Realistic plant growth modeling

1 . A method implemented using one or more processors, comprising:

generating, using one or more first machine learning models, an output corresponding to a first two-dimensional (2D) digital image taken at a first time and a second 2D digital image taken at a second time, wherein the first 2D digital image and the second 2D digital image capture one or more of a type of plant, the output to include a first annotation and a second annotation, the first annotation to correspond to one or more structural features of the type of plant at the first time and the second annotation to correspond to the one or more structural features of the type of plant at the second time;

extracting a temporal feature from the first annotation and the second annotation based on a difference between the first annotation and the second annotation;

identifying a correlation between the extracted temporal feature and environmental data corresponding to the first time and the second time, wherein the environmental data is associated with one or more geographic areas in which the first 2D digital image and the second 2D digital image were captured, wherein the correlation is determined based on a mapping between the environmental data and the extracted temporal feature;

generating a first parameter subspace based on the extracted temporal feature and the correlation, wherein a first parameter subspace embedding of the first parameter subspace includes a plant parameter corresponding to one or more extracted temporal features, the correlation, or the environmental data corresponding to the first time and the second time, wherein the first parameter subspace is part of a whole plant parameter space, and wherein the whole plant parameter space corresponds to a species of the type of plant and the first parameter subspace corresponds to the type of plant; and

generating one or more three-dimensional (3D) growth models of the type of plant, wherein the one or more 3D growth models learn the first parameter subspace embedding of the first parameter subspace to constrain a simulation of a growth of the type of plant, and wherein the generation of the one or more 3D growth models is based on selection of at least one plant parameter from the first parameter subspace.

2 . The method of claim 1 , further including:

graphically incorporating the one or more 3D growth models of the type of plant into 2D temporal synthetic training data; and

training one or more second machine learning models based on the 2D temporal synthetic training data.

3 . The method of claim 2 , wherein the 2D temporal synthetic training data includes one or more 2D training images that each captures a ground truth environment, and wherein each of the one or more 2D training images includes one or more superimposed instances of the one or more 3D growth models of the type of plant.

4 . The method of claim 3 , wherein the one or more of the 2D training images includes one or more overlaid 2D instances of a weed interspersed with the one or more superimposed instances of the one or more 3D growth models of the type of plant.

5 . The method of claim 1 , further including:

analyzing the output to detect one or more constituent plant parts in the first 2D digital image and the second 2D digital image;

generating a respective temporal sequence of 3D point clouds for each of the one or more detected constituent plant parts, wherein each of the respective temporal sequences of 3D point clouds captures growth of a corresponding constituent plant part over time; and

based on the respective temporal sequence of 3D point clouds, generate 3D growth models of the constituent plant part over time, wherein the 3D growth models learn a second parameter subspace of constituent plant part parameters to constrain the simulation of the type of plant.

6 . The method of claim 1 , wherein the extracting the temporal feature includes detecting one or more growth time intervals of the one or more structural features of the type of plant.

7 . The method of claim 6 , wherein the one or more growth time intervals include time for a flower to develop into a fruit or time for one of the structural features to grow from a first size to a second size.

8 . The method of claim 1 , wherein the first 2D digital image and the second 2D digital image include one or more videos, wherein each of the one or more videos captures a distinct instance of the type of plant.

9 . The method of claim 1 , wherein the generating the first parameter subspace includes applying a generator and a discriminator of a generative adversarial network (GAN) to search and validate parameters of the first parameter subspace.

10 . 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:

generate, using one or more first machine learning models, an output to correspond to digital imagery that captures at a first time and at a second time one or more of a type of plant, the output to include a first annotation and a second annotation, the first annotation to correspond to one or more structural features of the type of plant at the first time and the second annotation to correspond to the one or more structural features of the type of plant at the second time;

extract a temporal feature from the first annotation and the second annotation based on a difference between the first annotation and the second annotation;

identify a correlation between the extracted temporal feature and environmental data corresponding to the first time and the second time, wherein the environmental data is associated with one or more geographic areas in which the digital imagery was captured, wherein the correlation is determined based on a mapping between the environmental data and the extracted temporal feature;

