Irrigation system control with predictive water balance capabilities

1. A system for reinforcement learning-based irrigation control to maintain or increase a crop yield or reduce water use, comprising:

at least one computing device; and

program instructions stored in memory and executable by the at least one computing device that, when executed, direct the at least one computing device to:

determine an optimal irrigation schedule for at least one crop in at least one region of a field by executing a reinforcement learning (RL) routine, where, for a given state of a total soil moisture, the reinforcement learning (RL) routine is configured to:

perform an action, the action comprising waiting or irrigating the at least one crop;

execute a cascading neural network comprising a first neural network and a second neural network and generate the crop yield for the at least one crop based on the performed action by:

executing a first neural network that receives, as an input, at least one of irrigation data or weather data and determines a total soil water (TSW) value determined based at least in part on the input, wherein the first neural network receives training data from a crop growth computer model and is trained using the training data prior to generating the total soil water (TSW) value; and

executing a second neural network that receives the total soil water as an input and generates the crop yield based at least in part on the total soil water, wherein the second neural network receives training data from the crop growth computer model and is trained using the training data before generating the crop yield;

assign a reward to a state-action pair, the state-action pair comprising the given state of the total soil moisture and the action performed, wherein the reward is assigned based on a predicted crop yield generated by the cascading neural network; and

instruct an irrigation system to apply irrigation to at least one crop in accordance with the optimal irrigation schedule determined, wherein the optimal irrigation schedule comprises an amount of water and a determined time at which the amount of water should be applied.

2. The system of claim 1 , wherein the optimal irrigation schedule is determined using at least one of: a real-time soil moisture value, a near real-time soil moisture value, a predictive evapotranspiration (ET) metric, or a weather forecast metric.

3. The system of claim 1 , wherein the system further comprises a decision support system for agrotechnology transfer (DSSAT) computing device.

4. The system of claim 1 , wherein:

the at least one computing device is a microcontroller implemented in a mobile irrigation machine; and

the irrigation system is instructed to apply irrigation to at least one crop in accordance with the optimal irrigation schedule by converting the optimal irrigation schedule to a suitable signal for interpretation by the irrigation system.

5. The system of claim 1 , further comprising at least one soil moisture sensor positioned in the at least one region of the field; and

wherein the optimal irrigation schedule is determined based at least in part on a soil moisture measurement obtained by the at least one soil moisture sensor.

6. A computer-implemented method for reinforcement learning-based irrigation control to maintain or increase a crop yield or reduce water use, comprising:

determining, by at least one computing device, an optimal irrigation schedule for at least one crop in at least one region of a field by executing a reinforcement learning (RL) routine, where, for a given state of a total soil moisture, the reinforcement learning (RL) routine comprises:

simulating, by the at least one computing device, an action, the action comprising waiting or irrigating the at least one crop;

generating, by a cascading neural network comprising a first neural network and a second neural network, a crop yield for the at least one crop based on the simulated action by:

receiving, by the first neural network, as an input, at least one of irrigation data or weather data;

providing, by the first neural network, a total soil water (TSW) value determined based at least in part on the input, wherein the first neural network receives training data from a crop growth computer model and is trained using the training data prior to generating the total soil water (TSW) value;

receiving, by the second neural network, the total soil water as an input; and

generating, by the second neural network, the crop yield based at least in part on the total soil water, wherein the second neural network receives training data from the crop growth computer model and is trained using the training data before generating the crop yield; and

assigning, by the at least one computing device, a reward to a state-action pair, the state-action pair comprising the given state of the total soil moisture and the simulated action, wherein the reward is assigned based on a predicted crop yield generated by the cascading neural network; and

instructing an irrigation system to apply irrigation to at least one crop in accordance with the optimal irrigation schedule determined, wherein the optimal irrigation schedule comprises an amount of water and a determined time at which the amount of water should be applied.

7. The computer-implemented method of claim 6 , wherein the optimal irrigation schedule is determined using at least one of: a real-time soil moisture value, a near real-time soil moisture value, a predictive evapotranspiration (ET) metric, or a weather forecast metric.

