Development of an Online Tool for Tracking Soil Nitrogen to Improve the Environmental Performance of Maize Production

Abstract

Freshwater nitrogen (N) pollution is a significant sustainability concern in agriculture. In the U.S. Midwest, large precipitation events during winter and spring are a major driver of N losses. Uncertainty about the fate of applied N early in the growing season can prompt farmers to make additional N applications, increasing the risk of environmental N losses. New tools are needed to provide real-time estimates of soil inorganic N status for corn (Zea mays L.) production, especially considering projected increases in precipitation and N losses due to climate change. In this study, we describe the initial stages of developing an online tool for tracking soil N, which included, (i) implementing a network of field trials to monitor changes in soil N concentration during the winter and early growing season, (ii) calibrating and validating a process-based model for soil and crop N cycling, and (iii) developing a user-friendly and publicly available online decision support tool that could potentially assist N fertilizer management. The online tool can estimate real-time soil N availability by simulating corn growth, crop N uptake, soil organic matter mineralization, and N losses from assimilated soil data (from USDA gSSURGO soil database), hourly weather data (from National Weather Service Real-Time Mesoscale Analysis), and user-entered crop management information that is readily available for farmers. The assimilated data have a resolution of 2.5 km. Given limitations in prediction accuracy, however, we acknowledge that further work is needed to improve model performance, which is also critical for enabling adoption by potential users, such as agricultural producers, fertilizer industry, and researchers. We discuss the strengths and limitations of attempting to provide rapid and cost-effective estimates of soil N availability to support in-season N management decisions, specifically related to the need for supplemental N application. If barriers to adoption are overcome to facilitate broader use by farmers, such tools could balance the need for ensuring sufficient soil N supply while decreasing the risk of N losses, and helping increase N use efficiency, reduce pollution, and increase profits.

1. Introduction

Nutrient losses from agriculture are a significant concern from tile-drained landscapes in the U.S. Midwest [1,2]. Building on the work and recommendations of the Mississippi River/Gulf of Mexico Watershed Nutrient Task Force over the past two decades (https://www.epa.gov/ms-htf/history-hypoxia-task-force, accessed on 16 December 2020), individual states have developed nutrient loss reduction strategies outlining the most effective in-field and edge-of-field practices for reducing nitrogen (N) leaching losses from agriculture and other sources. Among the different recommendations, changes in N management are particularly important because of the combined influence on crop yields and water quality. Moreover, compared to other practices, there is potential for immediate impact through voluntary adoption because changes in N management can result in cost savings and have relatively greater stackability (the ability to pair management practices) and trackability (the ability to track practice implementation) [3]. By avoiding excessive N inputs, multiple studies have highlighted the opportunity to optimize crop yields while decreasing N losses [4,5].

A major factor influencing efforts to reduce N losses is precipitation [6]. Particularly in tile-drained fields, large precipitation events contribute to higher drainage volumes, consequently increasing nitrate export [7]. Studies at both the field-level [4,5] and regional-level [8] have highlighted the combined negative influence of increased precipitation and N inputs on water quality. More recently, using a dataset across the entire U.S. to develop empirical models, the variables of precipitation, land use, and N inputs explained nearly 70% of the variation in annual N loading to waterways [9]. Of concern is that precipitation is expected to become more variable and extreme under climate change in the U.S. Midwest, further increasing N losses [10]. One study estimated that N inputs would have to decrease by more than 30% to offset the anticipated increase in riverine N loading due to changes in future precipitation patterns, including extreme precipitation events [11].

There is an urgent need for site-specific N management tools to ensure sufficient soil N supply while minimizing the risks of N losses [12]. Soil N concentration is a key variable to monitor when trying to match soil N supply with crop demand, especially early in the growing season before the major period of crop N uptake. For example, the late spring soil nitrate test is used in Iowa to quantify the amount of soil N available relative to an established threshold, helping guide in-season N fertilizer management decisions [13]. In a modeling study comparing many different N management strategies, Mandrini et al. [14] found that soil N concentration at V5 corn growth stage was the most influential variable influencing the economic optimum N rate relative to other climate or soil factor. Meanwhile, increasing amounts and intensity of precipitation during springtime, which are becoming more frequent with climate change, can decrease soil N concentrations. Puntel et al. [15] reported that excess precipitation during early season corn establishment and growth (April–June) led to exponential increases in N losses through leaching and denitrification, thereby reducing soil N supply. In addition, excessive precipitation has been shown to negatively influence regional crop yields [16].

In the U.S. Midwest, farmers sometimes respond to wet weather early in the season by applying additional N fertilizer to corn, believing that some of the N previously applied was lost to leaching or denitrification. Thus, uncertainty in soil N supply can lead to unnecessary increased N inputs, directly conflicting with farm profitability and environmental goals. Model-based decision support tools which have the potential to estimate soil N concentration could reduce this uncertainty, while avoiding the high costs and time requirements associated with manual soil sampling [12,17]. Yet, simulating soil N concentration during the growing season is challenging due to weather uncertainties and interactions with many soil-crop processes [18]. Several modeling tools have been developed in the private sector and are available to farmers, but these are proprietary products, and there is a lack of published information describing the modeling mechanisms or validating model performance. We are currently unaware of an online, user-friendly, publicly available tool allowing farmers to enter management information and estimate soil N concentration for their fields in this region, which would be especially useful after periods of wet spring weather. Another benefit of public tools is the transparency in describing the steps of model development, calibration and validation, and assessment of uncertainty and limitations.

The objective of this study was to describe the development of an online tool for tracking soil N status in Illinois, a state that contributes around 20% of total annual N loads, from the Mississippi River Basin to the Gulf of Mexico [19]. Several stages of research were necessary to complete this process, each described below. First, soil N concentrations were monitored over many site-years to develop an empirical database and quantify the spatial and temporal variability in soil mineral N (SMN hereafter: nitrate-N + ammonium-N) concentration in response to different N management practices and soil and weather conditions. Second, model calibration and validation was performed to enable process-based simulations of SMN for specific fields, accounting for crop management practices and in-season weather conditions. Third, an online platform was developed to enable rapid estimates of SMN using a combination of inputs related to farmer-entered crop management, geospatial soil information, and real-time weather data to drive model simulations. The motivation of this study is to provide an example research pipeline that could be applied in other regions to develop similar tools by explaining the steps required for field work, crop modeling, and computer programming. This pipeline is also essential for understanding the limitations of model accuracy and identifying future research and outreach efforts necessary to improve model performance and chances of adoption.

2. Soil N Concentration Data Collection in Field Experiments

The methods for conducting field experiments to monitor SMN over time have been previously reported by Banger et al. [20] and Preza-Fontes et al. [21]. Briefly, the database of SMN concentrations contained a total of 1149 observations collected from 32 site-years of N management studies conducted at research stations and on-farm trials across Illinois between 2015 and 2018 (

Table A1. Geographical location of the field experiments in Illinois. Research trials were located at the University of Illinois Crop Sciences Research & Education centers near DeKalb (DKB), Monmouth (MON), Urbana (URB), Perry (ORR), Neoga, (NEO), Simpson (DSP), and Brownstown (BRT). On farm trials were located in Christian County (CHR), McLean County (MCL), Sangamon County (SNG), and Ford County (FRD).

Site	Latitude	Longitude	Year Present
DKB	41.93	−88.75	2015–2018
MON	40.91	−90.65	2015–2018
URB	40.11	−88.21	2015–2018
ORR	39.89	−90.75	2015–2018
CHR	39.71	−89.18	2015, 2017, and 2018
MCL	40.56	−88.96	2015–2018
SNG	39.88	−89.44	2015 and 2016
FRD	40.48	−88.17	2017 and 2018
NEO	39.24	−88.44	2016–2018
DSP	38.95	−88.96	2015
BRT	37.45	−88.72	2015

in Appendix A). In these experiments, SMN was measured 2–3 times after fall N application, starting in November and running through March. The sampling frequency increased in the spring, with soil samples taken at 10-d to 14-d intervals from spring N application to VT-R1 corn growth stage. At each sampling event, composite soil samples were collected from two replications of each treatment at two depths (0–30 and 30–60 cm) and analyzed for nitrate-N and ammonium-N concentrations. Soil N concentrations were averaged across depths, resulting in SMN (sum of nitrate-N and ammonium-N) at 0–60 cm depth. The N management treatments consisted of 12 different combinations of N fertilizer rates, timing, and source, plus a control treatment with zero N (see Preza-Fontes et al. [21] for a full description of N treatments). In addition to these treatments, the multiple locations across years created a range of SMN concentrations under different soil and weather conditions.

