Science Wednesday: no need to go off-road

November 17th, 2021

Getting ground truth data is a common problem in agtech applications, and cost-vs-quality tradeoffs can be a significant concern. In a paper in the International Journal of Earth Observation and geoinformation, Graincast early-career fellow Franz Waldner and colleagues compared monitoring cropland extent using locations scattered along roadsides (as recommended by the Joint Experiment on Crop Assessment and Monitoring network) with a gold standard, but slower, random sampling strategy and with transects along roads.

When used to map crop extent across sites in 4 countries, their results lend support to roadside sampling as :

  • A “roadside” sample of given size was always less representative of the landscape than a random sample of the same size, but this translated into only a small (~2%) decrease in the accuracy of machine-learning derived crop maps
  • At a given level of representativeness, “roadside” samples were as accurate as random ones
  • Sampling only on roads within a 5km wide transect, on the other hand, was much less representative and less accurate
  • Improvements in accuracy with sample size were modest after about 5000 ground-truth data points.

The industry deployment of CSIRO’s Graincast crop mapping & yield forecasting technology continues to use roadside sampling

Examples of random, roadside and transect ground-truth-collection strategies