits resistant to biotic and abiotic stresses. High throughput
phenotyping (HTP) is essential to meet the timely delivery of the
phenotypic metrics of a large number of plots. Conventional phenotyping
is made by notetaking to score the plant health conditions, which is
laborious and inconsistent. Image-based HTP was deployed to replace
manual scoring and achieve the rapid extraction of phenotypic metrics
using ground and aerial platforms. Plot-level metrics is required for
plant phenotyping on a large number of plots and extracted by defining a
region of interest (ROI) of the field boundary and processing sub-ROIs
aligned with rows and columns of the total number of plots, called
gridding. Grid-based image processing is offered by commercial software
(e.g., ArcGIS) but is limited to upright rectangular fields and manual
drawing of polygons.
Unmanned aircraft system (UAS) has been widely used to deliver a massive
volume of images in high resolution and extract metrics by examining
spectral signature and morphological features of the plants. Raw UAS
images are preprocessed for orthomosaicing through global positioning
system (GPS) and inertial measurement unit (IMU) information using
commercial software (e.g., Pix4Dmapper, Agisoft). The orthomosaiced
image is georeferenced to align the image top to north, whereas the
field orientation is often off the north, resulted by various field
layout and declination of magnetic north (used by IMU) from geographic
north up to 20 degrees [1]. In practice, no fields are truly
oriented to north. This misalignment of the field orientations occurs to
any airborne images not only from UASs but also from manned airplanes or
satellites, as their image products are all georeferenced. Due to the
misaligned orientation, gridding requires a preprocess of image rotation
to make the field orientation upright to be aligned with the grid
[2], because the computation pattern is sequenced by row (i) and
column (j) in image coordinates and performs metrics extraction based on
upright rectangular ROIs. Finding a rotation degree, however, takes
multiple adjustments to precisely align sub-ROIs with plots across the
field, which is a laborious time-consuming task and leads to a heavy
computational load especially on the big-sized (e.g., >1
GB) orthomosaic image. Image rotation also creates resampling errors due
to the changes of pixel values in repositioned geometry. To solve this
issue, gridding method must be generalized for the various field
orientations and sizes without changing the original image. The grid
rotation and metrics extraction on the rotated sub-ROIs are key
challenges in image processing.
This adaptive gridding method will help creating a GIS interface of the
grid by converting sub-ROIs to a shapefile that contains a list of plot
polygons with GPS coordinates.
The objective of the study is to develop open-source software that
provides a quick extraction of plot-level metrics of the field image
without the image rotation. Specific objectives are to 1) develop
algorithm to create a rotated grid that fits the field and plot
boundaries aligned with all sub-ROIs, 2) to extract metrics on the
rotated ROIs by geofencing algorithm, 3) publish the software available
to the public that automates the adaptive gridding process in graphic
user interface (GUI).
MATERIALS AND METHODS
UAS images are collected for field mapping and registered with GPS
coordinates and IMU data based on flight waypoints and signal
communication (Fig. 1). The UAS receives GPS signals from satellites and
correction signals from a real-time kinematic (RTK) base station or
NTRIP (Networked Transport of RTCM via Internet Protocol) service. The
raw UAS images are preprocessed for stitching tile images to an
orthomosaic image in field level.