Building footprint generation by using clean mask generation and received image data

1. A system comprising:

an image data source storing image data from a plurality of images;

a height map source storing height maps for areas of interest;

a building footprint module;

a memory storing program instructions; and

a building footprint processor, coupled to the memory, and in communication with the building footprint module and operative to execute the program instructions to:

receive image data for an area of interest (AOI);

receive a height map for the AOI;

execute a building segmentation module to generate a pixel-by-pixel building mask that indicates a presence of one or more buildings in the AOI;

apply at least one clean mask process to the generated pixel-by-pixel building mask to generate a clean mask, wherein for each pixel in the pixel-by-pixel building mask, the clean mask process: confirms a pixel assignment in a case the pixel assignment matches a predicted pixel assignment, and changes the pixel assignment in a case the pixel assignment does not match the predicted pixel assignment, to generate the clean mask;

receive the clean mask at an instance building segmentation module; and

execute the instance building segmentation module to generate at least one building footprint based on the clean mask and the received image data.

2. The system of claim 1 , further comprising program instructions to:

assign a building identifier to each generated building footprint.

3. The system of claim 1 , wherein the image data is data from multi-spectral images.

4. The system of claim 3 , further comprising process steps to:

extract red wavelengths, green wavelengths and blue wavelengths from the multi-spectral images to generate a red-green-blue (RGB) image.

5. The system of claim 1 , wherein the pixel-by-pixel building mask is a binary label representation of each pixel in the image data as a building or non-building.

6. The system of claim 1 , wherein the at least one clean mask process is a morphological operation, a small object removal filter, a small hole removal filter and a spectral analysis operation.

7. The system of claim 1 , wherein the instance building segmentation module includes a direction neural network and a distance transform neural network.

8. The system of claim 7 , wherein the instance building segmentation module is operative to predict whether each indicated building in the clean mask is one building or more than one building.

9. The system of claim 7 , wherein each of the direction neural network and the distance transform neural network are convolutional neural networks.

10. The system of claim 7 , wherein the direction neural network determines a direction to an edge of a building pixel group for each pixel in the building pixel group in the clean mask and generates a direction mask.

11. The system of claim 10 , wherein the direction mask provides an indication of the direction of a nearest boundary.

12. The system of claim 10 , wherein the distance transform neural network is operative to receive the direction mask and predict a distance from each pixel to the edge of the building pixel group.

13. A method comprising:

receiving multi-spectral image data for an area of interest (AOI);

receiving a height map for the AOI;

executing a building segmentation module to generate a pixel-by-pixel building mask that indicates a presence of one or more buildings in the AOI;

applying at least one clean mask process to the generated pixel-by-pixel building mask to generate a clean mask, wherein for each pixel in the pixel-by-pixel building mask, the clean mask process: confirms a pixel assignment in a case the pixel assignment matches a predicted pixel assignment, and changes the pixel assignment in a case the pixel assignment does not match the predicted pixel assignment, to generate the clean mask;

receiving the clean mask at an instance building segmentation module; and

executing the instance building segmentation module to generate at least one building footprint based on the clean mask and the received image data.

14. The method of claim 13 , further comprising:

assigning a building identifier to each generated building footprint.

15. The method of claim 13 , further comprising:

extracting red wavelengths, green wavelengths and blue wavelengths from the multi-spectral images to generate a red-green-blue (RGB) image.

16. The method of claim 13 , wherein the instance building segmentation module includes a direction neural network and a distance transform neural network, wherein the instance building segmentation module is operative to predict whether each indicated building in the clean mask is one building or more than one building.

17. The method of claim 16 , wherein the direction neural network determines a direction to an edge of a building pixel group for each pixel in the building pixel group in the clean mask and generates a direction mask.

18. A non-transient, computer-readable medium storing instructions to be executed by a processor to perform a method comprising:

receiving multi-spectral image data for an area of interest (AOI);

receiving a height map for the AOI;

executing a building segmentation module to generate a pixel-by-pixel building mask that indicates a presence of one or more buildings in the AOI;

applying at least one clean mask process to the generated pixel-by-pixel building mask to generate a clean mask, wherein for each pixel in the pixel-by-pixel building mask, the clean mask process: confirms a pixel assignment in a case the pixel assignment matches a predicted pixel assignment, and changes the pixel assignment in a case the pixel assignment does not match the predicted pixel assignment, to generate the clean mask;

receiving the clean mask at an instance building segmentation module; and

executing the instance building segmentation module to generate at least one building footprint based on the clean mask and the received image data.

19. The medium of claim 18 , further comprising:

assigning a building identifier to each generated building footprint.

20. The system of claim 1 , further comprising program instructions to:

assign a label to pixels that are not marked in the pixel-by-pixel building mask,

wherein the assignment is based on neighboring pixel areas.