Recipient Organization
PURDUE UNIVERSITY
(N/A)
WEST LAFAYETTE,IN 47907
Performing Department
Agronomy
Non Technical Summary
Nowadays, an increasing amountof attention isbeing paid to digital agriculture and its component technologies and techniques, such as precision agriculture, agricultural Big Data, remote and proximal sensing, and on-farm experimentation, and great hope has been placed in using these components to increase farm management efficiency.While those technologies and techniques are individually impressive, and despite public and commercial enthusiasm about them, numerous agriculturalists have expressed frustration with the current state of their use, and have recognized a need to bring them together systematically.The project proposed aims at developing such inclusing framework.This data-intensive farm management system will be based on on-farm precision experiments (OFPEs) (Bullock, et al. 2019; Mieno, et al. 2020; Trevisan, et al. 2020), which are large, field-sized randomized trials beginning to radically change agronomic research. Participating farmers will annually conduct OFPEs and use data to make economically and environmentally superior input management choices.
Animal Health Component
45%
Research Effort Categories
Basic
45%
Applied
45%
Developmental
10%
Goals / Objectives
Our first supporting objective is to lay the research groundwork needed to support an on-farm precision experimentation (OFPE) infrastructure, sufficiently automated to be scaled up to enable 1) the running of tens of thousands of OFPEs per year, worldwide;
2) the collection of field characteristics and weather data on the OFPE fields;
3) the processing and management of the resultant data, which will make it possible to apply advanced statistical and artificial intelligence methodologies to data analysis, thereby
4) providing farmers with management recommendations based on data from their own fields, as well as on data from other fields, worldwide.
The resultant increase in input use efficiency will not only
5) increase farm income, but also
6) enhance the nationâ¿¿s water quality, as more fertilizers enter plants as building blocks to growth and production instead of waterways (Martinez-Feria et al., 2018; Puntel et al., 2016).
Project Methods
We conceptualize the problem of crop input management in terms of four types of variables. The first variable, y, is crop yield. The second type we represent by a multi-element vector c, of unmanaged, spatially distributed "field characteristics." The third is a multi-element vector z, of unmanaged and temporally stochastic variables (principally, weather). The fourth is a multi-element vector x of "managed input variables," (e.g., fertilizer application rate, seeding rate). We conceptualize yield, y, as resulting from a natural process described by a function f, which depends on farmer choices, field characteristics, and weather: y = f(x, c, z).The fundamental research question for crop input management is,"What is f?"Agricultural scientists have been attempting to generate data to estimateffor major row crops for almost two hundred years (Odell, et al. 1984). Indeed, estimatingfwas motivated the advance of modern statistical theory, as R.A. Fischer (1935) developed his pioneering statistical research on randomization in experimental design to address estimating f using data from agronomic experiments (Antle 2019). The extraordinary efforts to learn aboutfhave been made becausefreveals how crop yields respond to producer choices, and how yield responses change with growing conditions. Better estimates offare key to better scientific support of farm management and all the social benefits that follow.Figure 1 provides the simplest of illustrations. Versions of figure 1 and explanations of the implications for producer management choices appear in most introductory microeconomics textbooks (e.g., Pindyck and Rubinfeld 2013, pp. 204-208), and are typically presented in the first few weeks of introductory microeconomics courses. It is assumed in figure 1 that there is only one managed input, x. The characteristics of a site A on which the crop is grown are cA. For simplicity, it is assumed that when a value of x is chosen, weather is assumed known and given as the vector of constant values z2019. The input price and output price are assumed constant at levels w and p. Elementary calculus shows that the profit-maximizing input application rate, x*A,19(p, w), is that value at which the slope of the yield response curve is equal to the price ratio, w/p. When multiple inputs are chosen, similar mathematical rules apply, optimizing over more than two dimensions. Space limitations prevent us from illustrating more complicated situations.Technical Feasibility of the WorkThe focus of the proposed project is on-farm precision experimentation (OFPE) and the analysis of OFPE data, which are illustrated in figure 2. The first panel of figure 2 shows the randomized design of a nitrogen rate OFPE on corn. In form, the design is like those of traditional