ASABE Logo

Article Request Page ASABE Journal Article

Forecasting of Canopy Temperatures Using Machine Learning Algorithms

Manuel A. Andrade1,*, Susan A. O'Shaughnessy2, Steven R. Evett2


Published in Journal of the ASABE 66(2): 297-305 (doi: 10.13031/ja.15213). 2023 American Society of Agricultural and Biological Engineers.


1Agriculture, Veterinary and Rangeland Sciences, University of Nevada Reno, Reno, Nevada, USA.

2USDA-ARS, Conservation and Production Research Laboratory, Bushland, Texas, USA.

* Correspondence: andradea@unr.edu

Submitted for review on 1 June 2022 as manuscript number NRES 15213; approved for publication as a Research Article and as part of the Artificial Intelligence Applied to Agricultural and Food Systems Collection by Associate Editor Dr. Srinivasulu Ale and Community Editor Dr. Yiannis Ampatzidis of the Natural Resources & Environmental Systems Community of ASABE on 12 January 2023.

Mention of company or trade names is for description only and does not imply endorsement by the USDA. The USDA is an equal opportunity provider and employer.

Highlights

Abstract. Recent advances can provide farmers with irrigation scheduling tools based on crop stress indicators to assist in the management of Variable Rate Irrigation (VRI) center pivot systems. These tools were integrated into an Irrigation Scheduling Supervisory Control and Data Acquisition System (ISSCADAS) developed by scientists with the USDA-Agricultural Research Service (ARS). The ISSCADAS automates the collection of data from a network of wireless infrared thermometers (IRTs) distributed on a center pivot’s lateral and in the field irrigated by the center pivot, as well as data from a wireless soil water sensor network and a microclimate weather station. This study analyzes the use of Artificial Neural Networks (ANNs), a type of machine learning algorithm, for the forecasting of canopy temperatures obtained by a wireless network of IRTs mounted on a three-span VRI center pivot irrigating corn near Bushland, TX, during the summer of 2017. Among the predictors used by the ANNs were weather variables relevant to the estimation of evapotranspiration (i.e., air temperature, relative humidity, solar irradiance, and wind speed), irrigation management variables (irrigation treatment, irrigation scheduling method, and the amount of water received during the last 5 days as irrigation or rainfall), and days after planting. Two case studies were conducted using data collected from periodic scans of the field performed during the growing season by running the pivot dry. In the first case, data from the first three scans were used to train an ANN, and canopy temperatures estimated using the ANN were then compared against canopy temperatures measured by the network of IRTs during the fourth scan. In the second case, data from the first six scans were used to train ANNs, and canopy temperatures estimated using the ANN were then compared against canopy temperatures measured by the network of IRTs during the seventh scan. The Root of the Mean Squared Error (RMSE) of ANN predictions in the first case ranged from 1.04°C to 2.49°C, whereas the RMSE of ANN predictions in the second case ranged from 2.14°C to 2.77°C. To assess the impact of ANN accuracy on irrigation management, estimated canopy temperatures were fed to a plant-stress-based irrigation scheduling method, and the resulting prescription maps were compared against prescription maps obtained by the same method using the canopy temperatures measured by the network of IRTs. In the first case, no difference was found between both prescription maps. In the second case, only one plot (out of 26) was assigned a different prescription. Results of this study suggest that machine learning techniques can be used to assist the ISSCADAS in situations where canopy temperatures cannot be measured by the network of IRTs due to poor visibility conditions, or because the center pivot cannot traverse the field within a reasonable amount of time.

Keywords. Artificial Neural Network, Canopy temperature sensing, Center pivot irrigation, Irrigation scheduling, Machine learning, Metamodeling, Variable rate irrigation.

