Deep learning system and method for predicting crop characteristics

Patent metadata

Field	Value
Publication number	US-2023389460-A1
Application number	US-202218056677-A
Country	US
Kind code	A1
Filing date	Nov 17, 2022
Priority date	Jun 3, 2022
Publication date	Dec 7, 2023
Grant date	—

Abstract

A deep learning system is used to predict crop characteristics from inputs that include crop variety features, environmental features, and field management features. The deep learning system includes domain-specific modules for each category of features. Some of the domain-specific modules are implemented as convolutional neural networks (CNN) while others are implemented as fully-connected neural networks. Interactions between different domains are captured with cross attention between respective embeddings. Embeddings from the multiple domain-specific modules are concatenated to create a deep neural network (DNN). The prediction generated by the DNN is a characteristic of the crop such as yield, height, or disease resistance. The DNN can be used to select a crop variety for planting in a field. For a crop that is planted, the DNN may be used to select a field management technique.

First claim

1 . A system comprising:
a processing unit; and a computer-readable medium having encoded thereon instructions, that when executed by the processing unit, cause the system to:
generate a soil embedding from soil features processed through a soil module that comprises a first neural network;
generate a weather embedding from weather features processed through a weather module that comprises a second neural network;
generate a field management embedding from field management features processed through a field management module that comprises a third neural network;
generate a crop variety embedding from crop variety features processed through a crop variety module that comprises a fourth neural network;
concatenate by a fusion module the soil embedding, the field management embedding, the weather embedding, and the crop variety embedding and provide to a deep neural network (DNN); and
receive from the DNN a predicted value for a crop characteristic of the crop variety.

2 . The system of claim 1 , wherein the instructions further cause the system to generate a variety-weather embedding by combining the variety embedding and the weather embedding with cross attention in a cross attention module and wherein the fusion module concatenates the variety-weather embedding with the soil embedding, the field management embedding, and the weather embedding.

3 . The system of claim 1 , wherein at least one of the first neural network, second neural network, third neural network, and fourth neural network is a convolutional neural network (CNN) and at least one is a fully-connected neural network.

4 . The system of claim 1 , wherein the soil features include a percentage of clay, percentage of sand, percentage of silt, percentage of organic matter, calcium content, magnesium content, phosphorus content, nitrate content, potassium content, sodium content, sulfate content, pH, soil conductivity, percentage of calcium saturation, percentage of hydrogen saturation, percentage of potassium saturation, percentage of magnesium saturation, and percentage of sodium saturation, and the first neural network is a fully connected neural network with at least two hidden layers.

5 . The system of claim 1 , wherein the field management features use of irrigation, irrigation amount, planting density, total amount of nitrogen (N) fertilizer, total amount of phosphorus (P) fertilizer, and total amount of potassium (K fertilizer and the second neural network is a fully connected neural network with at least two hidden layers.

6 . The system of claim 1 , wherein the weather features are a timeseries including solar radiation, vapor pressure, dewpoint, precipitation, maximum temperature, minimum temperature, wind speed, relative humidity, dewpoint, day length, and growing degree days (GDD) and the third neural network is a convolutional neural network (CNN) with at least two convolutional layers.

7 . The system of claim 1 , wherein the crop variety features comprise the genome of the crop variety and the fourth neural network is a CNN.

8 . The system of claim 1 , wherein the crop characteristic of the crop variety is one of grain yield, protein content, moisture content, fiber content, height, drought resistance, molecular or metabolic characteristic, and disease resistance.

9 . A method comprising:
a) obtaining soil features and weather features for a field; b) identifying a field management technique; c) receiving an indication of a first crop variety; d) generating, using neural networks, embeddings from the soil features, the weather features, field management features of the field management technique, and features of the first crop variety; e) providing a concatenation of the embeddings to a deep neural network (DNN) trained to predict a value for a crop characteristic; f) receiving a predicted value for the crop characteristic of the crop variety from the DNN; g) repeating operations c-f with a second crop variety; h) selecting either the first crop variety or the second crop variety as a selected crop variety based on the respective predicted values for the crop characteristic; and i) planting the selected crop variety in the field.

10 . The method of claim 9 , wherein a soil embedding is created from the soil features by a first neural network, a weather embedding is created from the weather features by a second neural network, a field management embedding is created from the field management features by a third neural network, and a crop variety embedding is created from the features of the crop variety by a fourth neural network.

11 . The method of claim 10 , wherein the first neural network is a fully-connected neural network, the second neural network is a CNN, the third neural network is a fully-connected neural network, and the fourth neural network is a CNN.

12 . The method of claim 10 , further comprising creating a variety-weather embedding by combining the variety embedding and the weather embedding with cross attention and wherein concatenating the embeddings comprises concatenating the soil embedding, the field management embedding, the weather embedding, and the variety-weather embedding.

13 . The method of claim 10 , wherein during repetition of operations c-f with the second crop variety the soil embedding, the weather embedding, and the field management embedding are reused.

14 . The method of claim 9 , wherein a plurality of field management techniques are identified and further comprising:
after choosing the selected crop variety, repeating operations b-f with each of the plurality of field management techniques; selecting a one of the plurality of field management techniques based on the predicted values for the crop characteristic; and managing the selected crop variety after planting according to the selected one of the field management techniques.

15 . A method comprising:
a) obtaining soil features and weather features for a field; b) receiving an indication of a crop variety planted in the field; c) identifying a first field management technique; d) generating embeddings using neural networks from the soil features, the weather features, features of the crop variety, and field management features of the field management technique; e) providing a concatenation of the embeddings to a deep neural network (DNN) trained to predict a value for a characteristic; f) receiving a predicted value for the characteristic of the crop variety from the DNN; g) repeating operations c-f with a second field management technique; h) selecting either the first field management technique or the second field management technique as a selected field management technique based on the respective predicted values for the characteristic; and i) managing the crop variety according to the selected field management technique.

16 . The method of claim 15 , wherein the field management features include at least one of use of irrigation, irrigation amount, planting density, total amount of nitrogen (N) fertilizer, total amount of phosphorus (P) fertilizer, and total amount of potassium (K fertilizer.

17 . The method of claim 15 , wherein a soil embedding is created from the soil features by a first neural network, a weather embedding is created from the weather features by a second neural network, a field management embedding is created from the field management features by a third neural network, and a variety embedding is created from features of the crop variety by a fourth neural network.

18 . The method of claim 17 , wherein the first ne

Assignees

• Microsoft Technology Licensing Llc

Inventors

• De Freitas Cunha Renato Luiz
• Badam Anirudh
• Buehler Patrick Bernd
• Chandra Ranveer
• Dan Debasis
• De Luis Blaguer Maria Angels
• Sharma Swati
• Aditi Fnu
• Maua Sara Malvar

Classifications

• A01B79/005 Primary

  Precision agriculture · CPC title

• A01B79/02 Primary

  combined with other agricultural processing, e.g. fertilising, planting · CPC title

• G06N3/0454

  Physics · mapped topic

• G06N3/08

  Learning methods · CPC title

• G01W1/02

  Instruments for indicating weather conditions by measuring two or more variables, e.g. humidity, pressure, temperature, cloud cover or wind speed (G01W1/10 takes precedence) · CPC title