small-plot trials, but covers a much wider area. The field in figure 2 is 37 ha in size. The second panel shows the trial being "put in the ground" using variable input rate application is accomplished by using GPS-based computer software to pre-program a variable application rate "plan" into a computer aboard farm machinery. That program "instructs" application equipment to apply inputs at the planned rates on the designated plots as the farmer just drives the farm equipment through the field in the usual manner. The final figure shows the resultant data.The aim is to build a research infrastructure and encourage the development of a commercial infrastructure that will permit tens of thousands of OFPEs to be designed and conducted farmers and their crop consultants annually, as well as the handling, processing, and analysis of the resulting data. The system would annually generate vastly more field trial data than has been generated since the first agronomic field trials were conducted in the first half of the nineteenth century.The main elements of the strategy to scale up data generation involve the creation of a cloud-based, "on-farm precision experiment design" software system. That system will allow crop consultants, who have received training from Extension personnel or others, to upload basic information about a farmer's field, such as a geo-referenced file of the field's border, and the sizes of the farmer's machinery, and then, with some "pointing and clicking," design a statistically legitimate agronomic experiment over the farmer's entire field. All data would be transferred wireless to and from farm machinery.The principal element of scaling up OFPE data analysis will be an automated "analytical engine," which can import OFPE data, and then with minimal "human-in-the-loop" effort, employ econometric analysis and a variety of machine learning methods to develop management recommendations. For example, methods based on reinforcement learning can be used to optimize prescriptions for seeding or applying fertilizers or herbicides. Learning algorithms such as those based on deep learning (e.g., convolutional networks or long short-term memories), locally (spatial or temporal) random forests, or Gaussian processes can be used to capture spatial and temporal properties in fields. Other techniques based on a variety of heuristics (local search) and metaheuristics (evolutionary and swarm-based search) in constraint satisfaction, Markov random fields, and even natural language models to describe agricultural processes can be applied to derive data-based farm management support.The statistical and machine learning models will be accompanied by responsive, interactive statistical visualizations to explain which data have the most impact on the predictions (the "why" of the statistical models), helping build the trust of farmers and consultants in the "black box" models used for prediction. In addition, OFPE data will be used to improve the calibrations of existing crop growth models, which in turn can be used to model yield response and estimate optimal management strategies. Automating OFPE data analysis will be a major challenge. We envision a future in which significant parts of crop science and agricultural economics academia devote themselves to this very task. But it will be feasible, and fascinating, to make a convincing start to this research endeavor in the five years of the proposed project.Automating the communication of the implications of a farm's OFPE data will involve the creation of a cloud-based "decision tool" software system, which displays in user-friendly, visual ways, OFPE data results and implications. Several such "decision tools" are currently available on commercial markets. Several of the 135 interviews of farmers, professional crop consultants, and extension personnel conducted by two of the writers of this proposal (Montesdeoca, et al. 2018) made clear that many farmers and consultants have limited faith in their effectiveness. The crop science and agricultural economics literature impels skepticism of current commercial decision tool software packages. They are based on long debunked yield-based input application algorithms (Rodriguez, et al. 2019) and crop growth models with parameters that in general cannot be well enough calibrated to allow the models to provide adequate management advice (Antle 2018).We believe that the scaling-up of that work is also feasible within the five-year timeline. But, certainly, the project faces risks. Data analysis is difficult, and to develop automated techniques for it that work well enough to provide farmers with profitable management recommendations will be a challenge. But this is why a multistate project is needed--we plan to put dozens of very skilled, highly talented researchers to the tasks at hand, for five years. We are confident that we can develop a workable infrastructure by then, and that the next generation of agronomic science and farm management rely upon, and continue to advance that infrastructure.