Figure 1 shows daily precipitation and SMN at the research stations for the treatments that received 224 kg N ha⁻¹ as anhydrous ammonia in the fall or spring, with or without a nitrification inhibitor (N.I.), plus the control. Research trials were located at the University of Illinois Crop Sciences Research & Education centers near DeKalb (DKB), Monmouth (MON), Urbana (URB), and Perry (ORR), in Illinois (Table A1). As expected, SMN varied considerably among sites under the fertilized treatments across the 4 year study period. For instance, SMN before corn planting in 2015 (the first sampling event in late April) was ~2.1-fold greater in DKB (northern Illinois) than in ORR (central Illinois) with the 224 N Spring–N.I. treatment (50 vs. 24 mg N kg⁻¹). Similarly, pre-plant SMN with the 224 N Spring + N.I. treatment was 47 mg N kg⁻¹ in ORR compared to 14 mg N kg⁻¹ in URB (west-central Illinois) in 2016.

Despite differences in SMN between the different N management treatments, the overall pattern of SMN was temporally similar within each site-year (Figure 1). In general, SMN was high between April and mid-May, due to previous N application and soil organic matter (SOM) mineralization, and started to decline after mid-May or early June as corn developed and plant N uptake increased. Because corn N uptake is relatively low early in the growing season (<17 kg N ha⁻¹ by V5–V6 growth stage) [22], either fertilizer- or soil-derived SMN is subject to losses by leaching and/or denitrification after heavy rainfall during this period.

Across the different locations, results showed the potential for early season SMN to drop substantially when cumulative rainfall was excessive in some years (Figure 1). However, field data did not show large decreases in SMN at other site-years, illustrating that the fate of applied N fertilizer is highly variable due to interactions among weather, soil properties, and soil N transformations. It is important to highlight that just because no change in SMN was observed at some sites, this does not necessarily mean that there was no loss of SMN within the top 60 cm. Rather, it is possible that the amount of N lost was approximately the same amount as was produced by the mineralization of SOM during this period. In other site-years, SMN increased during this time, especially with low to moderate amounts of cumulative precipitation. Because it is a biological process, the rate of mineralization depends mainly on temperature and moisture [23]. Thus, large increases in SMN were likely driven by a combination of sufficient moisture with warmer soils, for example during May 2016 and 2018 at some site-years.

3. Model Calibration and Validation Steps for Predicting Soil Mineral N Concentration

Using the extensive field dataset above, we carried out a first attempt to calibrate and validate a model for predicting SMN concentration under variable rainfall patterns, soil properties, and N fertilizer management scenarios in Illinois. For this, we selected a publicly available, widely used process-based model, the Decision Support System for Agrotechnology Transfer (DSSAT). Details about the modeling portion of this work have been recently published in two separate studies [20,24]. In brief, our approach relied on using calibration and validation data from 6–7 field experiments focusing on N management practices and from 49 commercial cornfields to estimate changes in soil N status during the growing season. A two-step strategy was used to evaluate model performance in predicting SMN concentrations [20]. The first step used three independent field experiments that contained SMN at the same interval as the experiments used for calibration. The second step used SMN observations from 49 commercial cornfields in Illinois. An additional validation exercise was conducted through scenario analysis to assess the effects of rainfall on the simulation results. For this, we evaluated SMN changes under different precipitation amounts across the commercial cornfields assuming similar planting date and N management practice.

Table 1. Model performance criteria for predicting soil mineral nitrogen (SMN) concentration during the first and second step calibration strategy. Parameters included normalized root mean square error (nRMSE), coefficient of residual mass (CRM), and index of agreement (D-index). Adapted from Banger et al. [20].

1st Step Calibration		2nd Step Calibration
Parameter	Value	Parameter	Value
nRMSE 1	25.7%	nRMSE	21.2%
CRM 2	−0.036	CRM	0.017
D-index 3	0.88	D-index	0.91

¹ nRMSE values between 20 and 30 indicate “fair” model performance. ² positive and negative values indicate the model underestimated and overestimated SMN concentrations, respectively. ³ Values close to 1 indicate model predictions are consistent.

shows the results of three indicators of model performance.

Table 2. Summary analysis of DSSAT model performance for predicting soil mineral nitrogen concentration for four different cumulative precipitation ¹ categories on 49 commercial cornfields in Illinois. N, number of replicated data; slope, the “a” from the y = ax regression; R², from the y = ax regression. Adapted from Banger et al. [20].