Agriculture, like many other economic sectors, is facing a rapid transformation resulting from the use of data analytics (Pham and Stack, 2018). Machine learning is a data analysis technique that can be used to solve a wide range of problems since it uses historical data to train computational ‘machines’ that can make predictions on new data with no human intervention. An Artificial Neural Network (ANN) is a type of machine learning algorithm inspired by the way that the human brain processes information to acquire knowledge. Hornik et al. (1989) characterized ANNs as “universal approximators,” meaning that ANNs can approximate any measurable function to any desired degree of accuracy. Therefore, ANNs are capable of handling complex associations between inputs and outputs. This characteristic makes ANNs particularly well-suited for solving agricultural problems where no known deterministic model may be available to simulate the complex interactions occurring between the many components involved in an agricultural system. In agricultural engineering, ANNs have been used to, among other applications, predict soil properties (Schaap et al., 1998; Farifteh et al., 2007), estimate crop yield (Kaul et al., 2005), model greenhouse environmental variables (Ferreira et al., 2002; He and Ma, 2010), analyze images for the discrimination of crop and weeds (Yang et al., 2000; Kavdir, 2004), predict reference canopy temperatures for the estimation of a Crop Water Stress Index (CWSI) (King et al., 2020), and estimate of evapotranspiration (Yamac and Todorovic, 2020). Additional details of ANNs can be found in several textbooks (Hagan et al., 1996; Kartam et al., 1997; Lingireddy and Brion, 2005).

This study uses ANNs to forecast canopy temperatures obtained by a wireless network of infrared thermometers (IRTs) mounted on a three-span Variable Rate Irrigation (VRI) center pivot irrigating corn near Bushland, TX, during the summer of 2017. The network of IRTs is part of an Irrigation Scheduling Supervisory Control and Data Acquisition System (ISSCADAS) (Evett et al., 2020). The ISSCADAS collects data from soil water, plant, and weather sensing systems and feeds those data to computerized irrigation scheduling algorithms based on plant stress to generate site-specific prescription maps. The ISSCADAS also integrates hardware functions to manage the submission of prescription maps for their application by VRI center pivot systems operated with a Pro2 control panel (Valmont Industries Inc., Valley, NE). A software package, named ARSPivot, was developed to simplify the operation of the ISSCADAS (Andrade et al., 2020a, b). The ISSCADAS was the result of experience accumulated over multiple studies conducted at Bushland, TX, using VRI center pivots and irrigation scheduling methods based on plant stress (O’Shaughnessy and Evett, 2010; O’Shaughnessy et al., 2012, 2013, 2015, 2017; Peters and Evett, 2008). These studies demonstrated that plant sensing feedback could be used to intensively manage irrigated crop production with little labor while still obtaining yield and crop water productivity (yield per unit of water consumed) values comparable to those obtained using labor-intensive Neutron Probe (NP) measurements (Evett et al., 2008). Evett et al. (2020) described the development of the plant-based and plant and soil water-based site-specific irrigation scheduling methods incorporated in the ISSCADAS and summarized relevant research preceding the development of these methods.

The main objective of this study is to analyze the feasibility of using ANNs to predict canopy temperatures based on (1) previous measurements of canopy temperatures made by the aforementioned network of IRTs as the center pivot moves through the field, (2) current weather information, and (3) recent water management records. Results obtained from this study can provide guidelines for the future integration of canopy temperature forecasting using ANNs, one of the tools available in ARSPivot. The availability of such a tool can add redundancy to the ISSCADAS so that site-specific prescription maps can be generated even if a direct measurement of canopy temperatures is not possible due to poor visibility conditions (e.g., during an overcast day) or because the center pivot cannot traverse the entire field within a reasonable amount of time.

Materials and Methods

Irrigation Management

In the summer of 2017, the ISSCADAS and the ARSPivot software were used for the integrated irrigation management of a three-span center pivot (131 m) located at the USDA-ARS Conservation and Production Research Laboratory, near Bushland, TX. The center pivot was equipped with a Pro2 control panel and a commercial VRI system (Valmont Industries Inc., Valley, NE). A midseason corn hybrid, Dupont Pioneer P1151AM, was planted on 15 May, day of year (DOY) 135. Experimental plots used in this study were located within the six outermost sprinkler zones in the field (fig. 1).

VRI zone control was used for the north-northwest (NNW) side of the field. With VRI zone control, variable irrigation amounts are controlled along the lateral pipe of an irrigation system by equidistant solenoid valves regulating the flow of groups of sprinklers. The field was divided into six control sectors of 28° each and six concentric control zones with a width of 9.14 m (30 ft) each, for a total of 36 management zones, each of which was considered an experimental plot. Plots were organized using a Latin square design (fig. 1). VRI speed control was used for the south-southeast (SSE) side of the field. With VRI speed control, variable irrigation amounts are controlled by changing the irrigation system’s speed. The field was divided into eight control sectors of 20° each and a single concentric control zone with a width of 54.9 m, for a total of eight management zones, each of which was considered an experimental plot (fig. 1).

