GR1010281B - Soil moisture collection and processing arrangment and method aiming at the optimal automatically-commanded irrigation in cultivations - Google Patents

Abstract

The present invention relates to a combined method consisting in the calculation of soil moisture at all times and various depths, using machine learning algorithms in the various phases and steps of the method’s organization. This method aims at eliminating existing problems that make unproductive and economically unviable the exclusive in-situ detailed recording of soil moisture in many places, in a cultivated area and at different depths, with use of sensors. This goal is achieved, in the first phase, by the combined use of: a) data derived from in-situ measurements obtained by a specific method; b) data extracted from special satellite receptions; c) recording and forecasting data of climatic conditions, used for the training of neural networks. In the second phase, there is achieved the gradual abolition of the in-situ measurements as well as the abolition of the satellite reception and recorded climatic conditions data. With the use of this method, the conditions for the abolition of on-situ measurements with the use of sensors are gradually shaped, drastically reducing the operating and maintenance costs of the process of estimating, at every moment, the soil moisture in a cultivated area.

Description

Method and device for optimizing processes for collecting & processing soil moisture data for use in automatically determining commands for optimal crop irrigation

Description

1 Technical field to which the invention refers

The use of special sensors in fields used for the on-site measurement of climatological data and in particular data related to the formation of soil moisture at various depths which are interconnected through networks (e.g. 4G, LoRa, Wi-Fi, 5G NR, C - band, 60 GHz, CBRS, etc.) to transfer the relevant data for processing to a central computing infrastructure, and in combination with other data related to the prediction of moisture variation (precipitation height, humidity and wind intensity , etc), in order to formulate the proposed optimal irrigation program for each crop, while it can theoretically constitute a practice that can lead with sufficient accuracy to the diagnosis of irrigation needs, however, in practice, it requires a large number of sensors, in order to be possible the recording of moisture conditions in several places within a field and at different depths (from 0 - 1 m), to the extent that the specific method becomes, m the financially viable.

As far as measurement sensors are concerned, their practical use has shown that special shielding against moisture and oxidative damage is required in order to be able to function uninterruptedly for a long time and to be effectively maintained, while the power and network connection supplies for sending and receiving data, create additional needs and costs for long-term maintenance. On the other hand, especially during the last 3-4 years, due to the great development of the relevant satellite technology, the quite detailed recording capability with footprint analysis that provides a moisture value per 100 square meters (or even for a smaller area) and with scanning frequency of each region for the recording of values every 3 days, gives new important possibilities to capture and utilize atmospheric and ground data. The aim of all these methods is to record the soil moisture as accurately as possible at various depths from the surface and up to 1-2 meters at any time, in order to estimate the need to irrigate each crop in order to accordingly plan the moment when the irrigation will take place , as well as the appropriate amount at each point of the irrigated area.

The present invention refers to a combined method of calculating soil moisture at any moment, at various depths, using machine learning algorithms in the various phases and steps of its organization, which aims to eliminate the aforementioned problems, which make it unproductive and economically exploitable the exclusively on-site analytical recording of soil moisture at several points in a cultivated area and at different depths using sensors. This purpose is achieved in the first phase by the combined use of - a) data from on-site measurements obtained by a specific method based on the use of sensors,

- b) data extracted from special satellite images and

- c) recording and forecasting data of climatic conditions, which are used for the training of neural networks and in a subsequent phase with the gradual abolition of on-site measurements and the use exclusively of data extracted from satellite images and data recording and forecasting of climatic conditions in an area.

With the use of the method, conditions are gradually formed for the abolition of on-site measurements using sensors, drastically reducing the cost of operation and maintenance of the process of estimating soil moisture at each moment in a cultivated area.

The method involves performing specific actions involving two types of neural networks. The first network is a cycleGAN network, which when trained creates enhanced humidity indicators. In its training, in situ measurements of soil moisture are used and the network is trained to improve the indices of the observation set based on these measurements.

