Dear all,
I would like to find a solution to separate green tree pixels from background using UAV images taken over forest area. Generally, forest floor/background is complex with containing soil, rocks, grass/shrub, moss and so forth. Moreover, many shaded ground areas casted by trees introduce more difficulty for image segmentation. All above mentioned induce big challenges for extracting green tree pixels from RGB even multispectral images.
Some years ago I published a conference paper (Fuzzy Logic Based Pesticide Sprayer For Smart Agricultural Drone) but in that case I detected the green surface in general; trees and grass. In your case; can you send me a sample of those images to see how may I help you "if I can"?
It might be worthwhile calculating additional bands (derived bands) representing the ratio between band pairs from the original imagery. Then stack these derived bands with your original bands before running your classification. This can prove useful when dealing with shadows.
The principle is that shadows represent a drop in light intensity received by the sensor (camera), but may not affect the ratio of intensities between bands. So R, G and B values for say grassy understory falling in the shadow of a tree canopy will be lower than that of grassy understory that is not in shadow. But, the ratio between bands, for example R/G, will tend to be similar between grassy understory under shadow and in full light.
Abdellatif Baba Sorry for the late. I have been rapid through your conference. The method you proposed seems unsuitable for my dataset. But of course, I am glad to share some photos with you to facillitate your test. Pleasw find them by the link: https://drive.google.com/open?id=18DipREcJfWandKQrFlnu8-_jEYxZ24I5.
Thanks in advance.
Linyuan
Varun Varma Hi, Varun. Thanks for your suggestion. The derived bands with ratio format might have capacity to alleviate the shadow effect. However, the largest problem inducing challenge for image segmentation is the color of background(grass) pixels is similar with that of tree crown pixels. Please see an example (screenshot file) attached in this post.
I tried Gabor filter + PCA; more specifically an array of Gabor Filters which are tuned to different wavelengths and orientations. One result is attached to this answer. I'm not sure if it is useful for you!
Big challenges indeed! Is it just my eyes that see trees, or is the concept of trees in my brain involved? If the former, you are lucky, if the latter, you might need an AI solution. Training a neural network will take considerable time...
Maybe you could try a band-filter in the frequency domain; the pine branches show some pattern.
Lambert Zijp Hi, Lambert. The concept of tree perfers the latter you mentioned. I will try to apply deep learning method to these images. Of course, training will cost a lot of time, especially when high accuracy is needed. Is it essential or useful to feed band-filtered image, in addition to R-G-B band, into a deep networks?
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing