Advanced cloud detection using neural networks and optimization techniques
Techniques for automatically determining, on a pixel by pixel basis, whether imagery includes ground images or is obscured by cloud cover. The techniques include training a Neural Network, making an initial determination of cloud or ground by using the Neural Network, and performing a max-flow, min-cut operation on the image to determine whether each pixel is a cloud or ground imagery.
1. A method for identifying cloud imagery and ground imagery in an image, the method comprising:
determining, by a processor, preliminary classifications for each of multiple portions in the image as primarily cloud or primarily ground;
identifying, by the processor, adjacent pixels within the image;
calculating, by the processor, a capacity between the identified adjacent pixels, the capacity corresponding to similarity between adjacent pixels; and
determining, by the processor, which portions of the image include the cloud imagery and which portions of the image include the ground imagery based on the preliminary classifications and based on the capacity.
2. The method of claim 1 , further comprising:
creating a score for each pixel to represent the likelihood that the pixel does or does not contain a cloud; and
creating a grid-graph of the score of each pixel with adjacency information associated with each set of adjacent pixels.
3. The method of claim 2 , further comprising connecting the pixels of the grid-graph to at least one of a source or a sink using the score as the capacity, wherein one of the source or the sink represents the cloud imagery and the other of the source or the sink represents the ground imagery, and wherein determining which portions of the image include the cloud imagery and which portions of the image include the ground imagery is further based on the grid-graph.
4. The method of claim 1 , wherein determining the preliminary classifications is performed using a neural network.
5. The method of claim 1 , wherein determining the preliminary classifications is performed by reconstructing each of the multiple portions in a cloud dictionary and in a ground dictionary, and determining the preliminary classifications based on which reconstruction is more accurate.
6. The method of claim 1 , further comprising:
applying a window, having a height and width that are less than a height and width of the image and that are the same as that of each of the multiple portions of the image, at least some of the multiple portions partially overlapping adjacent portions, in order to determine if each portion most likely contains the cloud imagery or the ground imagery; and
incrementing or decrementing a score for each pixel in the portion based on whether the determination made by applying the window contains the cloud imagery or contains the ground imagery.
7. The method of claim 1 , wherein the image is a satellite image obtained using a satellite.
8. The method of claim 1 , further comprising adding to metadata associated with each pixel an indication of whether each such pixel includes the cloud imagery.
9. The method of claim 8 , further comprising selecting, by the processor, pixels for an orthomosaic that is free of clouds based on the metadata.
10. The method of claim 1 , wherein determining which portions of the image include the cloud imagery and which portions of the image include the ground imagery is performed by performing a min-cut/max-flow segmentation on the image to define portions of the image that are believed to include the cloud imagery and portions of the image that are believed to include the ground imagery.
11. A method for creating cloud-free satellite imagery, the method comprising:
determining, by a processor, preliminary classifications for each of multiple portions in an image as primarily cloud or primarily ground;
confirming, by the processor, which portions of the image include cloud imagery and which portions of the image include ground imagery using an algorithm based on the preliminary classifications;
adding, by the processor, to metadata associated with each pixel of the image an indication of whether each such pixel includes the cloud imagery based on the confirming; and
selecting, by the processor, pixels for an orthomosaic that is free of clouds based on the metadata.
12. The method of claim 11 , wherein confirming which portions of the image include the cloud imagery and which portions of the image include the ground imagery includes:
identifying, by the processor, adjacent pixels within the image;
calculating, by the processor, a capacity between the identified adjacent pixels, the capacity corresponding to similarity between adjacent pixels; and
determining, by the processor, which portions of the image include the cloud imagery and which portions of the image include the ground imagery based on the preliminary classifications and based on the capacity.
13. The method of claim 12 , wherein confirming which portions of the image include the cloud imagery and which portions of the image include the ground imagery includes:
creating a score for each pixel to represent the likelihood that the pixel does or does not contain a cloud; and
creating a grid-graph of the score of each pixel with adjacency information associated with each set of adjacent pixels.
14. The method of claim 13 , wherein confirming which portions of the image include the cloud imagery and which portions of the image include the ground imagery includes connecting the pixels of the grid-graph to at least one of a source or a sink using the score as the capacity, wherein one of the source or the sink represents the cloud imagery and the other of the source or the sink represents the ground imagery, and wherein confirming which portions of the image include the cloud imagery and which portions of the image include the ground imagery is further based on the grid-graph.
15. The method of claim 11 , wherein confirming which portions of the image include the cloud imagery and which portions of the image include the ground imagery includes:
applying a window, having a height and width that are less than a height and width of the image and that are the same as that of each of the multiple portions of the image, at least some of the multiple portions partially overlapping adjacent portions, in order to determine if each portion most likely contains the cloud imagery or the ground imagery; and
incrementing or decrementing a score for each pixel in the portion based on whether the determination made by applying the window contains the cloud imagery or contains the ground imagery.