CycleGANs can capture the features of one dataset and translate them into the features of another dataset, without the need for paired examples between two datasets. This is very useful when we want to capture the information contained in the sensor measurement dataset and store it in the network parameters. This network can then take only remote sensing data and generate new observations that are enhanced by the benefit of the information embedded in the sensor data.

The second network is a recurrent neural network (RNN - ), which was trained using reinforcement learning methods. The network receives as input an observation and generates an energy command, in this case a command to irrigate the cultivated area per cultivated tree/plant.

2 State of the art and their evaluation Information on the spatiotemporal variability of soil moisture, i.e. the timely availability of accurate soil moisture information at regular intervals, is valuable for many purposes, including: monitoring drought management of water resources, irrigation planning, prediction of vegetation dynamics and agricultural yields, prediction of floods and heat waves, and understanding the impacts of climate change (Chawla, I., Karthikeyan, L, and Mishra, A K.: A review of remote sensing applications for water security: quantity, quality, and extremes,30 Journal of Hydrology, p. 124826, https://doi.org/ 10.1016/j .j hydrol.2020.124826, 2020.) .

The complete overview of the products and methods available for the generalized measurement of moisture in the upper soil layers is provided in the recent scientific publication: “Evaluation of 18 satellite- and model-based soil moisture products using in situ measurements from 826 sensors evaluated the temporal dynamics of 18 state-of-the-art (quasi)global near-surface soil moisture products, including six based on satellite retrievals, six based on models without satellite data assimilation (referred to hereafter as “open-loop” models), and six based on models that assimilate 5 satellite soil moisture or brightness temperature data (H.EBeck, M.Ran, D.IVSralles et al), Hydrology and Earth System Sciences, May 2020. This very recent publication highlights the significant potential of remote sensing measurements through satellites and the small degree of deviation they show in relation to on-site measurements

In recent decades, numerous products and methods have been developed to accurately measure and calculate soil moisture, each with advantages and disadvantages. The products differ in their design objective (general or local use), spatial resolution and coverage, data sources, calculation algorithm and delay in their availability (especially when derived from remote sensing). They can be broadly classified into three broad categories

- (i) Products derived from direct active or passive microwave satellite remote sensing

- (D) Hydrological models or land surface models without satellite data assimilation (open loop models)

- (iii) Hydrological or land surface models assimilating soil moisture values derived from satellite remote sensing or satellite microwave remote sensing and brightness temperature observation

Numerous recent studies have evaluated these soil moisture measurement products and methods using in situ measurements other independent soil moisture products remotely sensed green vegetation data or precipitation data. Strong differences in spatiotemporal dynamics and accuracy were found among products and methods, even among those from the same data source. In addition, several new or recently reprocessed products have not yet been thoroughly evaluated, while there is also uncertainty e.g. on the ineffectiveness of multi-sensor fusion techniques, the effect of model complexity on the accuracy of soil moisture simulations, and the extent to which model deficiencies and precipitation data quality affect the added value of data assimilation.

However, to date the use of methods to create soil moisture calculation products that utilize machine learning techniques and algorithms is limited.

The possibilities of productive integration of data from two sources that train a neural network to learn to create a new improved data set but also the possibilities provided by stochastic variational inference algorithms (CycleGAN algorithms as well as Variatonal AutoEncoder) are discoveries of recent years in the field of artificial intelligence. intelligence and machine learning made possible due to the significant progress of science in the field especially in the last 5-7 years, which is due to the possibility of collecting, processing and managing a large amount of data.

cycleGAN is a class of GAN (generative adversarial networks) algorithms that involves automatically training image-to-image translation/transformation models without paired examples. Network models are trained unsupervised using a collection of images from the source domain and the target domain, which are not necessarily related in any way.

