Your data is not in state space form (that I could easily imagine as applicable for the Kalman filter). Eg your state could be (x, y) coordinates of the root tips, and then perhaps you could use a random walk or an I(1) process (“velocity” changes randomly), but those coordinates would need to be extracted first from the pixels. AFAICT that is orthogonal to the application of the Kalman filter. (BTW, amazing stuff).