Rainfall Category	N	Slope	R2
500–600 mm	16	1.247	0.685
600–700 mm	11	0.996	0.883
700–800 mm	29	0.968	0.716
>800 mm	12	1.147	0.815

¹ Cumulative precipitation occurring between January and July in 2015 and 2016.

shows the results of linear regression between observed and measured SMN for four cumulative precipitation categories.

Overall, the model produced normalized root mean squared error (nRMSE) values of 21.2–25.7% for estimating SMN, indicating the ability to simulate interactions between SOM mineralization dynamics, crop N uptake, and environmental N losses (Table 1). Index of agreement (D-index) values of 0.88 and 0.91 indicated that predicted variation was close to the observed variation in SMN concentrations in both calibration steps. In the scenario analysis, the model also captured the variability in SMN concentrations across 49 commercial fields under different cumulative precipitation scenarios (Table 2; R² = 0.68–0.88; slope, 0.99–1.24). Of specific relevance to the present work was the sensitivity of model predictions to increasing cumulative rainfall amounts across the 49 on-farm study sites. We investigated this by running model simulations and comparing SMN concentration 60 days after corn planting at sites experiencing moderate versus high precipitation. When rainfall was between 500 and 600 mm during January through July, the average SMN concentration was 23.7 ± 9.6 mg N kg⁻¹ soil. In contrast, this value decreased by 46% on the same date (16.1 mg N kg⁻¹ soil) for sites where cumulative rainfall exceeded 800 mm. These results indicate the potential to improve N management decisions based on tracking SMN concentrations and potential losses early in the growing season. Given the nRMSE values above 20%, however, we also recognize that further work is needed to improve model accuracy and understand what confidence level is required by farmers to use decision support tools for N management. For a deeper description of model uncertainty and assumptions, as well as considerations for adoption, please see [20].

In addition to the SMN data from field sites described above, we also calibrated and validated the model for predicting corn yield under different N management scenarios [24]. Specifically, the effects of N fertilizer amount and timing on maize grain yield were validated using 15 independent field experiments. Model performance criteria based on nRMSE (17.5%) and the D-index (0.91) suggested the ability of the model was “good” in predicting corn grain yield (

Table 3. Summary analysis of DSSAT model performance for predicting corn grain yields. Simulations included 15 field experiments including different N fertilizer amounts and timing treatments. N, number of replicated data; slope, the “a” from the y = ax regression ¹; R², from the y = ax regression; normalized root mean square error (nRMSE), coefficient of residual mass (CRM), and index of agreement (D-index). Adapted from Banger et al. [24].

N	Slope	R2	nRMSE	D-Index	CRM
66	1.0149	0.849	17.5%	0.91	−0.01

¹ Linear regression between simulated and observed corn grain yield (Mg ha⁻¹).

). In contrast, a coefficient of residual mass (CRM) index value of –0.01 indicated slight overestimation for model predictions compared to observed values.

In the same study, multiple scenarios were investigated using a gridded model simulation approach to understand the spatial and temporal variability of outcomes at the watershed-scale across different weather years [24]. One of these included shifting N fertilizer application timing from fall to spring, while simultaneously reducing N rates, with the goal of decreasing N pollution without negatively impacting yield. Results showed that this approach either simultaneously maintained or increased grain yield and N use efficiency less than half of the time (29–50% of cases) when changes in N timing from fall N application (224 kg N ha⁻¹) was combined with an N rate reduction of 25% in two spring N application scenarios (168 kg N ha⁻¹) across 2011–2015. Similar to the SMN simulations described above, the largest differences in crop yield and N use efficiency among fall and spring treatments occurred when winter and early spring precipitation was excessive, triggering environmental N losses. Even when N rates were reduced by 15%, rather than 25%, positive impacts on both grain yield and N use efficiency occurred in only 60% of simulations, highlighting the challenge of simultaneously improving yield and nutrient use efficiency through N rate reductions in this region.