The irrigation of plots on the NNW side was triggered by either the integrated Crop Water Stress Index (iCWSI) method described by O’Shaughnessy et al. (2017) or by weekly neutron probe (NP) (model 503DR1.5, Instrotek, Campbell Pacific Nuclear, Concord, CA) measurements. Each of these plots was assigned one of the following irrigation levels: 80%, 50%, or 30% of full irrigation. Full irrigation was defined as the irrigation required to refill the soil water content in the root zone to field capacity. The combination of two irrigation scheduling methods and three irrigation levels resulted in six treatments with six replicates per treatment (fig. 1). Plots irrigated with the iCWSI method are labeled in figure 1 as C80, C50, or C30, where ‘C’ stands for iCWSI-based control and numbers correspond to irrigation levels. Similarly, plots irrigated with the NP method are labeled in figure 1 as U80, U50, or U30, where ‘U’ indicates that irrigation scheduling is controlled by the user.

Plots on the SSE side were all assigned a single irrigation level of 80%; their irrigation was triggered by either the iCWSI method, or by a hybrid method using the iCWSI method and an average soil water depletion in the root zone (SWDr) calculated as described by Andrade et al. (2020a) using sets of three time domain reflectometer (TDR) sensors (model 315L, Acclima, Meridian, ID) buried at depths of 15 cm, 30 cm, and 45 cm. The hybrid method used a two-step approach for irrigation scheduling. During the first step, the SWDr was compared against pre-determined lower and upper SWDr thresholds. No irrigation was assigned if the SWDr was lower than 0.1 (lower threshold), and an irrigation depth of 30.5 mm (1.2 in) was assigned if the SWDr was higher than 0.5 (upper threshold). If the SWDr fell between these values, the iCWSI method was used during a second step to determine its prescription. Plots irrigated with the hybrid method are labeled in figure 1 as H80. Additional details of the hybrid method can be found in Andrade et al. (2020a).

Figure 1. Experimental setup as displayed in ARSPivot software. Numbers inside of plots preceded by letter ‘p’ indicate numbers used to identify experimental plots. Green squares represent approximate location of soil water sensing stations based on Time Domain Reflectometry (TDRs). Two-small yellow circles inside well irrigated area (w1) indicate approximate location of field Infrared Thermometers (IRTs). Red line represents position of center pivot and small yellow triangles next to line indicate location of IRTs mounted on center pivot. VRI zone control was used for NNW side of field delimited by angles of 244° and 52°. VRI speed control was used for SSE side of field that is delimited by angles of 228° and 68°.

The iCWSI method is based on the calculation of the theoretical Crop Water Stress Index (CWSI) (Jackson et al., 1981) at discrete intervals during daylight hours:

(1)

where

iCWSIx = integrated Crop Water Stress Index (iCWSI) calculated at any given location x inside experimental plots

Tc = canopy temperature at location x during time t estimated using a temperature scaling algorithm by Peters and Evett (2004)

Ta = air temperature at time t

(Tc - Ta)ll = lower limit of the difference between Tc and Ta occurring for a well-watered crop

(Tc - Ta)ul = upper limit of the same difference occurring for a severely stressed crop.

The lower and upper limits of the difference between canopy and air temperatures were calculated using formulas by O’Shaughnessy et al. (2012).

