Machine learning techniques for identifying clouds and cloud shadows in satellite imagery
Systems and methods for identifying clouds and cloud shadows in satellite imagery are described herein. In an embodiment, a system receives a plurality of images of agronomic fields produced using one or more frequency bands. The system also receives corresponding data identifying cloud and cloud shadow locations in the images. The system trains. a machine learning system to identify at least cloud locations using the images as inputs and at least data identifying pixels as cloud pixels or non-cloud pixels as outputs. When the system receives one or more particular images of a particular agronomic field produced using the one or more frequency bands, the system uses the one or more particular images as inputs into the machine learning system to identify a plurality of pixels in the one or more particular images as particular cloud locations.
1. A system comprising:
one or more processors;
a memory storing instructions executable by the one or more processors to cause the system to:
receive a plurality of images of agronomic fields produced using one or more frequency bands;
receive data identifying cloud locations and cloud shadow locations in the plurality of images;
train a machine learning system to identify cloud locations using the plurality of images as inputs and data identifying pixels as cloud pixels or non-cloud pixels as outputs;
receive one or more particular images of a particular agronomic field produced using the one or more frequency bands;
use the one or more particular images as inputs into the machine learning system, identify a plurality of pixels in the one or more particular images as particular cloud pixels;
generate a subset of the particular images, wherein the subset of the particular images comprises images containing at least a threshold percentage of pixels identified by the machine learning system as particular cloud pixels; and
remove the subset of the particular images from the one or more particular images to generate a set of modeling images containing less than the threshold percentage of pixels identified by the machine learning system as particular cloud pixels.
2. The system of claim 1 , wherein the machine learning system is a convolutional encoder-decoder comprising a plurality of convolutional encoding steps interspersed with pooling steps and a plurality of convolutional deciding steps interspersed with upsampling steps.
3. The system of claim 1 , wherein the machine learning system is a convolutional encoder-decoder that has been customized to accept inputs of different types using a customized machine learning layer which shapes a build of a vector based on a symbolic shape of an output tensor.
4. The system of claim 1 , wherein each input of the inputs used to train the machine learning system comprise a plurality of stacked matrices, each of the plurality of stacked matrices generated from images captured using different frequency bands.
5. The system of claim 1 , wherein the threshold percentage of pixels is thirty percent.
6. The system of claim 1 , wherein the memory further stores instructions executable by the one or more processors to cause the system to generate an agronomic yield model of a field, wherein the agronomic field model is configured to model a yield of a crop based on the modeling images in the particular images, and wherein the agronomic yield model is a function of at least an average normalized difference vegetation index (NDVI) value or other index values across the field.
7. The system of claim 1 , wherein the memory further stores instructions executable by the one or more processors to cause the system to:
identify at least one ideal image of a field from the particular images, wherein the at least one ideal image does not contain any cloud pixels; and
display the ideal image on a computing device.
8. The system of claim 7 , wherein the memory further stores instructions executable by the one or more processors to cause the system to:
receive a request from a user for display of a current image of the particular agronomic field;
generate a current image of the particular agronomic field;
determine that the current image contains more than the threshold percentage of pixels;
generate an overlayed image by overlaying at least a portion of the current image with at least a portion of the ideal image; and
display the overlayed image on a computing device.
9. A computer-implemented method comprising:
receiving a plurality of images of agronomic fields produced using one or more frequency bands;
receiving data identifying cloud locations and cloud shadow locations in the plurality of images;
training a machine learning system to identify cloud locations using the plurality of images as inputs and data identifying pixels as cloud pixels or non-cloud pixels as outputs;
receiving one or more particular images of a particular agronomic field produced using the one or more frequency bands;
using the one or more particular images as inputs into the machine learning system, identify a plurality of pixels in the one or more particular images as particular cloud pixels;
generating a subset of the particular images, wherein the subset of the particular images comprises images containing at least a threshold percentage of pixels identified by the machine learning system as particular cloud pixels; and
removing the subset of the particular images from the one or more particular images to generate a set of modeling images containing less than the threshold percentage of pixels identified by the machine learning system as particular cloud pixels.
10. The computer-implemented method of claim 9 , wherein the machine learning system is a convolutional encoder-decoder comprising a plurality of convolutional encoding steps interspersed with pooling steps and a plurality of convolutional deciding steps interspersed with upsampling steps.
11. The computer-implemented method of claim 9 , wherein the machine learning system is a convolutional encoder-decoder that has been customized to accept inputs of different types using a customized machine learning layer which shapes a build of a vector based on a symbolic shape of an output tensor.
12. The computer-implemented method of claim 9 , wherein each input of the inputs used to train the machine learning system comprise a plurality of stacked matrices, each of the plurality of stacked matrices generated from images captured using different frequency bands.
13. The computer-implemented method of claim 9 , wherein the threshold percentage of pixels is thirty percent.
14. The computer-implemented method of claim 9 , further comprising:
generating an agronomic yield model of a field, wherein the agronomic field model is configured to model a yield of a crop based on the modeling images in the particular images, and
wherein the agronomic yield model is a function of at least an average normalized difference vegetation index (NDVI) value or other index values across the field.
15. The computer-implemented method of claim 9 , further comprising identifying at least one ideal image of a field from the particular images, wherein the at least one ideal image does not contain any cloud pixels.
16. The computer-implemented method of 15 , further comprising:
receiving a request from a user for display of a current image of the particular agronomic field;
generating a current image of the particular agronomic field;
determining that the current image contains more than the threshold percentage of pixels;
generating an overlayed image by overlaying at least a portion of the current image with at least a portion of the ideal image; and
displaying the overlayed image on a computing device.