Carbon estimation
Training an estimation model using soft labels includes receiving an image. It further includes generating a continuous target map corresponding to the image that includes hard labels and soft labels. A model is trained using the corresponding continuous target map.
1. A system, comprising:
a processor configured to:
receive an image comprising a plurality of pixels;
receive a partial label map including hard labels corresponding to a first subset of pixels of the image;
determine soft labels corresponding to a second subset of pixels of the image that is unlabeled, wherein determining a soft label for an unlabeled pixel in the second subset of pixels of the image comprises comparing a set of characteristics of the unlabeled pixel against a set of characteristics of a labeled pixel in the first subset of pixels of the image;
generate an output label map that includes both the hard labels included in the partial label map corresponding to the first subset of pixels of the image and the soft labels determined for the second subset of pixels of the image; and
train a model using both the image and the output label map that includes both the hard labels and the soft labels; and
a memory coupled to the processor and configured to provide the processor with instructions.
2. The system of claim 1 , wherein the soft label for the unlabeled pixel is determined based at least in part on a determined similarity of the unlabeled pixel to the labeled pixel in the first subset of pixels of the image.
3. The system of claim 2 , wherein a corresponding hard label of the labeled pixel is assigned as the soft label for the unlabeled pixel.
4. The system of claim 2 , wherein the similarity is determined based at least in part on a determined spectral similarity.
5. The system of claim 4 , wherein the spectral similarity is determined based at least in part on a cosine similarity.
6. The system of claim 1 , wherein the image comprises a multi-band satellite image.
7. The system of claim 1 , wherein during training of the model, a weight is applied to a contribution of the soft label in a loss function.
8. The system of claim 7 , wherein the weight applied to the contribution of the soft label in the loss function is different from a weight applied to a contribution of a hard label in the loss function.
9. The system of claim 7 , wherein the weight applied to the contribution of the soft label is adjusted based at least in part on a stage of the training.
10. A method, comprising:
receiving an image comprising a plurality of pixels;
receiving a partial label map including hard labels corresponding to a first subset of pixels of the image;
determining soft labels corresponding to a second subset of pixels of the image that is unlabeled, wherein determining a soft label for an unlabeled pixel in the second subset of pixels of the image comprises comparing a set of characteristics of the unlabeled pixel against a set of characteristics of a labeled pixel in the first subset of pixels of the image;
generating an output label map that includes both the hard labels included in the partial label map corresponding to the first subset of pixels of the image and the soft labels determined for the second subset of pixels of the image; and
training a model using both the image and the output label map that includes both the hard labels and the soft labels.
11. The method of claim 10 , wherein the soft label for the unlabeled pixel is determined based at least in part on a determined similarity of the unlabeled pixel to the labeled pixel in the first subset of pixels of the image.
12. The method of claim 11 , wherein a corresponding hard label of the labeled pixel is assigned as the soft label for the unlabeled pixel.
13. The method of claim 11 , wherein the similarity is determined based at least in part on a determined spectral similarity.
14. The method of claim 13 , wherein the spectral similarity is determined based at least in part on a cosine similarity.
15. The method of claim 10 , wherein the image comprises a multi-band satellite image.
16. The method of claim 10 , wherein during training of the model, a weight is applied to a contribution of the soft label in a loss function.
17. The method of claim 16 , wherein the weight applied to the contribution of the soft label in the loss function is different from a weight applied to a contribution of a hard label in the loss function.
18. The method of claim 16 , wherein the weight applied to the contribution of the soft label is adjusted based at least in part on a stage of the training.
19. A computer program product embodied in a non-transitory computer readable medium and comprising computer instructions for:
receiving an image comprising a plurality of pixels;
receiving a partial label map including hard labels corresponding to a first subset of pixels of the image;
determining soft labels corresponding to a second subset of pixels of the image that is unlabeled, wherein determining a soft label for an unlabeled pixel in the second subset of pixels of the image comprises comparing a set of characteristics of the unlabeled pixel against a set of characteristics of a labeled pixel in the first subset of pixels of the image;
generating an output label map that includes both the hard labels included in the partial label map corresponding to the first subset of pixels of the image and the soft labels determined for the second subset of pixels of the image; and
training a model using both the image and the output label map that includes both the hard labels and the soft labels.