generate a first parameter subspace based on the extracted temporal feature and the correlation, wherein a first parameter subspace embedding of the first parameter subspace includes a plant parameter corresponding to one or more extracted temporal features, the correlation, or the environmental data corresponding to the first time and the second time, wherein the first parameter subspace is part of a whole plant parameter space, and wherein the whole plant parameter space corresponds to a species of the type of plant and the first parameter subspace corresponds to the type of plant; and

generate one or more three-dimensional (3D) growth models of the type of plant, wherein the one or more 3D growth models learn the first parameter subspace embedding of the first parameter subspace to constrain a simulation of a growth of the type of plant, and wherein the generation of the one or more 3D growth models is based on selection of at least one plant parameter from the first parameter subspace.

11 . The system of claim 10 , further including instructions to:

graphically incorporate the one or more 3D growth models of the type of plant into 2D temporal synthetic training data; and

train one or more second machine learning models based on the 2D temporal synthetic training data.

12 . The system of claim 11 , wherein the 2D temporal synthetic training data includes one or more 2D training images that each captures a ground truth environment, and wherein each of the one or more 2D training images includes one or more superimposed instances of the one or more 3D growth models of the type of plant.

13 . The system of claim 10 , further including instructions to:

analyze the output to detect one or more constituent plant parts in the digital imagery;

generate a respective temporal sequence of 3D point clouds for each of the one or more detected constituent plant parts, wherein each of the respective temporal sequences of 3D point clouds captures growth of a corresponding constituent plant part over time; and

based on the respective temporal sequence of 3D point clouds, generate 3D growth models of the constituent plant part over time, wherein the 3D growth models learn a second parameter subspace of constituent plant part parameters to constrain the simulation of the type of plant.

14 . The system of claim 10 , including instructions to detect one or more growth time intervals of the one or more structural features of the type of plant.

15 . The system of claim 14 , wherein the one or more structural features include one or more of leaves, flowers, branches, stems, fruit, or petioles.

16 . The system of claim 14 , wherein the one or more growth time intervals include time for a flower to develop into a fruit.

17 . The system of claim 14 , wherein the one or more growth time intervals include time for one of the structural features to grow from a first size to a second size.

18 . The system of claim 10 , wherein the digital imagery includes one or more videos, wherein each of the one or more videos captures a distinct instance of the type of plant.

19 . A non-transitory computer-readable medium comprising instructions that, in response to execution of the instructions by a processor, cause the processor to:

generate, using one or more machine learning models, an output to correspond to digital imagery that captures at a first time and a second time one or more of a type of plant, the output to include a first annotation and a second annotation, the first annotation to correspond to one or more structural features of the type of plant at the first time and the second annotation to correspond to the one or more structural features of the type of plant at the second time;

extract a temporal feature from the first annotation and the second annotation based on a difference between the first annotation and the second annotation;

identify a correlation between the extracted temporal feature and environmental data corresponding to the first time and the second time, wherein the environmental data is associated with one or more geographic areas in which the digital imagery was captured, wherein the correlation is determined based on a mapping between the environmental data and the extracted temporal feature;

generate a first parameter subspace based on the extracted temporal feature and the correlation, wherein a first parameter subspace embedding of the first parameter subspace includes a plant parameter corresponding to one or more extracted temporal features, the correlation, or the environmental data corresponding to the first time and the second time, wherein the first parameter subspace is part of a whole plant parameter space, and wherein the whole plant parameter space corresponds to a species of the type of plant and the first parameter subspace corresponds to the type of plant; and

generate one or more three-dimensional (3D) growth models of the type of plant wherein the one or more 3D growth models learn the first parameter subspace embedding of the first parameter subspace to constrain a simulation of a growth of the type of plant, and wherein the generation of the one or more 3D growth models is based on selection of at least one plant parameter from the first parameter subspace.

20 . The method of claim 1 , wherein the generation of the one or more 3D growth models based on the selection of the at least one plant parameter from the first parameter subspace further includes processing the at least one plant parameter from the first parameter subspace using a time series machine learning model, wherein the time series machine learning model outputs a predicted dimension of a plant part of the type of plant.