The original proposal of CydeGAN is made in “Unpaired lmage-to-Image Translation using Cyde-Cbnsistent Adversarial Networks, Zhu et al, Berkeley Al Research Laboratory, 2017 (https://arxiv.org/pdf/1703.10593.pdf , lEEE Xplore, International Conference on Computer Vision) and in 'Bros, J.-Y. Z (2020). Unpaired Image-to-Image Translation using Cyde-COnsistent Adversarial Networks Retrieved from https://arxiv.Org/abs/1 703.1 0593# »

Accordingly, Auto-Encoding Variational Bayes (https://arxiv.org/pdf/1312.6114.pdf, Kingma & Welling 2014, University of Amsterdam) and the Sochastic Backpropagation and approximate inference in deep generative models method inference in deep regenerative learning models) provide combinatorially meaningful capabilities and introduce new recognition models to approximate posterior distributions by acting as a stochastic encoder of the data.

With the significant evolution of the possibilities of precise recording through remote sensing provided by satellite infrastructures such as that of Copernicus (https://www.copemicus.eu/el/ypiresies/xira), combined with the capabilities provided by computers today to perform of Artificial Intelligence and Machine Learning algorithms by oxidizing large data sets, new fields of research and innovation are opening up in relation to:

- the observation of natural phenomena,

- the creation of models for predicting the evolution of various indicators and parameters that determine the growth and performance of a crop and - taking preventive measures to improve the result.

Under these technological conditions with the combined use of the most appropriate machine learning algorithms, the disclosed method was formulated, which is innovative and directly applicable in the field, and has already been developed experimentally and works in real conditions, using electronic and computational devices to solve the problem of irrigation optimization in open field tree plantations.

1 Detailed Description - disclosure of the invention

The presented method consists of a series of actions organized in steps and phases, performed in a technical infrastructure of computing and other electronic devices. In the composition and configuration of the method, computer/network devices in which developed software works.

Each arrangement constitutes one of the subsystems that make up the technical infrastructure (the overall system) necessary to implement the method in practice (Figure 1).

These provisions are organized into the following subsystems:

- Y1. Electronic stage of initial management and measurement routing hardware, with integrated electronic processing devices that scale according to the cultivated area placed in a protection box (Measurement Unit), which is connected to an array of analog sensors. The data from the various measurement units end up at the first level in a routing unit, through which they are sent to the Internet and then to Y4. - Y2. Arrays of analog sensors that are placed inside the ground and receive data

- Y3. A wide area network consisting of a wireless local area network and an Internet access network or cloud computing infrastructure

- Y4. Computational Processing Unit, which is installed in the form of a virtual machine (the array of them) in a Cloud Computing Infrastructure (cloud - YYN) and in which the Database operates in which the data is stored and the various algorithms are executed in the context of the application of the method.

- Y5. Irrigation management unit. Interfaces via internet or via Y3 to one or more standard electric flow controlled irrigation valves

In the context of the method, the data used and processed by the above subsystems fall into the following categories:

- D1. Satellite and remote sensing data. These include Raw satellite data (data from bands B04, B08 and B10)

Exported indices such as NDVI (Normalized Difference Vegetation Index) and NDMI (Normalized Difference Moisture Index)

- D2. Meteorological data:

o Historical: average, high and low temperature, humidity percentage, daily precipitation (in millimeters), atmospheric pressure, wind speed and direction, dew point temperature.

o Forecast data: day and next days forecast including temperature (maximum - minimum), precipitation height, atmospheric pressure, wind speed and direction, dew point temperature

- D3. Cultivated area data: Geographical coordinates of polygonal area, altitude, slopes and identification of critical points - D4. Tree-crop data: Crop identity, soil moisture needs in the various phases of the life cycle.

The system takes as input a set of observational data that includes

- (a) the satellite land remote sensing dataset (from the Sentinel-2 satellite the data from bands B04, B08 and B10, NDVI and NDMI indices) and

- (b) the set of meteorological data (temperature, humidity, precipitation, wind speed and direction, dew point).

When the system is queried to produce a result, the last 30 days of satellite data, indices and meteorological data are collected and integrated into a batch. In this batch there should be 6 distinct groups of raw satellite data (1 every 5 days), indicators and at least 30 days of meteorological data.

The method that is the object of the present invention consists of a series of phases that include individual steps and actions that utilize the data and execute the algorithms. They finally arrive at the calculation of the need to irrigate the cultivated area with the specific type of crop, which in practice takes the form of an order given to controlled irrigation systems to supply a specific amount of water, finally measured in the amount of water per tree/cultivated plant.