A unique approach in this modeling framework was that the calibration and validation steps occurred using many different site-years of field experiment data. Large amounts of variability present a challenge in selecting appropriate model parameters during calibration. As a result, many crop modeling studies generally focus on investigating processes in one or two field sites, where calibration can be achieved with a single set of parameters, but this then limits the inference space for model application. To handle the observed variability in SMN across sites, we developed eight sets of SOM decomposition parameters based on combinations of soil organic carbon content and soil drainage rates [20]. We made this decision because the initial steps in our calibration process indicated that SMN concentration was most sensitive to these two factors. For example, with the original SOM decomposition parameters, the predicted SMN concentration was extremely low in most cases, especially where organic carbon was <1% and soils were classified as “well-drained”. Developing different sets of parameters was a crucial step in making the model more widely applicable, helping us meet the long-term goal of developing an online decision support tool.

4. Development of an Online Decision-Support Tool

4.1. Weather Data and Methods

Based on the validated DSSAT model, we developed an online Internet tool that provides user-friendly interfaces and operations for predicting SMN availability, while requiring minimal user inputs. Our project was built on the premise that such information could help support in-season N management decisions by understanding the risk of soil N losses, thereby improving the economic and environmental sustainability of N management in the U.S. Corn Belt. An important first step was to incorporate an improved weather data source compared to the previously published studies abovementioned [20,24]. This earlier version was driven by data inputs from local weather stations and soil parameters from the Gridded Soil Survey Geographic (gSSURGO) dataset [25]. The soil dataset has a spatial resolution of 10 m. In the original model, weather parameters such as daily solar radiation, maximum and minimum air temperature, rainfall amount, and relative humidity were obtained from 19 weather stations in the Illinois Climate Network (http://www.isws.illinois.edu/warm/datatype.asp, accessed on 16 December 2020). Below we describe how datasets for real-time weather were integrated with field-level management and soil information into an online platform to track soil N status during the growing season in cornfields throughout Illinois.

Inputs to the online tool include soil property data, daily weather data, soil conditions before planting, and crop management information (planting date, N application dates, and N rate). Soil property data were directly extracted from the gSSURGO dataset [25]. The drained upper and lower limits, saturation, drainage coefficients, and runoff curve number were estimated based on soil profile properties [26].

The online tool uses the National Weather Service (NWS) weather data (Real-Time Mesoscale Analysis, 2.5 km, and 1 h resolution) downloaded automatically to the server daily. The data are saved on the server and processed to CSV format daily to decrease processing time when users request to run the tool. The data include hourly 2 m air temperature, 10 m wind speed and direction, cloud cover, and precipitation. The DSSAT needs daily total solar radiation, daily maximum and minimum air temperature, rainfall, longitude, latitude, elevation, long-term average air temperature, and amplitude of the warmest and coolest monthly long-term average temperatures, and wind height and temperature data.

Total solar radiation is calculated using equations in [27] based on longitude, latitude, elevation, and cloud cover. The longitude and latitude are input data from the online tool user. Elevation data of 1 km resolution is obtained from MOD03 data (MODIS satellite) (https://modaps.modaps.eosdis.nasa.gov/services/about/products/c6/MOD03.html, accessed on 16 December 2020).

The server contains NWS data from 2005 to the present. The amplitudes of warmest and coolest monthly long-term average temperatures are calculated from these historical data.

Crop management information is provided by the user and includes hybrid maturity in growing degree days, planting date, date of N fertilizer application, and rate. In most cases, the user will be the farmer; hence this information is readily available.

In addition to soil property and weather data, DSSAT requires information about initial soil conditions (soil moisture and N content) at the site. This information is not readily available, however. We used initial soil N conditions for seven sites across Illinois each year from 2015 to 2018 [24] to set the initial conditions, using the nearest site to the user’s location. This information can also be updated later as more data become available.

4.2. Online Tool Structure

The online tool tracks real-time (daily frequency) SMN content (in lb N acre⁻¹ = kg ha⁻¹ × 0.89) in cornfields for user-defined locations in Illinois (flowchart in Figure 2). As highlighted above, SMN concentration is considered a useful indicator to guide crop management decisions. It is the net balance of multiple processes governing crop response to N fertilizer, including SOM mineralization, N fertilizer transformations, crop N uptake, and environmental losses.

The online tool can be accessed at http://rsetserver.sws.uiuc.edu/ntrack (accessed on 16 December 2020). A user registers first (name and email address). After logging in, the user can inquire about SMN availability in real-time for their field by inputting N application (fertilizer history), crop information (growing degree days for maturity, planting date), simulation end date (default: current day), and location (latitude and longitude or by clicking on Googlemap) (Figure 3). In Illinois, because precipitation is usually sufficient, most cornfields are rainfed; therefore, the interface does not include irrigation as a management option. Each request is stored in a queue, and the online tool is ready to receive other requests. The queued requests are processed on a first-in, first-out (FIFO) basis.

