Machine learning methods for predicting site-specific profitability from sensor-based nitrogen applications

Machine Learning (ML) techniques were applied to predict the site-specific profitability of a sensor-based N application using data from the SENSE (Sensors for Efficient Nitrogen Use and Stewardship of the Environment) project. Our previous study found the profitability of implementing sensor-based N fertilizing varied from one study site to another and changed from one location to another within the same site. To better understand the relationship between site conditions and SENSE profitability, we developed a method to analyze SENSE project data at a finer scale. We first partitioned study strips into 20 feet long paired (Grower vs. SENSE) study units. We then created a 200-feet long aggregation block to move along each study strip to integrate N application data, yield data, and site condition data. Within each block, we compared yields and marginal net returns (MNR) from Grower and SENSE treatments and quantified the profitability of the sensor-based N application as the difference between SENSE MNR and Grower MNR. Three types of features were constructed to characterize site conditions within each aggregation block: 1) features derived from DEM, TWI, 2) features constructed to quantify the complexity of soil combination, and 3) features extracted from the SSURGO database on soil properties. The data processing generated 12000+ samples with 70+ features. Multiple supervised methods were applied to model the relationship between the site features and the profitability of the sensor-based N application (target outputs are -1 for nonprofitable, 1 for profitable). We compared the ML models and found that the Random Forest classification model performed the best with a close to 90% test accuracy.