those location points. Csillik used the known sample classes
to teach the CNN to differentiate between classes by learning
the specific features of each class. A quick study, the CNN
completed its training in 13 minutes.
Csillik then tested the newly trained CNN to identify and
delineate individual citrus trees. Each 40x40 pixel region
was submitted to the network to obtain a likelihood that
this image patch contains a citrus tree. Moving the analysis
region like a sliding-window across the orthomosaic, it took
eCognition's CNN two minutes to produce a probability "heat
map", a grayscale image of likely tree locations.
In a final step, the software used another newly integrated
feature called superpixel segmentation to further process
the heat map to remove multiple crown-detection errors
and refine its tree delineation, improving the detection and
location of single citrus trees.
In total, eCognition analyzed, identified and delineated
3,105 individual trees in 30 minutes with 96.2 percent
accuracy.
Left: Tree detection analysis. White crosses indicate automated
tree locations and classification using eCogniation.
Bottom left: Initial CNN-generated probability heatmap of trees
(black to white) and its refined stages to produce results that
better matched the ground reference samples (i).
Above: Study area location near Visalia, California (left)
and a false color image acquired by the UAS with a spatial
resolution of 0.12 m and covering 64.6 ha (right)