Next, the real-time daily downloaded weather data and soil data are automatically prepared for the DSSAT v4.6 model. The server will assimilate the data and user inputs to simulate plant growth, N uptake, and N loss pathways (including leaching and denitrification) using the DSSAT v4.6 model. The simulation process takes 3–5 min, and the real-time SMN results are provided to the user by email. After successful processing, the request is deleted from the queue.

The simulation results of SMN are provided in a time-series graph with date on the X-axis and SMN content on the Y-axis. Figure 4 shows an example of the simulation results using the crop management information in Figure 3, with starting simulation date 1 November 2017 and end date 15 July 2018, and 200 lbs N acre⁻¹ (224 kg N ha⁻¹) applied on 8 November 2017. The results include SMN content (lbs N acre⁻¹) in three depths: 0–1 feet (red line), 1–2 feet (blue line), and 0–2 feet (light blue line) (1 feet = 30.5 cm).

5. Discussion

5.1. Challenge: Managing N Fertilization for Sustainable Crop Production

Illinois and other states in the U.S. Midwest have developed nutrient loss reduction strategies to meet nitrate loss reduction targets of 15% by 2025 and 45% ultimately [19]. In these strategies, management decisions related to N fertilizer application are highlighted as an important opportunity to address water quality concerns without negatively affecting crop yields. However, managing N fertilizer to meet both production and environmental goals is challenging, in part because of the year-to-year variability in weather and soil conditions. The increasingly common occurrence of wet early-spring weather can contribute to uncertainty about soil N status, which in turn, increases the tendency for farmers to apply additional N. Previous studies have shown that early-season SMN is positively correlated with corn yield [13,28,29], suggesting that low SMN is associated with a risk of a yield penalty. In contrast, applying N above the recommended rates to maximize crop productivity can increase N losses and economic costs [21]. While there is a delicate balance between profitability and environmental N losses, a recent survey study in Illinois found that most of the participants (farmers) were concerned about the economics and environmental impacts of nutrient loss, with some participants reporting making changes in their production practices to minimize nutrient losses in their fields [30].

5.2. Online Tool: Potential Benefits and Contributions

The web tool presented here transforms the complicated operating procedures of DSSAT into a user-friendly interface that producers can run with relatively few management inputs. Moreover, real-time weather and soil data that can be difficult to prepare in the DSSAT model are automatically assimilated into the tool. Freely available online tools could provide new information in helping farmers understand how changing weather conditions and management practices influence soil N availability. For instance, when farmers are trying to determine whether soil N status has declined before the period of maximum crop N uptake, which would present a risk of crop yield penalties, model simulations could therefore provide rapid and cost-effective information, compared to manual soil N sampling. In situations where soil N status remains stable, such information could help limit supplemental N applications that are not warranted, which would also compromise profitability [31].

The need to improve N management is also important from an environmental perspective. With sufficient modeling accuracy, the potential value in the present tool is not only providing timely information to assist farmer’s management decisions, but possibly providing information that would help farmers avoid additional N applications when they are not necessary. Recent studies have shown that increased N inputs correspond to greater nitrate leaching losses and nitrous oxide emissions in corn-based systems, particularly when N rate exceeds plant N demand [5,32]. In addition, N inputs are the largest component of carbon footprint in the U.S. Midwestern cornfields [33]. Recent work has documented the potential for decreasing N losses when using adaptive N management strategies that incorporate site-specific soil information and in-season weather conditions [34]. Although there are several modeling platforms that provide soil N simulations for cornfields, particularly in the private sector, there is, as far as we can determine, no real-time soil N tracking tool publicly available for Illinois. Importantly, the online tool introduced here is driven by a calibrated and validated model that was documented through peer-reviewed publications, providing transparency about model performance and relative uncertainty in predictions.

5.3. Limitations and Future Work

It is important to highlight that the tool presented here is not the finished product but rather in the early stage of development. Given the limitations in model accuracy for predicting SMN in this study, future research is critical to continue improving model coefficients and calibration mechanisms across a wide variety of environmental conditions. This should include further field sampling for SMN at new sites, as well as investigating other crop modeling platforms that might better represent the study system and improve SMN prediction accuracy. In addition, model accuracy could be improved if information on initial SMN level is available [18]. Using the Agricultural Production Systems Simulator (APSIM) process-based model for predicting SMN, Archontoulis et al. [18] found that the model was most sensitive to initial conditions, including initial soil nitrate-N, soil water, and previous crop root C/N ratio. In future research steps, this option can be included in the tool that allows farmers to enter SMN information, if applicable.