The following is a more specific description of the layout and organization of the subsystems involved in the development and application of the method and then the actions that make up the method, organized into steps per phase.

1 Description of Subsystems

Y1. Electronic stage of initial management and measurement routing hardware with integrated processing electronics housed in a protection box

The self-contained electronic hardware stage Y1 integrates processing and communication electronics, and is housed in an IP65 weatherproof enclosure. The unit is based on a microcontroller that has a dual-core 32-bit processor, speed up to 600 DIMPs, built-in RAM memory 520 KB, RAM expansion 4 MB, and external flash memory 8 MB. It is powered by a constant voltage of 3.5-5.5VDC. It has a multitude of inputs/outputs and peripherals for interfacing with external analog (ADC) and digital sensors, supporting the UART, I2C, SPI protocols, while incorporating communication layers for the telecommunication protocols WiFi 802.11b/g/n 16Mbps, Bluetooth (Classic & LE ), LoRa and Sigfox.

The stage is powered by a rechargeable battery with a nominal voltage of 3.7-4.2V which ensures its uninterrupted operation for at least one (1) year,

The stage is wired to an array of 3-4 analog ground sensors, which are placed at depths of 5, 30, 60 and 90 cm respectively, and performs periodic measurements, with a frequency of 1 recording per hour. Intermittently it enters a deep hypnosis state for maximum energy saving and automatically switches to the active state using an internal clock.

In the active state, the following processes take place:

1) The voltage/current supply to the ground sensors is activated for a duration of 250 msec, using a suitable electronic device controlled by a digital micro-controller.

2) Through another analog type electronic device, the electrical conductivity level at the different depths is measured and retrieved, and then the power supply to the sensors is cut off.

3) Measurements are then taken from an external temperature sensor placed in the ground at a depth of 30 cm, an internal temperature/humidity sensor placed on the electronic board, as well as metadata such as battery level.

4) The set of data/metadata is formatted using an internally defined protocol into a data packet (bytes), in order to minimize the transmission time based on a certain desired accuracy.

5) An end-to-end interface with a remote private LoRaWan network infrastructure installed in a public Cloud is realized through an integrated telecommunication layer compatible with the LoRaWAN protocol. The connection is made using cryptographic keys that ensure the integrity and security of the communication.

6) The data packet is sent from the remote node through the LoRaWAN wireless interface, it is first received by a Data Routing Unit installed in the wider area (a commercial such unit is that of the RAK company, WisGate Edge Lite RAK7258 https://store. rakwireless.com/products/rak7258-microgateway?variant=39942876430534), and is forwarded through a Public Data Network of 4G technology and the use of the UDP/IP protocol in a Cloud infrastructure, in which Y2 operates. After confirming a successful transmission the remote node enters a deep sleep state until it restarts for a repeated recording cycle.

7) The data packet is received by a LoRaWAN network server application, decoded to retrieve the measurement data and accompanying metadata, and then forwarded using the open MQTT protocol to the routing unit where the data is stored.

8) A network application undertakes to present to the interested user (farmer, agronomist) all the history of the measurements taken, in the form of Percent Volumetric Quantity of Height. This metric is derived from electrical conductivity measurements using mathematical equations provided by the sensor manufacturer. The results are accompanied by environmental temperature/relative humidity readings taken from respective sensors.

Y2. Sensor arrays

Commercially available analog temperature and soil moisture measurement sensors are used. In the context of the experimental set-up developed, relevant sensors from the companies metergroup (ECH20 EC-5 https://www.metergroup.com/ environment/products/ec-5-soil-moisturesensor/) and maxim integrated (DS18B20 https:// gr.mouser.com/datasheet/2/256/maxim%20integrated%20products_ds18b20-1178755.pdf)

Y3. Wide area network

The wide area network includes two sub-infrastructures:

- LoRaWan network infrastructure, which covers data traffic between the units that make up Y1

- 4G network infrastructure that can be provided by any telecommunications provider or other infrastructure (eg WiFi or wired Ethernet) through which data is routed to the Internet

Y4. Subsystem - Computational Processing Unit (or array thereof) This subsystem consists of four databases and two computing elements (Figure 2). The bases are the following:

• Weather Data Base

• Satellite Data Base and file system

· Sensor Data Base

• Model Database

The aforementioned bases accept data from external online meteorological and satellite data services as well as sensor data (via Y1 and Y2). While the model base records the trained models.

