Field segmentation and classification
Implementations relate to improved crop field segmentation and crop classification in which boundaries between crop fields are more accurately detected. In various implementations, high-elevation image(s) that capture an area containing multiple demarcated fields may be applied as input across one or more machine learning models to generate a boundary enhancement channel. Each pixel of the boundary enhancement channel may be spatially aligned with a corresponding pixel of the one or more high-elevation images. Moreover, each pixel of the boundary enhancement channel may be classified with a unit angle to a reference location of the field of the multiple demarcated fields that contains the pixel. Based on the boundary enhancement channel, pixel-wise field memberships of pixels of the one or more high-elevation images in the multiple demarcated fields may be determined.
1. A method implemented using one or more processors, comprising:
obtaining one or more satellite images that capture an area containing multiple demarcated fields;
applying the satellite images as input across one or more machine learning models to generate a boundary enhancement channel, wherein each pixel of the boundary enhancement channel is spatially aligned with a corresponding pixel of the one or more satellite images, and wherein each pixel of the boundary enhancement channel is classified with a unit angle to a reference location of the field of the multiple demarcated fields that contains the pixel, wherein the reference location comprises a nearest boundary of the field that contains the pixel; and
based on the boundary enhancement channel, determining, and storing in memory, pixel-wise field memberships of pixels of the one or more satellite images in the multiple demarcated fields.
2. The method of claim 1 , further comprising upsampling the boundary enhancement channel to enhance spatial resolution of at least some boundaries between the multiple demarcated fields.
3. The method of claim 1 , wherein the determining comprises clustering pixels of the boundary enhancement channel together along boundaries where the unit angles align.
4. A system comprising one or more processors and memory storing instructions that, in response to execution of the instructions by the one or more processors, cause the one or more processors to:
obtain one or more satellite images that capture an area containing multiple demarcated fields;
apply the satellite images as input across one or more machine learning models to generate a boundary enhancement channel, wherein each pixel of the boundary enhancement channel is spatially aligned with a corresponding pixel of the one or more satellite images, and wherein each pixel of the boundary enhancement channel is classified with a unit angle to a reference location of the field of the multiple demarcated fields that contains the pixel, wherein the reference location comprises a nearest boundary of the field that contains the pixel; and
based on the boundary enhancement channel, determine, and store in the memory, pixel-wise field memberships of pixels of the one or more satellite images in the multiple demarcated fields.
5. The system of claim 4 , further comprising instructions to upsample the boundary enhancement channel to enhance spatial resolution of at least some boundaries between the multiple demarcated fields.
6. The system of claim 4 , wherein determination of the pixel-wise field memberships includes clustering pixels of the boundary enhancement channel together along boundaries where the unit angles align.
7. A method implemented using one or more processors, the method comprising:
obtaining one or more satellite images that capture an area containing multiple demarcated fields, wherein the one or more satellite images each include a plurality of spectral channels;
applying the satellite images as inputs across one or more machine learning models to generate a boundary enhancement channel, wherein each pixel of the boundary enhancement channel is spatially aligned with a corresponding pixel of the one or more satellite images, and wherein each pixel is classified with an angle to a reference location of the field of the multiple demarcated fields that contains the pixel, wherein the reference location comprises a nearest boundary of the field that contains the pixel;
comparing the boundary enhancement channel to ground truth segmentation data; and
based on the comparing, training one or more of the machine learning models.
8. The method of claim 7 , wherein the ground truth segmentation data comprises one or more of crop data layer (“CDL”) data, common land unit (“CLU”) data, crop map of England (“CROME”) data, or European Union Land Parcel Identification System (“LPIS”) data.
9. A method implemented using one or more processors, comprising:
obtaining one or more satellite images that capture an area containing multiple demarcated fields;
applying the satellite images as input across one or more machine learning models to generate a boundary enhancement channel, wherein each pixel of the boundary enhancement channel is spatially aligned with a corresponding pixel of the one or more satellite images, and wherein each pixel of the boundary enhancement channel is classified with a unit angle to a reference location of the field of the multiple demarcated fields that contains the pixel; and
based on the boundary enhancement channel, determining, and storing in memory, pixel-wise field memberships of pixels of the one or more satellite images in the multiple demarcated fields;
wherein the one or more machine learning models comprise first and second convolutional neural networks in series, wherein output of the first convolutional neural network is applied as input across the second convolutional neural network; and
wherein output of one of the first and second convolutional neural networks comprises pixel-wise crop type classifications, and output of the other of the first and second convolutional neural networks comprises the pixel-wise field memberships.
10. The method of claim 9 , wherein the one or more machine learning models comprise an atrous spatial pyramid pooling (“ASPP”) layer.
11. A system comprising one or more processors and memory storing instructions that, in response to execution of the instructions by the one or more processors, cause the one or more processors to:
obtain one or more satellite images that capture an area containing multiple demarcated fields;
apply the satellite images as input across one or more machine learning models to generate a boundary enhancement channel, wherein each pixel of the boundary enhancement channel is spatially aligned with a corresponding pixel of the one or more satellite images, and wherein each pixel of the boundary enhancement channel is classified with a unit angle to a reference location of the field of the multiple demarcated fields that contains the pixel; and
based on the boundary enhancement channel, determine, and store in the memory, pixel-wise field memberships of pixels of the one or more satellite images in the multiple demarcated fields;
wherein the one or more machine learning models comprise first and second convolutional neural networks in series, wherein output of the first convolutional neural network is applied as input across the second convolutional neural network; and
wherein the output of one of the first and second convolutional neural networks comprises pixel-wise crop type classifications, and output of the other of the first and second convolutional neural networks comprises the pixel-wise field memberships.
12. The system of claim 11 , wherein the one or more machine learning models comprise an atrous spatial pyramid pooling (“ASPP”) layer.
13. A method implemented using one or more processors, comprising:
obtaining one or more satellite images that capture an area containing multiple demarcated fields;
applying the satellite images as input across one or more machine learning models to generate a boundary enhancement channel, wherein each pixel of the boundary enhancement channel is spatially aligned with a corresponding pixel of the one or more satellite images, and wherein each pixel of the boundary enhancement channel is classified with a unit angle to a reference location of the field of the multiple demarcated fields that contains the pixel; and
based on the boundary enhancement channel, determining, and storing in memory, pixel-wise field memberships of pixels of the one or more satellite images in the multiple demarcated fields;
wherein the one or more machine learning models are trained using a weighted combination of two or more of cosine loss on boundary unit angles, cross-entropy loss for boundary pixels, and cross-entropy loss for pixel-wise crop type classifications.
14. A system comprising one or more processors and memory storing instructions that, in response to execution of the instructions by the one or more processors, cause the one or more processors to:
obtain one or more satellite images that capture an area containing multiple demarcated fields;
apply the satellite images as input across one or more machine learning models to generate a boundary enhancement channel, wherein each pixel of the boundary enhancement channel is spatially aligned with a corresponding pixel of the one or more satellite images, and wherein each pixel of the boundary enhancement channel is classified with a unit angle to a reference location of the field of the multiple demarcated fields that contains the pixel; and
based on the boundary enhancement channel, determine, and store in the memory, pixel-wise field memberships of pixels of the one or more satellite images in the multiple demarcated fields;
wherein the one or more machine learning models are trained using a weighted combination of two or more of cosine loss on boundary unit angles, cross-entropy loss for boundary pixels, and cross-entropy loss for pixel-wise crop type classifications.