Precision Agriculture Applied to Organic Systems

Objective 1) Greenhouse Experiment 

The goal of this experiment was to determine optimum green manure (GM) seeding rates to maximize biomass production and following kamut (Triticum turanicum) yield as a result of nitrogen fixed by the GM and made available to kamut. Garden boxes measuring 50cm by 50cm and 25cm deep were filled with sterilized soil mixed with 50% sand in order to better emulate low nitrogen and sandy conditions typical of agricultural settings in north central Montana. The green manure tested was Arvika green peas, a common Montana crop and variety. Inoculated seeds were planted 3cm deep in rows 16cm apart; similar to how they would be planted in field conditions. Seeding rates encompassed a range well below and above typical farmer chosen rates and included pea rates at 0, 42, 78, 204, and 486 kg/ha (0, 7, 13, 34, and 81 seeds (150mg each) per box). Pea plants were grown for approximately eight weeks at optimal light and temperature growth conditions. Plants were terminated at first flowering stage, when roughly half the plants were flowering, which is the recommended termination stage to maximize N production and minimize water loss in arid conditions (McCauley et al., 2012). At termination, a subset of plants was collected, dried, and weighed to estimate biomass of the remaining plants which were incorporated into the soil. Pea plants were mixed into the soil and allowed to decompose over six weeks; soil was kept moist and occasionally stirred with a shovel.

Kamut seed was planted into the boxes at four different rates: 25, 50, 75, 100 seeds/box (30, 60, 90, 120 kg/ha). Each kamut rate was planted across each pea rate to create a full factorial design. Kamut was grown until maturity, which took approximately twelve weeks, then each head was harvested and weighed to determine crop yield by plant and by box. Soil was sampled to track nitrogen gain and loss in the soil at multiple time points: at pea planting, pea termination, kamut planting, and again following kamut harvest. An experimental oversight was noted in that pea seeds themselves contain approximately 4% nitrogen (personal communication, Perry Miller, 2022). The highest rate of pea seeding rate therefore saw an increase in nitrogen availability despite significantly reduced plant growth due to competition. Because we were seeking to understand how nitrogen fixation would be affected by plant densities, this top rate (which is well above any recommended seeding rate), was dropped from most of the analyses and the figures shown in results. When only nitrogen change in the soil was assessed we could still use these data points by accounting for the nitrogen added to the soil via the seeds themselves. Data was analyzed with regression analysis using generalized linear models to determine the relationship between GM seeding rate and kamut seeding rate and final kamut biomass and yield response.

Objective 2) On Farm Precision Experimentation of Seeding Rates and Harvest Yield

In this project on farm precision experimentation (OFPE) was used to adaptively manage inputs on organic farms. OFPE was used to optimize seeding rates of nitrogen-fixing GM crops, and non-GM cash crops such as wheat, barley, or kamut, and on one farm bloodmeal was varied to determine its optimum input rate. Rates were varied randomly across the field in order to understand within field spatial variation based on continuous georeferenced yield and protein monitor combine data. As all organic farmers face weed challenges, variation in weed density was mapped using a range of methods (Maxwell et al., 2005).

A total of four separate Montana certified organic farms collaborated on this project, and each farm conducted experiments on at least one field. Farmer collaborators include Bob Quinn and Seth Goodman, Casey Bailey, Ole Norgaard, and Ty O’Connor. To maintain anonymity farms henceforth are simply referred to, in no particular order, as farms A:D. Farm A is near Big Sandy, farm B near Fort Benton, farm C near Shonkin, and finally farm D near Ekalaka. The first three farms conducted OFPE with seeding rates of GMs and cash crops, and farm D experimented with varied rates of bloodmeal, an organically approved source of nitrogen, on wheat cash crops. The acreages of the fields vary from 32 to 93 hectares.