The training computing component (Trainer) accepts as input data a series of primary satellite data (B04, B08, BIO), secondary data NDVI (Normalize Difference Vegetation Index), NDMI (Normalize Difference Moisture Index), weather data and sensor data . The training computer is responsible for training the guidance computer (Agent). When the subsystem is called to issue a command, then the Agent requests the last month's satellite data and the last month's weather data. The result of the Agent is a watering command, which includes the amount of water and the time and duration for which the command must be executed.

The computer guidance element (Agent) consists of two neural networks. The first network is a cycleGAN (Efros, 2020). This network is trained with the data it receives (a- from the on-site measuring stations placed in the field for one year and b - from the Sentinel satellites) in order to give modified secondary satellite data, detailed and adjusted based on the data from the field that have been used during his training.

During its training phase, measurements from sensors in the field are used along with other data. The cycleGAN neural network is a class of networks that are capable of capturing features of one data set, and translating them into features of another data set, without the need for examples of pairs between the two sets.

A version of the cycleGAN neural network and the algorithm used to extract the features of the data from the sensors and transfer them to the secondary data of the satellites are integrated into the subsystem, adapted accordingly according to the format of the data. By using cycleGan, having only the satellite data, specialized data can be produced as if the sensor data were also present.

The second network implemented in the subsystem is an RNN (recurrent neural network), which is trained with reinforcement learning methods. It receives as input the results of cycleGan and outputs the irrigation time and amount commands to Y5.

Y5. Subsystem - Irrigation Management & Control Unit.

Various commercially available products that can be connected to LoRaWan networks can be used for the implementation of Y5, such as e.g. STREGA LoRa wireless smart valve fhttp://www.streaat echnoloaies.com /products/ wireless- smart-valve/) or the Milesight LoRaWAN® Solenoid Valve Controller (https://www.milesiaht-iot.com/lorawan/ controller/ uc51x/) in which appropriate commands are provided through the system to 'open' and 'close' the automatic electric valve for specific intervals in order to release the required volume of water per tree at a time.

1 Method Phases

The method is developed in phases and in each phase there are steps in which the various actions are included, performed in the subsystems presented in the previous section.

Phase A: Data Collection and Preparation

During Phase A, data is collected and initial processing and preparation is done. They are divided into the following individual subgroups:

Remote sensing and meteorology data sets (observation data set)

To create this ensemble, satellite and meteorological data provided by satellites (specifically Sentinel 2) are combined. Sentinel 2 passes and captures the same geographic point every 3-5 days (over the course of a year it passes about 80-100 times over the same point).

The steps followed in this method are as follows:

- Step 1.1: The boundaries and vertices of the polygon of the cultivated area are captured in the form of geographic coordinates and the elevation curves within the area and the points with local minimum and maximum heights, as well as the points

- Bia 1.2:. Through the appropriate web interface, a query is sent to the web service of the European satellite data provision system 'Copernicus' (Copernicus Open Access Hub). The question asks for the geographic coordinates of the arable area polygon and a five-year period, in order to retrieve the raw data from the Sentinel 2 satellite and download and store it in the Y4 Database.

- Example 1.3: From the downloaded data, bands B04, B08 and B10 are selected which are the first three channels from the remote sensing data set.