CWSI values were calculated using equation 1 for each location x in the field at each time interval t. The air temperature and other relevant weather parameters (relative humidity, solar irradiance, wind speed, and wind direction) were sampled every 5 s and averaged and stored every minute at a weather station (Campbell Scientific, Logan, UT) located next to the pivot point. Crop canopy temperatures were measured at two fixed locations in the field using wireless IRTs (model SapIP-IRT, Dynamax Inc., Houston, TX). At each of these two fixed locations, two static IRTs were placed with opposing views of the same well-irrigated area to provide a reference canopy temperature curve for a well-watered crop (fig. 1). A network of 12 mobile wireless IRTs was mounted on the center pivot to measure site-specific canopy temperatures inside the experimental area (fig. 1). The scaling algorithm by Peters and Evett (2004) was used to scale one-time-of-day canopy temperature measurements obtained from mobile IRTs against the reference temperature curve obtained from the static IRTs in a well-watered area of the field to estimate the canopy temperature (Tc) at any given location and time during daylight hours (from 2 h after sunrise to 2 h before sunset). Canopy temperatures obtained with this scaling algorithm were then used to calculate an iCWSI using equation 1 for locations in the field where a canopy temperature measurement was obtained using the mobile IRTs. The mobile IRTs were located forward of the drop hoses, at a 45° angle from the nadir and a 45° angle relative to the sprinkler lateral axis and turned towards the center of a sprinkler control zone. The average of the data collected from two mobile IRTs with opposing views of a sprinkler control zone was the primary datum every minute for each sprinkler zone. Pairing the IRTs with opposite views of the same sprinkler control zone and positioning them at the specified angles minimizes viewing of the soil background and minimizes sun angle effects (Evett et al., 2020). Pairs of mobile IRTs arranged in such a way are referred to hereinafter as ‘IRT groups.’ Additional details of this experiment can be found in Andrade et al. (2020b).

Artificial Neural Networks Forecasting Canopy Temperatures

Scans of the field were performed weekly through the growing season by running the center pivot dry. Weather data and canopy temperatures—measured by the network of stationary IRTs in the field and on the center pivot—collected during scans were used to train ANNs to estimate the average canopy temperatures obtained by a given IRT group, i.e., by a pair of IRTs with opposing views of a sprinkler zone. Two case studies were conducted to analyze the feasibility of using ANNs for this purpose. In the first case, six types of ANNs (one for each of the six IRT groups located on the center pivot) were trained using the data collected during the first three scans that took place on 26 June (DOY 177), 7 July (DOY 188), and 11 July (DOY 192). Since the training of ANNs is a semi-random process that yields different results every time, 50 ANNs were trained for each ANN type, and the best performing ANN among them was then selected to be used for the forecasting of average canopy temperatures that would be measured by the corresponding IRT group during the following scan (12 July, DOY 193). The accuracy of the best ANN selected for ANN type n was then assessed by predicting the average canopy temperatures that would be measured by IRT group n on this date. In the second case, six types of ANNs were trained using data collected during the first six scans that, in addition to the previous dates, took place on 17 July (DOY 198), and 20 July (DOY 201). 50 ANNs were also trained for each ANN type, with the best performing ANN selected to be used for the forecasting of average canopy temperatures that would be measured by the corresponding IRT group during the following scan (24 July, DOY 205).

Figure 2. Arrangement of elements in ANNs used in study. Elements in input, hidden, and output layers allocate 10, 12, and 1 neurons, respectively. Number of input neurons (represented by blue circles) correspond to number of variables considered relevant for estimation of average canopy temperatures estimated by given IRT group mounted on center pivot. Only output neuron (represented by green triangle) in ANN allocates such average temperatures. Hidden neurons are represented by orange squares.

The typical structure of an ANN (also known as its architecture) is composed of at least three layers of nodes (usually referred to as neurons) and the links between these layers (fig. 2) (Hagan et al., 1996). The first layer is the input layer, the last is the output layer, and the remaining layers are referred as hidden layers. Nodes in these layers are referred to as input neurons, output neurons, and hidden neurons, respectively. The ANNs used in this study were three-layered feed-forward networks consisting of sigmoid hidden neurons and linear output neurons. The number of input neurons was 10, corresponding to the number of variables that were considered relevant for the estimation of average crop canopy temperatures estimated by a given IRT group n mounted on the center pivot (fig. 1). These variables were: (1) air temperature measured at time t during a scan, (2) relative humidity at time t, (3) solar irradiance at time t, (4) wind direction at time t, (5) wind speed at time t, (6) average canopy temperature measured by stationary IRTs at time t, (7) irrigation level (%) assigned to the experimental plot p being scanned by IRT group n at time t, (8) irrigation scheduling method assigned to plot p, (9) the number of days passed since planting at the time of the scan, and (10) cumulative irrigation (including precipitation) received by experimental plot p during the last five days preceding time t. Wind direction is a relevant variable for the highly advective semi-arid conditions in Bushland, where wind direction has been shown to affect soil evaporation under wet soil conditions (Agam et al., 2012). The irrigation scheduling method and irrigation level assigned to a plot were considered to be relevant variables because together they can provide information to the ANN regarding the irrigation management received by the crop since the beginning of the growing season.