8. The computer-implemented method of claim 6 , wherein the at least one computing device is a microcontroller implemented in a mobile irrigation machine; and

further comprising applying, by the microcontroller, irrigation to at least one crop in accordance with the optimal irrigation schedule by converting the optimal irrigation schedule to a suitable signal for interpretation by the irrigation system.

9. The computer-implemented method of claim 6 , wherein the optimal irrigation schedule is determined based at least in part on a soil moisture measurement obtained by at least one soil moisture sensor positioned in the at least one region of the field.

10. The system of claim 1 , wherein the reward assigned to the state-action pair is one of an immediate reward or a delayed reward to the state-action pair.

11. The system of claim 10 , wherein:

the reward assigned to the state-action pair comprises the delayed reward; and

the at least one computing device is further directed to execute a temporal difference learning algorithm to account for the delayed reward.

12. The computer-implemented method of claim 6 , wherein assigning, by the at least one computing device, the reward to the state-action pair further comprises assigning one of an immediate reward or a delayed reward to the state-action pair.

13. The computer-implemented method of claim 12 , wherein:

the reward assigned to the state-action pair comprises the delayed reward; and

the computer-implemented method further comprises executing a temporal difference learning algorithm to account for the delayed reward.

14. The computer-implemented method of claim 6 , wherein the at least one crop is at least one of wheat or maize.

15. A non-transitory computer readable storage medium storing software thereon, the software comprising instructions configured to cause at least one processor of at least one computer to perform steps comprising:

determining an optimal irrigation schedule for at least one crop in at least one region of a field by executing a reinforcement learning (RL) routine, where, for a given state of a total soil moisture, the reinforcement learning (RL) routine comprises:

simulating an action, the action comprising waiting or irrigating the at least one crop;

generating, by a cascading neural network comprising a first neural network and a second neural network, a crop yield for the at least one crop based on the simulated action by:

receiving, by the first neural network, as an input, at least one of irrigation data or weather data;

providing, by the first neural network, a total soil water (TSW) value determined based at least in part on the input, wherein the first neural network receives training data from a crop growth computer model and is trained using the training data prior to generating the total soil water (TSW) value;

receiving, by the second neural network, the total soil water as an input; and

generating, by the second neural network, the crop yield based at least in part on the total soil water, wherein the second neural network receives training data from the crop growth computer model and is trained using the training data before generating the crop yield; and

assigning a reward to a state-action pair, the state-action pair comprising the given state of the total soil moisture and the simulated action, wherein the reward is assigned based on a predicted crop yield generated by the cascading neural network; and

instructing an irrigation system to apply irrigation to at least one crop in accordance with the optimal irrigation schedule determined, wherein the optimal irrigation schedule comprises an amount of water and a determined time at which the amount of water should be applied.

16. The non-transitory computer readable storage medium of claim 15 , wherein the optimal irrigation schedule is determined using at least one of: a real-time soil moisture value, a near real-time soil moisture value, a predictive evapotranspiration (ET) metric, or a weather forecast metric.

17. The non-transitory computer readable storage medium of claim 15 , wherein the optimal irrigation schedule is determined based at least in part on a soil moisture measurement obtained by at least one soil moisture sensor positioned in the at least one region of the field.

18. The non-transitory computer readable storage medium of claim 15 , wherein assigning the reward to the state-action pair further comprises assigning one of an immediate reward or a delayed reward to the state-action pair.

19. The non-transitory computer readable storage medium of claim 18 , wherein:

the reward assigned to the state-action pair comprises the delayed reward; and

the steps further comprise executing a temporal difference learning algorithm to account for the delayed reward.

20. The non-transitory computer readable storage medium of claim 18 , wherein the steps further comprise applying irrigation to at least one crop in accordance with the optimal irrigation schedule by converting the optimal irrigation schedule to a suitable signal for interpretation by an irrigation system.