Product Placement: Evolution from Models to Digital Ag Advisor

Machine learning and optimization models are digitally connecting our R&D pipeline to maximize value for farmers and helping Bayer redesign the testing network.

Overview

The 2018 discovery science and modeling efforts shifted from a regional corn product placement model, a key component of the current Seed Advisor from FieldView product, to a field-level seed placement tool. This shift enabled modeling to better account for field variability in environmental characteristics (e.g., soils) and farmer operations (e.g., crop rotation and tillage). The models were built on Bayer R&D pipeline data, which includes small research plots managed by the Breeding Organization and large strips managed by the Technology Development Organization. The objective of this blog is to highlight the different machine learning and optimization models driving seed selection decisions for the 2019 farmer alpha program and collaborations with Bayer Crop Science enterprise. Seed recommendation for farmers is driven by an integration of multiple models.

Successful testing of the regional level placement model paved the way for Seed Advisor

The regional-level placement model has been validated in the field with select farmers in Iowa, Illinois, and Minnesota in 2017 and 2018. Split planter trials were used as a demonstration of model predictiveness. These trials are configured as a Seed Advisor selected product planted side by side against a farmer selected product. Trials were successful nearly 80% of the time with an average >6 bushel per acre of lift using the Seed Advisor selected product.

New machine learning and optimization models will take Seed Advisor to field level in 2019

Several new models will enable Seed Advisor to take product recommendations down to the field level. This is highly valuable for farmers as it takes into account many key features that are of importance to them. These include features like pH, organic matter, and cation exchange capacity (CEC), elevation, drainage, tillage, and crop rotation. In total, there are >150 different features that the field level model uses to create a product recommendation.

Under the hood, there are currently four models that capture key decisions for recommending products:

1. The placement model is a machine learning model that ranks Bayer product performance on farmer fields using field characteristics, operations and management. Through past, head-to-head, independent year cross validations, the placement model has demonstrated an average of 7.8 bu/ac of statistically significant yield lift across >9600 research fields.
2. The relative maturity risk (RM risk) model is a machine learning model that predicts the probability to harvest a product the 20 - 25% kernel moisture range prior to the fall frost.
3. The portfolio model is an optimization model that uses the outputs from the above two models to recommend a portfolio across a farmer’s operation. Potential products are derived from the list of products available in the region of interest.
4. The field assignment model is an optimization model that is implemented after farmers make the final seed order. This step maximizes the overall yield outcome for the farmer’s entire corn operation by selecting top product for fields.

When conditions change, field guides ensure that farmers can select high performing products even if they cannot plant the top product from field assignment - for example, if that top product is unavailable.

The Alpha program

The alpha program is a pre-commercial program that takes models that have been built (using research data) and validated (against FieldView^TM data) and tests them on farm with farmers in the FieldView Innovators  program. Each farmer’s operation is different, and recommendations are built using knowledge about the specific farmer’s operation. The alpha program is a season long partnership that feeds insights back into the Climate LLC science team to refine models.

The alpha program started in 2017 with one farmer in central Illinois (~20K acres). It expanded in 2018 to Illinois, Iowa, and Minnesota for a total of ~100K acres. This year, the program expanded geographically to include Illinois, Iowa, Minnesota, Wisconsin, Indiana, and Missouri for a total of ~100K acres.

Feature Based Testing for Bayer Crop Science R&D

We worked across Bayer Crop Science and initiated internal use of hybrid selection model insights in Breeding and Tech Development & Agronomy for 2019 testing shifts. This work started by identifying top predictive features from field level seed placement model and a deep understanding of the data coverage by feature class in the past R&D testing. We then leveraged FieldView data insights to metricize coverage by feature class in testing networks for opportunity on site selection shifts. With the excellent collaboration, Breeding has identified additional new locations to better capture conditions in FieldView farmer fields.  

Bayer research pipeline network aligning with field seed selection model predictors in 2019