ANNs with a single output neuron are expected to be better estimators than ANNs with multiple output neurons (Andrade et al., 2016), and thus a single output neuron was selected for ANNs used in this study. The only output neuron allocated is the average canopy temperature measured by a given IRT group mounted on the center pivot (fig. 2). Using a single output neuron for ANNs in this study offers the additional advantage of allowing ANNs to account for conditions that may be exclusive to a single IRT group, such as scanning a sprinkler zone with a clogged nozzle. Furthermore, in a center pivot, IRT groups in inner spans collect less canopy temperature data than pairs of IRTs in outer spans because the latter travel a longer distance. As a result, training an ANN with a single output neuron for each IRT group aids in avoiding the potential bias that can occur when an ANN is trained with data from multiple IRT groups. The number of neurons in the hidden layer was selected as 12 after running a series of preliminary tests with 2, 4, 6, …, up to 18 hidden neurons. These numbers were tested following rules of thumb for the selection of the number of hidden neurons (Heaton, 2008). Additional preliminary tests were performed to determine if the addition of a second hidden layer improved the accuracy of ANNs, but their accuracy did not show to be consistently better than ANNs with a single hidden layer. The training of ANNs was performed using Matlab’s neural network toolbox software (Mathlab, 2022). The Resilient Backpropagation method was selected for the training of ANNs since it performed consistently better than other training algorithms tested during preliminary runs. The other algorithms tested were Levenberg-Marquardt, Bayesian Regularization, and Conjugate Gradient.

The datasets used for the training of ANNs in the first case study can be represented by an input matrix with dimensions M by N and an output vector with M elements, where M is the total number of one-minute intervals occurring during days when the first three scans were performed in the growing season, and N is the number of input variables in the ANNs, i.e., 10 (fig. 2). The total number of one-minute intervals occurring during daylight hours (from 9:00 h to 19:00 h in this study) of a scan day was thus 600 (10 h times 60 one-minute intervals). The first row in the input matrix contained the values recorded for each input variable during the first one-minute interval, the second row contained the values recorded during the second interval, and so on. The output vector, on the other hand, contained the average canopy temperatures measured by an IRT group at each one-minute interval. The data contained in the training datasets obtained in this way were normalized using a Z-score normalization to account for the differences in the magnitudes of the input variables. The percentages of data in these datasets allocated for the training, validation, and testing of ANNs were 70%, 15%, and 15%, respectively.

Results and Discussion

The time series of the average crop canopy temperatures estimated by ANNs and measured by IRT groups mounted on the center pivot is displayed for the first and second case studies in figures 3 and 4, respectively. On 12 July (DOY 193), the scan started at 11.3 h at an angle of 227º (fig. 1). The center pivot then advanced in a counter-clockwise direction through the SSE side of the field and entered the NNW side at approximately 13 h. The scan was completed at 14.2 h when the pivot reached 248°. Since all IRT groups scanned experimental plots on the SSE side (where the highest irrigation level was assigned to all plots) before 13 h, measured canopy temperatures before this time tended to be lower than those obtained on the NNW side (where irrigation levels varied) after this time (fig. 3). Nevertheless, ANNs could approximate the oscillating pattern displayed by measured canopy temperatures through the scan (fig. 3), with a Root Mean Squared Error (RMSE) that ranged from 1.04°C to 2.49°C (table 1). The mean difference between the canopy temperatures measured by pairs of IRTs and the canopy temperatures estimated by ANNs were found to be significantly different for IRT groups 1, 5, and 6 at the 0.05 probability level based on a two-sample t-test (table 1). Differences between canopy temperatures measured by IRTs and estimated by ANNs for IRT groups 1 (fig. 3a), 5 (fig. 3e), and 6 (fig. 3f) were more pronounced between 13 h and 14.5 h, when the center pivot traversed the NNW side of the field—where different irrigation treatments were applied. This result suggests that ANNs trained to estimate canopy temperatures can be more accurate when a single irrigation treatment is applied to a field.

To assess the impact of using ANNs for irrigation management, their estimated canopy temperatures were used by the iCWSI and hybrid methods to recalculate the prescriptions of experimental plots using these methods. No difference was found between the prescription map obtained with canopy temperatures estimated by ANNs and the prescription map obtained with canopy temperatures measured by IRTs. Hence, the accuracy of all ANNs tested in the first case study can be deemed satisfactory for irrigation scheduling based on the quantification of plant stress.