- Step 1.4: NDVI (normalized difference vegetation index) and NDMI (normalized difference moisture index) are calculated for the dates for which data are available for the 3 bands,

B08 - B04

NDVI =

B08 B04

B10 - B04

NDMI =

B10 B04

These are the other two channels that complete the remote sensing data set

Step 1.5: The collection is carried out by downloading from an external meteorological data distribution system all the data concerning the arable area to form the last channel participating in the remote sensing data set (indicatively the meteorological data source https://openweathermap.org is mentioned /). The data that is inherited concerns the variables:

<■>Mean, highest, lowest temperature - mean, highest, lowest temperature (in Celsius)

<■>Humidity percentage - humidity percentage

<■>Rain volume (in mm)

<■>Atmospheric pressure - pressure (in hPa)

<■>Wind speed - wind speed (in Km/h)

<■>Wind direction - wind direction (in degrees)

<■>Dew point temperature (in Celsius)

Step 1.6: For each available date (every 3-5 days), historically for 5 years, the raw satellite data and the calculated indices are stacked, creating a 3D tensor, in a similar way that a color (three-channel) is coded - RGB) image. For the same intervals and for the specific days that the satellite data are available (every 3 - 5 days) and for the same interval of the previous 5 years, the meteorological data is stacked in a vector of length 5. At the end, the pairing is done ) of each layer of satellite data with the corresponding vector of meteorological data. The above steps are performed and coordinated by software running on Y4, specifically in Data Collection.

Data sets of field measurements using sensors (sensor Data Set)

To create the dataset of field measurements, soil moisture and temperature data are collected over a period of one year.

- Step 1.7: Soil moisture measurement sensors are placed at a depth of 30, 60, 90 cm and soil temperature measurement sensors at a depth of 30 cm.

- Measurements are taken hourly. The placement of the sensors in the arable area is done with the following method: a) the points in the arable area are selected that are located at the deepest and highest points in altitude, as well as 1-3 intermediate points b) at these points the nearest trunk is located and placed one group of sensors at a distance of 30 cm from the trunk and one at the border of the crown (about 3 meters from the trunk) and at a point not shaded by the crown.

- Step 1.8. For each layer at each date, the data from each sensor are modeled as vectors of length 24.

Phase B: Initial System Training

During this phase, the neural network configuration algorithm (cycleGAN) is trained to learn to generate upgraded and improved remote sensing data, from the original remote sensing data and the data derived from the field measurements with the sensors.

The method adapts the cycleGAN algorithm that leverages remote sensing data and sensor measurement data by following these steps:

- Step 2.1 The two sets of data (remote sensing and measurements) are aligned in time to refer to the same dates and times. For each vector of satellite data, there is a vector of length 5 with meteorological data, and a vector of length 120 with measurement data for the period in which measurements are taken.

- Step 2.2 ̈ Batches of monthly sequences are created. In a batch there are 6 satellite data tensors with a total of five distinct dates, 6 weather data vectors with a total of 30 distinct dates, and 6 sensor data vectors with a total of 720 distinct timestamps. - Step 2.3 ̈. Batches are used to train the CycleGAN network. The result of the network each time has the same structure as the remote sensing data, but the values of the NDVI and NDMI index segments are augmented with the data of the measurements from the sensors, that is, the sensor data is merged into the remote sensing data.

- The raw satellite data and weather data within the observational data should be identical between the input and output of the network.

Finally when the network is fully trained, the information collected from the measurement data with the sensors is stored in the network parameters. When the network is only fed with remote sensing data, it can now generate improved observations and therefore there is no need for further data collection with sensor measurements.

Figure 3 shows the steps included in the method in Phase B.

Phase C: Continuous operation and training - identification of irrigation actions

In this Phase a Recurrent Neural Network (RNN) is a class of neural networks where the connections between nodes form a directed graph along a time sequence. This allows it to exhibit temporally dynamic behavior. Derived from feedforward neural networks, RNNs can use their internal state (memory) to process input sequences of variable length. The method in Phase C includes the following steps:

- Step 3.1 The initial remote sensing and a data set of meteorological data vector sequences are created. - Example 3.2: This initial remote sensing data is fed into the now trained cycleGAN and an upgraded and improved remote sensing dataset is now born.

- Step 3.3: The improved observation data feeds the recurrent neural network (RNN) to generate an action command.

- Step 3.4: The improved remote sensing data and the action command are fed to the variational autoencoder (VAE) algorithm and thus the new next state is generated. VAE is a neural network trained using an unsupervised method, where it is given remote sensing data and an action command and outputs new remote sensing data.

