Assessing the impact of agricultural field traffic on corn grain yield using remote sensing and machine learning

•A framework to predict crop grain yield at a centimeter scale was developed.•A cubist algorithm predicted corn grain yield with the highest accuracy.•Findings based on high-resolution yield maps were consistent with hand-harvest data.•Corn grain yield influenced by soil types but not by planter induced compaction. With the increase in the physical size of agricultural field machinery, producers can experience negative effects of machinery-induced soil compaction on soil health and crop productivity. Traditionally, the impact of compaction from machinery traffic on crop yield has been examined either by hand harvesting of crops or by using yield maps generated using yield monitors mounted on combines at the time of harvest. While a hand harvesting approach is time consuming, laborious and subjective, yield maps generated from yield monitors are too coarse in resolution to truly assess the yield differences between individual crop rows. Remote sensing technology offers a cost- and time-effective approach to generate high-resolution yield maps that can be used to assess crop yields of individual rows or areas within a row and, thus, the impacts of field traffic on crop yield. The objectives of this study were to: (1) develop a framework to predict crop grain yield at both the centimeter scale and individual row level by integrating remote sensing and field collected data, and (2) use the predicted yield map to assess the impact of field traffic during planting on corn grain yield. Field experiments were conducted in three corn fields, each in 2016, 2017, and 2018, at the Molly Caren Agricultural Center, London, Ohio, to evaluate yield impact of field traffic at planting. Data collected included corn grain yield based on yield monitors and hand-harvesting, and aerial (i.e., visible and multispectral) imagery collected during the corn growing season. Corn grain yield was predicted using six machine learning algorithms (i.e., linear regression, random forest regression, support vector machine, stochastic gradient boosting model, neural network, and cubist). The results from the high-resolution yield maps were consistent with hand-harvest and yield-monitor data. Models using a cubist algorithm predicted corn grain yield with R2 of 0.76, 0.61 and 0.88 for fields F11, F8A and F7, respectively. Based on high-resolution corn grain yield maps, yield differences between rows that were most impacted by tires (i.e., pinch rows) and least impacted (i.e., non-pinch rows)