Regarding the second case study, the scan started on 24 July (DOY 205) at 11 h at an angle of 52º (fig. 1). The center pivot then advanced in a counter-clockwise direction through the NNW side of the field and entered the SSE side at approximately 12.5 h. The scan was completed at 13.7 h when the pivot reached 68°. Similar to the first case study, measured canopy temperatures tended to be lower as the center pivot advanced through the SSE side of the field, i.e.,

after 12.5 h. As in the first case, ANNs were capable of approximating the oscillating pattern displayed by canopy temperatures through the scan (fig. 4), with an RMSE that ranged from 2.14°C to 2.77°C (table 1). The mean difference between the canopy temperatures measured by pairs of IRTs and the canopy temperatures estimated by ANNs was found to be significantly different for all IRT groups at the 0.05 probability level based on a two-sample t-test (table 1). This result may be due to a higher variability of canopy temperatures measured by IRTs for the second case study, where the standard deviation of canopy temperatures ranged from 1.84 to 2.35 (table 1), whereas the standard deviation of canopy temperatures for the first case study ranged from 1.24 to 2.24 (table 1). The largest errors made by ANNs trained for the second case study occurred at the beginning of the scan, when the position of the center pivot was near 52° at 11 h, and near the middle of the scan, when the position of the center pivot was near 228° at 12.5 h (fig. 4). Angles of 52° and 228° are the boundaries of the east and west roads in the field, respectively (fig. 4). Hence, the peak canopy temperatures measured by IRT groups near these angles may have considered a portion of bare soil, which led to the measurement of high temperatures near these angles, which in turn were difficult to estimate by the ANNs (fig. 4). In spite of this, when comparing the prescription maps obtained with canopy temperatures estimated by ANNs and canopy temperatures measured by IRTs, only one plot (out of 26 assigned either the iCWSI or hybrid methods) was assigned a different prescription (fig. 5). Therefore, the accuracy of all ANNs tested in the second case study can also be deemed satisfactory for irrigation scheduling based on the quantification of plant stress.

It should be noted that the methodology followed in this study was motivated by the need to support the generation of site-specific irrigation prescription maps using the iCWSI and hybrid methods incorporated by the ISSCADAS when canopy temperatures cannot be collected by IRTs mounted on a center pivot. Hence, the input data selected for the ANNs used in this study were selected based on input data that is relevant to both irrigation scheduling methods. Results from this study corroborate that the selected input data—consisting of current canopy temperatures measured by static IRTs, current weather information, and recent weather management records—are relevant predictors for the estimation of canopy temperatures when these cannot be measured by mobile IRTs mounted on a center pivot. The incorporation of additional input data to the ANNs, such as soil water sensing data, may improve the canopy temperatures estimated by the ANNs, but the iCWSI method does not require soil water sensing data, so this type of data was not considered in this study. Similarly, the proposed methodology was applied to estimate the canopy temperatures of corn. Although the same methodology can be applied to different crops, further studies are recommended to determine if ANNs or other machine learning algorithms can be trained to accurately forecast the canopy temperatures of multiple crops. Although this approach was developed to make the ISSCADAS more robust in the event of data loss, it could be applied to other irrigation scheduling methods that incorporate canopy temperatures into their decision support algorithm.

Figure 5. Prescription maps generated by ARSPivot using canopy temperatures (a) measured by network of wireless IRTs mounted on center pivot and (b) estimated by ANNs using data collected on 24 July 2017 (DOY 205). Prescriptions are displayed as percentages of pre-specified maximum irrigation depth of 30.5 mm (1.2 in). Only one plot (p8) received different prescription when using canopy temperatures estimated by ANNs.