Initially one field per farm was extensively soil sampled (approximately one sample for every two acres). Farm B and farm D added a second field to the experimentation and these fields were not soil sampled due to cost restrictions. The samples were analyzed for nitrogen content in Montana State University’s Environmental Analytics Laboratory (EAL) and were used to build a fertility map of each field. Wherever possible, the fertility map, previous yields, and other geographic variables were used to stratify GM seeding rate strips across the field to maximize variability within each randomized replicated strip.

Experiments were initially designed to incorporate five rates of pea at 0X, 1.5X, 1.0X, 0.75X, and 0.5X farmer selected rates, these were to be planted with standard planters owned by each farmer. Given differences in equipment and ever-changing field and weather conditions the exact number and level of different rates was not followed precisely. In year two, differing seed rates of cereal crop were to be planted on top of the GM variable strips such that each GM rate will have every level of the cereal crop overlaid. Cereal seed rates were to similarly be 1.25X, 1.0X, 0.75X, and 0.5X of recommended organic rates. These were guidelines, and in practice rates were adjusted by farmers’ specific needs as seen below.

The farms joined the project at different times and experiments therefore started in different years. In 2019 farmer A planted peas on a 32 hectare field (soil sampled) at rates of 0, 60, 90, and 120kg/ha, and tilled them in after approximately eight weeks. In 2020 they planted barley over the same field at 60, 90, and 120 kg/ha, unfortunately a storm caused hail damage that wiped out 90% of the crop, and the remaining plant matter was removed as hay bales. Barley was again planted at the same rates as in 2020 in the spring of 2021, and harvested at the end of July 2021. In 2019 farmer B planted winter wheat at 60, 90, and 120 kg/ha across a 40 hectare field site (soil sampled); this field was severely wind damaged to the point where spring wheat was reseeded in a patchwork that made yield data collection based on seeding rate variation meaningless. In spring they also planted peas at 70, 90, 110, 130 kg/ha on a different 32 hectare field (not soil sampled), and planted winter wheat in fall 2020. The farmer chose not to vary the winter wheat seeding rates, but the experiment still provided important data on the effects of the varied pea rates. In spring 2020 farmer C planted peas at rates of 80, 100, 120, and 140 kg/ha across a 93 hectare field (soil sampled) and tilled them in approximately eight weeks later. Winter wheat was planted in fall 2020 across the field at varied rates of 55, 70, 85, and 100 kg/ha. Farmer D planted spring wheat at a uniform rate across a 32 hectare field (soil sampled) and a 67 hectare field (not soil sampled). Bloodmeal was applied across both fields at nitrogen rates of 0, 3.5, 7, 15, 20 kg/ha. Yield data were collected from the combine monitors. Farmer D was set to repeat this experiment in 2021 however suffered severe crop loss across their farm due to drought. The other three farms reported harvest data from cereal crops after harvest in late summer 2021.

Crop and weed data were collected during the growing season to track how the different seeding rates affected crop and weed growth. Approximately 60 points were chosen at random across each field, and stratified on seeding rate, in order to non-destructively survey plant growth. A wire frame with an area of 0.25m² was placed on the ground at the point and two researchers viewed its contents to estimate crop and weed cover. Each researcher estimated the amount of crop canopy coverage in the circle from 0 to 100 percent. To reduce bias the researchers then shared their numbers and recorded the average. The height of an average looking crop plant was then also recorded, and the percent cover and height was multiplied in order to estimate the volume that the crop occupies. Finally, several crop plants were measured, then collected and dried to create a regression of plant volume to biomass in order to convert plant volumes into estimates of plant biomass. This method was additionally conducted for the three most dominant weed species within the wire frame at each of the points across the field. Each species can have up to a total of 100% canopy cover. To further verify seeding rate affect across the field, crop counts were collected from across the field to determine germination rate and actual plant number within each seeding rate across the field.

Each farmer used their own field equipment, thus dimensions of experimental strips for each farm were slightly different depending on width of seeder and combine. GPS units on tractors and combines, and geospatially referenced yield monitors on the combines were already in place on the farmer’s equipment. Automatically aggregated data for each field includes as-applied maps of variable seeding rates, yield points, NDVI, and topographic information (elevation, slope, aspect). As-applied maps were collected from the seeder computer, yield points were collected from combine mounted yield monitors, and NDVI and topographic information were compiled via google earth engine of satellite information from NASA’s Modis and Landsat 8 instruments (He et al., 2018). Data were analyzed using linear and non-linear regression and random forest with R statistics packages in R Studio (Version 1.1.453 – © 2009-2018 RStudio, Inc.). Maps for producers and other reports including this one were produced in QGIS, and R Studio.

Modeling:

A data framework, already functioning in our lab for nitrogen fertilizer optimization across whole fields in conventional systems, was adjusted and updated for organic agriculture . The current R package that our lab has developed to produce an economic comparison among different nitrogen fertilizer management strategies (e.g. site-specific variable rate, no fertilizer, uniform farmer selected rate, uniform optimum rate, and organic with no fertilizer) was updated to compare site-specific variable GM seeding rate, no GM (tillage based fallow), uniform farmer selected GM seeding rate, and optimum uniform GM seeding rate based on farmer net return (https://paulhegedus.github.io/OFPE-Website/).

Statistical Analysis:

Point data from the seeder (as applied map) and combine (yield map) were added to a square 10m x 10m grid of the field. Means of points were taken when more than one point existed per cell. To remove border effects, headlands, as measured by 30 m from the field edge, were removed from analysis. From this accumulated data set random forest models were constructed to find optimized seeding rates for minimized weed presence and maximized yield. Moran’s I test, linear models and random forest models were tested in R, on data sets assembled in QGIS. Moran’s I test was used to verify that the seeding rate variable interacted spatially and affected weeds and yields differently in different parts of the field. Likely variables to be included in random forest models were verified using R package Boruta (Kursa and Rudniki, 2010). The random forest model used for each field varied based on input data, but a typical model is shown below in equation one.

[Equation 1]     Yield ~ pea_SR, barley_SR, elevation, x, y, slope, aspect, TPI, weed2019, weed2020, weed2021, NDVI2py, NDVIpy, NDVIcy, precipCY, sat_bulkdensity, sat_clay, sat_pH, sat_water, sat_carbon, nitrate6, nitrate24

The random forest model seen above included yield as the response variable to seeding rate of pea and wheat, and topographic variables: elevation, x and y coordinate points, slope, aspect, and Topographic Position Index (TPI); the season’s weed values as found from interpolated sample data; satellite data including: Normalized Difference Vegetation Index (NDVI) from the harvest year, the year prior to harvest, and two years prior to harvest, the current year precipitation, bulk density, clay content, pH, water content, and soil carbon; and soil sample data including: nitrate to six inches, and nitrate to 24 inches. Based on the predictions of this model ideal seeding rates were estimated for maximized yield across each grid cell. As an example, a subset of the variables collected for the above method is shown in maps in Figure 1 including topographic variables (1A), satellite variables (1B), manually sampled variables (1C), and machine variables (1D).

Economic Analysis:

Net returns were calculated based on farmers’ stated costs of seed and price received for product. Prices received in our tested harvest years ranged from $13 a bushel to $23 a bushel. To compare these to historic conditions a data set tracking organic prices (received from elevators) over the last twenty years was used, we also included varied fixed rates over these years as collected from USDA (see table 1). To ascertain the probability that one method would outcompete another, a Monte Carlo simulation analysis was conducted using the twenty years of economic data such that the simulation outcome of yield and recommended seeding rates was tested across every year of economic data. In this way we state the probability that one method (farmer chosen rates, optimum average rates, or variable seeding rate) is better than the others given varied economic conditions.

Table 1: Twenty years of economic data collected from local Montana elevators from 2000-2021.

Figure 1A – 80 acre On Farm Precision Experimentation Field showing topographic input variables for modeling including panel 1) elevation, panel 2) slope, panel 3) aspect, and panel 4) topographic position index (TPI).

