Associate shadows to their source in image analysis

Question

I've got two rasters. The first one corresponds to an cloud mask, that is an image with 1 when the pixel corresponds to a cloud and 0 when it does not. The second is a shadow mask, 1 if the pixel is classified as a shadow, else 0.

To reduce the error associated with the shadow classification, one can use the fact that a shadow must be associated to a cloud; there must be a shadow within x meters of each cloud in a certain direction (that can be retrieve thanks to the solar angle).

Any ideas on how to implement this association step?

Here is a snapshot of a raw image, the cloud mask (white) and the shadow mask (black) enter image description here

What exactly is your goal? Are you trying to create the shadow mask? Because it sounds like you start with it. If not, then what are you trying to classify? — Pace, Feb 11 '14 at 23:55
What I'm trying to do it to associate a shadow to its cloud to reduce the errors in the shadow mask — WAF, Feb 12 '14 at 07:42

score 1 · Answer 1 · answered Feb 12 '14 at 16:03

I've seen this done with a queue. Essentially the pseudo-code looks something like:

points = new list()
queue = new queue()
for x,y in image_coordinates:
  if is_cloud(x,y):
    point = new point(x=x, y=y, distance=0, cloud_x=x, cloud_y=y)
    queue.push(point)
  else:
    point = new point(x=x, y=y, distance=null, could_x=null, cloud_y=null)
  points.push(point)

while(!queue.is_empty()):
  point = queue.pop()
  for neighbor in point.neighbors():
    if angle_is_correct(point.cloud_x, point.cloud_y, neighbor.point.x, neighbor.point.y):
      if neighbor.is_direct(): //N,S,E,W
        new_distance = point.distance + 1
      else: //NE, SE, SW, NW
        new_distance = point.distance + SQRT_2
      if neighbor.point.distance == null or neighbor.point.distance > new_distance:
        neighbor.point.distance = new_distance
        neighbor.point.cloud_x = point.cloud_x
        neighbor.point.cloud_y = point.cloud_y
      queue.push(neighbor.point)

When the run is finished points will be a list of x,y coordinates, their distance to the nearest cloud, and the x,y coordinate of the nearest cloud (which you probably don't need). You should use the angle_is_correct function to make sure to only consider clouds which are in the correct direction. You may also be able to further optimize it to stop adding points to the queue if their distance exceeds the max distance.

I'm not entirely sure the algorithmic complexity but I suspect one could devise a proof to show this is O(n) or O(log(n)). All I know is that it worked quickly for me when I needed it.

Associate shadows to their source in image analysis

1 Answers1