- We identified four pixel classes in our images: green (leaves), purple (leaves and flowers), residue, and soil (Figure 1, left window).
- To facilitate sampling pixels from images, we wrote a macro for ImageJ that records x and y position and red, green and blue color density for each pixel selected (Figure 1) (link to macro - "Color Saver").
- We sampled the color of 10-15 pixels per class per image for 2-4 images from a given batch (same plant species, images acquired on the same day). We intentionally selected pixels to represent the full range of each of these color classes.
- Samples of pixels from each class revealed differences in coloration that could be used to classify pixels (Figures 2).
- We graphed the sampled pixels by class using various combinations of colors and color ratios. We found a broadly similar pattern for most of the images we sampled, though the exact threshold values varied depending on crop growth, lighting conditions and exposure, and camera.
Figure 1
Screenshot of pixel sampling in ImageJ using the "Color Saver" macro
Figure 2 A
Green (and some purple) pixels can be separated by low red:green ratio. They also typically have intermediate brightness. (Remember, we aren't separating gren from purple, but these 2 classes are different enough that they don't easily separate from the rest using a single rule).
Figure 2 B
The remaining purple pixels have a low green:blue ratio, and intermediate brightness.
Figure 2 C
We can separate out plant residue and soil based on red color and brightness; soil is generally quite dark, and residue quite bright. In many cases there is a region where these classes overlap. We consider this are to be ambiguous, and leave these pixels unassigned.