- Step 3.5: The next state is fed to the reward function and thus the reward is generated. The reward function consists of two parts. The first part yields a positive reward if the NDVI and NDMI values are within a predefined range and a negative reward if they are outside the range. The second part aims to penalize the agent if he uses extensive amounts of water and if he produces too frequent watering instructions. The two sections stack as a single reward for simultaneously training the network to maintain soil moisture in a healthy range, with the least amount of water applied to the most efficient number of irrigations.

- Step 3.6: The new improved remote sensing data, action, reward and next state are stored in the replay buffer.

- Step 3.7: In the next state the meteorological data is replaced by the actual new meteorological data provided by the relevant external source. This creates a new overview.

- Step 3.8:. The new overview data again feeds the cycleGAN input and the cycle continues to repeat for a predetermined number of steps.

- Step 3.9: When the cycle is completed, the data collected in the replay buffer is used to train the recurrent neural network (RNN). During the training process, the RNN parameterization is updated in such a way that the reward is maximized.

When the training is completed, a new cycle of simulations is started and new data is collected in the repetition memory and the network is trained further.

VAE (Welling, 2014), (Wierstra, 2014) networks, are automatic probabilistic encoders. A VAE maps its input data to the parameters of a probabilistic distribution, such as mean and variance. It can create a continuous structured latent dataset, i.e. a dataset that is not directly observed but inferred from others that are observed.

This makes them ideal for generating synthetic data and using it as a transition function within an augmented learning facility environment.

Figure 4 shows the steps included in the method during Phase C:

1 References

Efros, J.-Y. Z. (2020). Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks. Retrieved from https://arxiv.org/abs/ 1703.10593#

Welling, D. P. (2014). Auto-Encoding Variational Bayes. Retrieved from https://arxiv.org/abs/ 1312.6114

Wierstra, D. J. (2014). Stochastic Backpropagation and Approximate Inference in Deep Generative Models. Retrieved from https://arxiv.org/abs/ 1401 .4082

Claims (1)

Claims Main Method Claim Economical method of determining the quantity and timing of irrigation based on observational data (satellite remote sensing and meteorological), with the aim of achieving ideal soil moisture conditions for crops, in order not to require incremental on-site measurements, which includes the following phases: - Phase A : Data Collection and Preparation - Phase B : Initial System Training - Phase C: Continuous operation and training - identification of irrigation actions Characterized by being performed on a physical infrastructure (System) that includes specific types of subsystems, in which the actions included in the steps defined in each phase are performed. The main actions concern the initial and ongoing training of specific types of neural networks that perform functions of improving observation data sets, and extracting optimal irrigation commands, given to flow-controlled automatic irrigation systems. Dependent claim 1 Method applied to optimize the accuracy of satellite ground remote sensing data using cycleGAN algorithms Dependent claim 2 Method implemented as an environment transition function for initializing the reinforcement learning algorithm Dependent claim 3 A method of using a reinforcement learning algorithm to train a recurrent neural network that generates commands for irrigation amounts and intervals. Method Implementation System Main Claim System composition including subsystems and means for implementing the method, according to claim 1, characterized by including the following subsystems: - Y1. Electronic material stage with built-in electronic processing stages that scale according to the cultivated area placed in a protection box which is connected to an array of analog sensors (Measurement Unit). The data from the various measuring units end up at the first level in a routing unit, through which they are sent to the Internet and then to Y4. - Y2. Arrays of analog sensors that are placed inside the ground and receive data - Y3. A wide area network consisting of a wireless local area network and an Internet access network or cloud computing infrastructure - Y4. Subsystem - Computational Processing Unit, which is installed in the form of a virtual machine (the array of them) in a cloud computing infrastructure (cloud - YYN) and in which the Database operates in which the data is stored and the various algorithms are executed in the context of the application of the method . - Y5. Subsystem - Irrigation Management Unit. Interfaces via internet or via Y3 to one or more standard electric flow controlled irrigation valves