Another limitation is the uncertainties related to confidence level for adoption. The nRMSE (relative difference between simulated and measured observations) values for predicting SMN were between 21% and 26%, indicating “fair” model performance. As we described earlier [20], many research efforts evaluating model performance for different soil and plant variables typically report ranges of 10% to 30% nRMSE. Still, it is not known whether some information (i.e., estimate of SMN produced by a freely available model) would hold more value compared to no information (i.e., no soil testing). In this context, we urge for greater transparency in discussing model accuracy and limitations in the private sector, considering the increasing emphasis on model-based decision support tools for N management. To our knowledge, models in the private sector display some uncertainty related to weather predictions, but do not explicitly report uncertainties related to model accuracy based on comparisons with measured observations.

Another important constraint to consider when developing decision support tools is farmer adoption. Because of its interaction with many soil-crop processes, SMN has large uncertainty and error within the context of simulating soil and crop variables [18]. On the other hand, one of the limitations of using soil testing for N recommendation is that it can be time-consuming and costly (i.e., sampling collection and lab analysis), in addition to in-field spatial variability of SMN with manual soil sampling. Given limitations with both approaches, it is not clear what level of accuracy is acceptable when developing agronomic recommendations for farmers. The threshold required by farmers to trust model simulation and associated N management recommendations is an important question that needs to be answered in future research. Previous studies has suggested a focus on farmer education and social learning for participatory development is needed for ensuring effective delivery and adoption of decision support tools [35]. Although the present model performance may be acceptable for researchers to investigate interactions among the soil-plant-water processes that govern N cycling in agroecosystems, or to evaluate different management and weather scenarios, we speculate that the confidence level required by farmers in estimating these variables is much higher. Farmers are managing risk in a highly uncertain context with large economic implications in rainfed systems. Even without specific knowledge of SMN, much research has advocated for reducing N rates, but considerable agronomic and economic consequences may exist at the field-level for farmers if negative yield impacts occur [36], indicating there are existing barriers to consider when proposing changes in N management.

The tool currently does not provide recommendations for additional N application, but this could be an area for future research. To translate our framework into a N rate recommendation tool, calibration curves would need to be established for the relationships between modeled soil N predictions and observed yield response, similar to the empirical relationships developed for the late spring soil nitrate test in Iowa [13]. Doing so would require future field research using a different experimental design to answer different research questions. Specifically, it remains unclear whether crop yields will respond to supplemental N addition in fields with low SMN values, especially applications during late vegetative growth stages; and moreover, how much additional N fertilizer should be applied to different levels of SMN deficiency. Therefore, compared to other research efforts focused on identifying optimal N rates, the current scope of our tool is simulating SMN as a first step in understanding N fertilizer requirements.

6. Conclusions

Farmers currently lack knowledge regarding the impacts of excessive precipitation on early-season soil N availability, which could prompt additional N application that conflicts with profitability and environmental goals. In this study, we describe the initial stages of developing an online tool for tracking soil N, which included, (i) implementing a network of field trials to monitor changes in soil N concentration during the winter and early growing season, (ii) calibrating and validating a process-based model for soil and crop N cycling, and (iii) developing a user-friendly and publicly available online decision support tool that could potentially assist N fertilizer management. The developed online tool can estimate soil N status dynamics in real-time by integrating a publicly available crop model (DSSAT) with producer-entered management information and gridded soil and climate data in a geospatial framework. Given limitations in prediction accuracy, however, we acknowledge further work is needed to improve model performance, which is also critical for enabling adoption by potential users, such as agricultural producers, the fertilizer industry, and researchers. If barriers to model accuracy and adoption are overcome to facilitate broader use by farmers, such tools could balance the need for ensuring sufficient soil N supply while decreasing the risk of N losses, and helping increase N use efficiency, reduce pollution, and increase profits. As N losses continue to be a growing concern in Illinois and throughout the U.S. Midwest, this online tool could be used to enhance our understanding of N management strategies for minimizing N losses while optimizing crop yield.