Figure 1B – 80 acre On Farm Precision Experimentation Field showing satellite input variables for modeling including panel 1) pH, panel 2) water content, panel 3) carbon content (%), and panel 4) clay content (%).

Figure 1C – 80 acre On Farm Precision Experimentation Field showing on the ground collected data for modeling including panel 1) nitrate 0-6”, panel 2) nitrate 6-24”, and panel 3) sampled weed biomass in cubic meters per square meter.

Figure 1D – 80 acre On Farm Precision Experimentation Field showing on the machine collected data for modeling including panel 1) experimental pea seeding rates from 2019, panel 2) experimental barley seeding rates from 2020 (repeated in 2021 due to hail 2020), and panel 3) combine yield results from barley harvest 2021.

Objective 3) Computer Model and App

A data framework, already functioning in our lab for nitrogen fertilizer optimization across whole fields in conventional systems, was adjusted and updated for organic agriculture . The current R package that our lab has developed to produce an economic comparison among different nitrogen fertilizer management strategies (e.g. site-specific variable rate, no fertilizer, uniform farmer selected rate, uniform optimum rate, and organic with no fertilizer) was updated to compare site-specific variable GM seeding rate, no GM (tillage based fallow), uniform farmer selected GM seeding rate, and optimum uniform GM seeding rate based on farmer net return (https://paulhegedus.github.io/OFPE-Website/). The model compares different wheat seeding rates on top of each of the GM seeding rate strategies (Figure 1 under Methods). A final output of the model generates a profit maximizing site-specific GM seeding rate and wheat seeding rate map (Figure 2 under Results). This study is intended to be the beginning of a much longer study, as the predictive power of an OFPE based model increases over time. Crucial to the long term implementation of this methodology is automation of OFPE and minimizing farmer knowledgebase and time commitment required to apply OFPE results (https://sites.google.com/site/ofpeframework/home). The longer study will maintain GM and wheat seeding rate experimentation, and incorporate any other input experiments farmers are interested in pursuing.

All code is part of a package built in the R language and environment. The first step in producing experimental maps for farmers was to access a communal lab database. Within this database historical spatial data was stored including: farm boundaries and field boundaries, yield and protein data from combine monitors, and any experimental input rates from sprayers, spreaders, and seeders. Many other data types for each field are also freely collected via Google Earth Engine comprising topographical information including slope, aspect, and topographic position index, weather data including precipitation, and growing degree days, vegetation indices including NDVI, NDRE, and CIRE, and soil moisture from SSURGO. A script built in R was used to aggregate data from the database for a particular field. The aggregated data include: field name, farmer name, spatial coordinates, previous experimental year yield, previous experimental year seeding rate, topographic information, current year precipitation up to March 1 (point of decision making), precipitation of the previous year, growing degree days of the current year up to March 1, vegetation index of the current year up to March 1, vegetation index of the previous year, and vegetation index of two years previous.

            This aggregated data file was then used to generate field prescriptions based on the desired type of experiment. Optimum rates were chosen at each point on the field based on maximized net returns, and experimental rates were organized in random blocks over portions of the field to increase optimization power in the future. For the current version of the model visit: https://github.com/paulhegedus/OFPE. Moving forward, new experiments can be generated that use the previous years of seeding rate experiments to find optimum field rates.

Objective 4) Educational Outreach

Progress and results of this project have been presented at multiple meetings and conferences, and as a guest lecture in classroom settings to students, farmers, and agriculture professionals, and will continue to be presented electronically and at meetings as appropriate.

Results beyond those already published are targeted to be published in Agriculture Ecosystems and the Environment, Nutrient Cycling in Agro Ecosystems, Agriculture, and Precision Agriculture. Proposed manuscript titles include:
1. Variable seeding rate of green manures to affect nitrogen production
2. Variable application of wheat seed to maximize yield in an organic rotation
3. Precision agriculture as a method to increase quality and quantity of organically produced wheat