Conclusions

Machine learning is a data analysis technique that can be used to solve a wide range of problems since it uses previous

data to train computational machines that can make predictions on new data with no human intervention. This study analyzes the feasibility of using Artificial Neural Networks (ANN), a machine learning technique, to estimate canopy temperatures in a field irrigated by a Variable Rate Irrigation (VRI) center pivot system. Results indicate that the machine learning technique tested can predict canopy temperatures with satisfactory accuracy for plant stress-based irrigation scheduling. The Root Mean Squared Error (RMSE) of ANN predictions in the first case ranged from 1.04°C to 2.49°C, whereas the RMSE of ANN predictions in the second case ranged from 2.14°C to 2.77°C. Machine learning technology can be useful to add redundancy to an Irrigation Scheduling Supervisory Control and Data Acquisition System (ISSCADAS) patented by scientists from ARS (Bushland, Texas). In addition to providing canopy temperature data when it is missing, the addition of machine learning capabilities to the ISSCADAS can also assist users when poor visibility conditions prevent the correct estimation of canopy temperatures using the network of sensing systems incorporated by the ISSCADAS.

Acknowledgments

This material is based upon work that was supported by the National Institute of Food and Agriculture, U.S. Department of Agriculture, under award number 2016-67021-24420. The work described in this paper was also completed as part of a cooperative research and development agreement between USDA-ARS and Valmont Industries, Inc., Valley, NE (Agreement No.: 58-3K95-0-1455-M), as well as with assistance from the University of Nevada, Reno, Office of Research and Innovation and the Nevada Agricultural Experiment Station (NAES) of the College of Agriculture, Biotechnology, and Natural Resources.

References

Agam, N., Evett, S. R., Tolk, J. A., Kustas, W. P., Colaizzi, P. D., Alfieri, J. G.,... Chávez, J. L. (2012). Evaporative loss from irrigated interrows in a highly advective semi-arid agricultural area. Adv. Water Resour., 50, 20-30. https://doi.org/10.1016/j.advwatres.2012.07.010

Andrade, M. A., Choi, C. Y., Lansey, K., & Jung, D. (2016). Enhanced artificial neural networks estimating water quality constraints for the optimal water distribution systems design. J. Water Resour. Plann. Manag., 142(9), 04016024. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000663

Andrade, M. A., O’Shaughnessy, S. A., & Evett, S. R. (2020a). ARSPivot, a sensor-based decision support software for variable-rate irrigation center pivot systems: Part A. Development. Trans. ASABE, 63(5), 1521-1533. https://doi.org/10.13031/trans.13907

Andrade, M. A., O’Shaughnessy, S. A., & Evett, S. R. (2020b). ARSPivot, a sensor-based decision support software for variable-rate irrigation center pivot systems: Part B. Application. Trans. ASABE, 63(5), 1535-1547. https://doi.org/10.13031/trans.13908

Evett, S. R., Heng, L. K., Moutonnet, P., & Nguyen, M. L. (2008). Field estimation of soil water content: A practical guide to methods, instrumentation, and sensor technology. Vienna: International Atomic Energy Agency.

Evett, S. R., O’Shaughnessy, S. A., Andrade, M. A., Colaizzi, P. D., Schwartz, R. C., Schomberg, H. S.,... Sui, R. (2020). Theory and development of a VRI decision support system: The USDA-ARS ISSCADA approach. Trans. ASABE, 63(5), 1507-1519. https://doi.org/10.13031/trans.13922

Farifteh, J., Van der Meer, F., Atzberger, C., & Carranza, E. J. (2007). Quantitative analysis of salt-affected soil reflectance spectra: A comparison of two adaptive methods (PLSR and ANN). Remote Sens. Environ., 110(1), 59-78. https://doi.org/10.1016/j.rse.2007.02.005

Ferreira, P. M., Faria, E. A., & Ruano, A. E. (2002). Neural network models in greenhouse air temperature prediction. Neurocomputing, 43(1), 51-75. https://doi.org/10.1016/S0925-2312(01)00620-8

Hagan, M., Demuth, H., & Beale, M. (1996). Neural network design. Boston, MA: PWS Publishing Company.

He, F., & Ma, C. (2010). Modeling greenhouse air humidity by means of artificial neural network and principal component analysis. Comput. Electron. Agric., 71(Suppl. 1), S19-S23. https://doi.org/10.1016/j.compag.2009.07.011

Heaton, J. (2008). Introduction to Neural Networks with Java (2nd ed.). St. Louis, MO: Heaton Research, Inc.

Hornik, K., Stinchcombe, M., & White, H. (1989). Multilayer feedforward networks are universal approximators. Neural Networks, 2(5), 359-366. https://doi.org/10.1016/0893-6080(89)90020-8

Jackson, R. D., Idso, S. B., Reginato, R. J., & Pinter Jr., P. J. (1981). Canopy temperature as a crop water stress indicator. Water Resour. Res., 17(4), 1133-1138. https://doi.org/10.1029/WR017i004p01133

Kartam, N., Flood, I., & Garret, J. (1997). Artificial neural networks for civil engineers: Fundamentals and applications. Reston, VA: American Society of Civil Engineers.

Kaul, M., Hill, R. L., & Walthall, C. (2005). Artificial neural networks for corn and soybean yield prediction. Agric. Syst., 85(1), 1-18. https://doi.org/10.1016/j.agsy.2004.07.009

Kavdir, I. (2004). Discrimination of sunflower, weed and soil by artificial neural networks. Comput. Electron. Agric., 44(2), 153-160. https://doi.org/10.1016/j.compag.2004.03.006

King, B. A., Shellie, K. C., Tarkalson, D. D., Levin, A. D., Sharma, V., & Bjorneberg, D. L. (2020). Data-driven models for canopy temperature-based irrigation scheduling. Trans. ASABE, 63(5), 1579-1592. https://doi.org/10.13031/trans.13901

Lingireddy, S., & Brion, G. M. (2005). Artificial neural networks in water supply engineering. American Society of Civil Engineers: Reston, VA.

Matlab. (2022). Statistics and machine learning toolbox. Retrieved from https://www.mathworks.com/help/stats/?index.html?s_tid=CRUX_lftnav

O’Shaughnessy, S. A., Andrade, M. A., & Evett, S. R. (2017). Using an integrated crop water stress index for irrigation scheduling of two corn hybrids in a semi-arid region. Irrig. Sci., 35(5), 451-467. https://doi.org/10.1007/s00271-017-0552-x

O’Shaughnessy, S. A., Evett, S. R., & Colaizzi, P. D. (2015). Dynamic prescription maps for site-specific variable rate irrigation of cotton. Agric. Water Manag., 159, 123-138. https://doi.org/10.1016/j.agwat.2015.06.001

O’Shaughnessy, S. A., Evett, S. R., Colaizzi, P. D., & Howell, T. A. (2013). Wireless sensor network effectively controls center pivot irrigation of sorghum. Appl. Eng. Agric., 29(6), 853-864. https://doi.org/10.13031/aea.29.9921

O’Shaughnessy, S. A., & Evett, S. R. (2010). Canopy temperature based system effectively schedules and controls center pivot irrigation of cotton. Agric. Water Manag., 97(9), 1310-1316. https://doi.org/10.1016/j.agwat.2010.03.012

O’Shaughnessy, S. A., Evett, S. R., Colaizzi, P. D., & Howell, T. A. (2012). A crop water stress index and time threshold for automatic irrigation scheduling of grain sorghum. Agric. Water Manag., 107, 122-132. https://doi.org/10.1016/j.agwat.2012.01.018

Peters, R. T., & Evett, S. R. (2004). Modeling diurnal canopy temperature dynamics using one-time-of-day measurements and a reference temperature curve. Agron. J., 96(6), 1553-1561. https://doi.org/10.2134/agronj2004.1553

Peters, R. T., & Evett, S. R. (2008). Automation of a center pivot using the temperature-time-threshold method of irrigation scheduling. J. Irrig. Drain. Eng., 134(3), 286-291. https://doi.org/10.1061/(ASCE)0733-9437(2008)134:3(286)

Pham, X., & Stack, M. (2018). How data analytics is transforming agriculture. Bus. Horiz., 61(1), 125-133. https://doi.org/10.1016/j.bushor.2017.09.011

Schaap, M. G., Leij, F. J., & van Genuchten, M. T. (1998). Neural network analysis for hierarchical prediction of soil hydraulic properties. Soil Sci. Soc. Am. J., 62(4), 847-855. https://doi.org/10.2136/sssaj1998.03615995006200040001x

Yamaç, S. S., & Todorovic, M. (2020). Estimation of daily potato crop evapotranspiration using three different machine learning algorithms and four scenarios of available meteorological data. Agric. Water Manag., 228, 105875. https://doi.org/10.1016/j.agwat.2019.105875

Yang, C. C., Prasher, S. O., Landry, J. A., & DiTommaso, A. (2000). Application of artificial neural networks in image recognition and classification of crop and weeds. Can. Agric. Eng., 42(3), 147-152.